Awesome-AI-Memory

【中文 | English】

👋 Introduction

Large Language Models (LLMs) have rapidly evolved into powerful general-purpose reasoning and generation engines. Nevertheless, despite their continuously advancing capabilities, LLMs remain fundamentally constrained by a critical limitation: the finite length of their context window. This constraint defines the scope of information directly accessible during a single inference process, endowing models with only short-term memory capabilities. Consequently, they struggle to support extended conversations, personalized interactions, continuous learning, and complex multi-stage tasks.

To transcend the inherent limitations of context windows, AI memory and memory systems for LLMs have emerged as a vital and active research and engineering frontier. By introducing external, persistent, and controllable memory structures beyond model parameters, these systems enable large models to store, retrieve, compress, and manage historical information during generation processes. This capability allows models to continuously leverage long-term experiences within limited context windows, achieving cross-session consistency and continuous reasoning abilities.

Awesome-AI-Memory is a comprehensive repository dedicated to AI memory and memory systems for large language models, systematically curating relevant research papers, framework tools, and practical implementations. This repository endeavors to map the rapidly evolving research landscape in LLM memory systems, bridging multiple disciplines including natural language processing, information retrieval, intelligent agent systems, and cognitive science.

---

🎯 Goal of Repository

Our mission is to establish a centralized, continuously evolving knowledge base that serves as a valuable reference for researchers and practitioners, ultimately accelerating the development of intelligent systems capable of long-term memory retention, sustained reasoning, and adaptive evolution over time.

---

📏 Project Scope

This repository focuses on Agent Memory and memory mechanisms inside language models. Memory must be a central research question, with an explicit language-model or multimodal-language-model subject.

• Include: long-term, episodic, semantic, shared and experience-derived procedural agent memory; storage, retention, addressing, recall and forgetting inside language models; evaluations, security studies and surveys specifically about those mechanisms.
• Procedural skills must study persistent records formed from execution experience, later reuse, and validation or maintenance. Static skill packages, prompt optimization and general self-evolution do not qualify through keywords alone.
• Exclude: pure vision or low-level robot state, hardware/KV-cache efficiency, general RAG/long-context/tool routing, and adjacent work supported only by speculative transfer to LLMs. High-level language-agent memory in embodied or multimodal settings can qualify.
• Hold uncertain candidates until the memory user, content, operations, central contribution and supporting evidence are established. State the actual reading depth; abstract verification is not full-text review.

See the screening policy and publication gate.

---

🔔 Recent hot research and news

• 2026-11-05 - 🎉 Updated 72 papers, including 49 on Framework & Methods, 15 on Datasets & Benchmark, 8 on Systems & Models
• 2026-11-01 - 🎉 Updated 39 papers, including 31 on Framework & Methods, 5 on Datasets & Benchmark, 3 on Systems & Models
• 2026-10-27 - 🎉 Updated 34 papers, including 2 on Survey, 26 on Framework & Methods, 5 on Datasets & Benchmark, 1 on Systems & Models
• 2026-10-23 - 🎉 Updated 109 papers, including 1 on Survey, 70 on Framework & Methods, 28 on Datasets & Benchmark, 10 on Systems & Models
• 2026-10-18 - 🎉 Updated 32 papers, including 2 on Survey, 22 on Framework & Methods, 6 on Datasets & Benchmark, 2 on Systems & Models
• 2026-10-14 - 🎉 Updated 47 papers, including 33 on Framework & Methods, 11 on Datasets & Benchmark, 3 on Systems & Models
• 2026-10-11 - 🎉 Updated 53 papers, including 1 on Survey, 40 on Framework & Methods, 10 on Datasets & Benchmark, 2 on Systems & Models
• 2026-10-06 - 🎉 Updated 67 papers, including 43 on Framework & Methods, 15 on Datasets & Benchmark, 9 on Systems & Models
• 2026-10-03 - 🎉 Updated 69 papers, including 1 on Survey, 51 on Framework & Methods, 11 on Datasets & Benchmark, 6 on Systems & Models
• 2026-09-29 - 🎉 Updated 67 papers, including 1 on Survey, 37 on Framework & Methods, 17 on Datasets & Benchmark, 12 on Systems & Models
• 2026-09-26 - 🎉 Updated 17 papers, including 12 on Framework & Methods, 2 on Datasets & Benchmark, 3 on Systems & Models
• 2026-09-23 - 🎉 Updated 34 papers, including 19 on Framework & Methods, 7 on Datasets & Benchmark, 8 on Systems & Models
• 2026-09-20 - 🎉 Updated 11 papers, including 8 on Framework & Methods, 2 on Datasets & Benchmark, 1 on Systems & Models
• 2026-09-15 - 🎉 Updated 222 papers, including 3 on Survey, 153 on Framework & Methods, 32 on Datasets & Benchmark, 34 on Systems & Models
• 2026-07-06 - 🎉 Updated 25 papers, including 4 on Datasets & Benchmark, and 23 on Framework & Methods
• 2026-06-14 - 🎉 Updated 24 papers, including 2 on Survey, 4 on Systems & Models, 2 on Datasets & Benchmark, and 16 on Framework & Methods
• 2026-06-06 - 🎉 Updated 45 papers, including 1 on Survey, 6 on Systems & Models, 12 on Datasets & Benchmark, and 26 on Framework & Methods
• 2026-05-10 - 🎉 Updated 16 papers, including 3 on systems and models, 1 on benchmarks, and 12 on methods; also added 1 new project under systems and open sources
• 2026-05-06 - 🎉 Updated 16 papers, including 2 on systems and models, 2 on benchmarks, and 12 on methods
• 2026-04-27 - 🎉 Updated 15 papers, including 2 on survey, 3 on systems and models, and 10 on methods
• 2026-04-17 - 🎉 Updated 46 papers, including 1 on survey, 5 on systems and models, 3 on benchmarks, and 37 on methods
• 2026-04-07 - 🎉 Updated 16 papers, including 15 on methods, and 1 on benchmarks
• 2026-03-15 - 🎉 Updated 14 papers, including 14 on methods
• 2026-03-08 - 🎉 Updated 15 papers, including 3 on survey, 2 on systems and models, 5 on benchmarks, and 5 on methods
• 2026-03-02 - 🎉 Add a new code agent to this repo
• 2026-02-27 - 🎉 Updated 20 papers, including 1 on survey, 2 on systems and models, 2 on benchmarks, and 15 on methods
• 2026-02-26 - 🎉 Updated 14 papers, including 14 on methods
• 2026-02-14 - 🎉 Updated 15 papers, including 1 on survey, 12 on methods, 1 on benchmarks, and 1 on systems and models
• 2026-02-09 - 🎉 Updated 15 papers
• 2026-02-01 - 🎉 Updated 16 papers, including 9 on methods, 4 on benchmarks, and 3 on systems and models
• 2025-12-24 – 🎉 Release Repository V(1.0)
• 2025-12-10 – 🎉 Initial Repo

---

🗺️ Table of Contents

• Awesome-AI-Memory
  • 👋 Introduction
  • 🎯 Goal of Repository
  • 📏 Project Scope
  • 🔔 Recent hot research and news
  • 🧠 Core Concepts
  • 📚 Paper List
  • 🧰 Resources
    • 📊 Benchmarks and Tasks
    • 💻 Systems and Open Sources
    • 🎥 Multi-media resource
    • 🧠 Adam Framework
  • 🤝 Make a Contribution
  • 💬 Community & Support
  • 🌟 Star Trends

---

🧠 Core Concepts

• LLM Memory: A fusion of implicit knowledge encoded within parameters (acquired during training) and explicit storage outside parameters (retrieved at runtime), enabling models to transcend token limitations and possess human-like abilities to "remember the past, understand the present, and predict the future."

• Memory System: The complete technical stack implementing memory functionality for large language models, comprising four core components:

  • Memory Storage Layer: Vector databases (e.g., Chroma, Weaviate), graph databases, or hybrid storage solutions
  • Memory Processing Layer: Embedding models, summarization generators, and memory segmenters
  • Memory Retrieval Layer: Multi-stage retrievers, reranking modules, and context injectors
  • Memory Control Layer: Memory prioritization managers, forgetting controllers, and consistency coordinators

• Memory Operations: Atomic memory operations executed through tool calling in memory systems:

  • Writing: Converting dialogue content into vectors for storage, often combined with summarization to reduce noise
  • Retrieval: Generating queries based on current context to obtain Top-K relevant memories
  • Updating: Finding relevant memories via vector similarity and replacing or enhancing them
  • Deletion: Removing specific memories based on user instructions or automatic policies (e.g., privacy expiration)
  • Compression: Merging multiple related memories into summaries to free storage space

• Memory Management: The methodology for managing memories within memory systems, including:

  • Memory Lifecycle: End-to-end management from creation, active usage, infrequent access, to archiving/deletion
  • Conflict Resolution: Arbitration mechanisms for contradictory information (e.g., timestamp priority, source credibility weighting)
  • Resource Budgeting: Allocating memory quotas to different users/tasks to prevent resource abuse
  • Security Governance: Automatic detection and de-identification of PII (Personally Identifiable Information)

• Memory Classification: A multi-dimensional classification system unique to memory systems:

  • By Access Frequency: Working memory (current tasks), frequent memory (personal preferences), archived memory (historical records)
  • By Structured Degree: Structured memory (database records), semi-structured memory (dialogue summaries), unstructured memory (raw conversations)
  • By Sharing Scope: Personal memory (single user), team memory (collaborative spaces), public memory (shared knowledge bases)
  • By Temporal Validity: Permanent memory (core facts), temporary memory (conversation context), time-sensitive memory (e.g., "user is in a bad mood today")

• Memory Mechanisms: Core technical components enabling memory system functionality:

  • Retrieval-Augmented Generation (RAG): Enhancing generation by retrieving relevant information from knowledge bases
  • Memory Reflection Loop: Models periodically "review" conversation history to generate high-level summaries
  • Memory Routing: Automatically selecting retrieval sources based on query type (personal memory/public knowledge base)

• Explicit Memory: Memory stored as raw text outside the model, implemented through vector databases with hybrid indexing strategies:

  • Dense Vector Indexing: Handling semantic similarity queries
  • Sparse Keyword Indexing: Processing exact match queries
  • Multi-vector Indexing: Segmenting long documents into multiple parts, each independently indexed

• Parametric Memory: Knowledge and capabilities stored within the fixed weights of a language model's architecture, characterized by:

  • Serving as the model's core long-term semantic memory carrier
  • Being activatable without external retrieval or explicit contextual support
  • Providing the foundational capability for zero-shot reasoning, general responses, and language generation

• Long-Term Memory: Key information designed for persistent storage, typically implemented as external knowledge bases with capabilities including:

  • Automatic Summarization: Distilling multi-turn dialogues into structured memory
  • Context Binding: Recording memory context to prevent erroneous generalization
  • Multimodal Storage: Simultaneously preserving text, images, audio, and other multimodal memories

• Short-Term Memory: Active information within the LLM's context window, constrained by attention mechanisms. Key techniques include:

  • KV Cache Management: Reusing key-value caches to reduce redundant computation
  • Context Compression: Using summaries instead of detailed history (e.g., "the previous 5 dialogue rounds discussed project budget")
  • Sliding Window Attention: Focusing only on the most recent N tokens while preserving special markers
  • Memory Summary Injection: Dynamically inserting summaries of long-term memory into short-term context

• Episodic Memory: Memory type recording specific user interaction history, fundamental to personalized AI:

  • User Identity Recognition: Identifying the same user across sessions
  • Interaction Trajectory Recording: Preserving user decision paths and feedback
  • Emotional State Tracking: Recording patterns of user mood changes
  • Preference Evolution Modeling: Capturing long-term changes in user interests

• Memory Forgetting: Deliberately designed forgetting mechanisms in large models, including:

  • Selective Forgetting (Machine Unlearning): Removing the influence of specific information from training data, such as covering specific knowledge with forgetting layers
  • Privacy-Driven Forgetting: Automatically identifying and deleting PII information, or setting automatic expiration
  • Memory Decay: Automatically lowering the priority of infrequently accessed memories based on usage frequency
  • Conflict-Driven Forgetting: Strategically updating or discarding old memories when new evidence conflicts with them

• Memory Retrieval: The complex process of precisely locating relevant information from massive memory repositories:

  • Semantic Pre-filtering: Vector similarity matching to obtain Top-100 candidates
  • Contextual Reranking: Reordering results based on current query context
  • Temporal Filtering: Prioritizing the most recent relevant information

• Memory Compression: A collection of techniques maximizing memory utility under limited resources:

  • Content-level Compression: Extracting core information while discarding redundant details
  • Representation-level Compression: Vector quantization (e.g., PQ coding), dimensionality reduction
  • Organization-level Compression: Clustering similar memories, building hierarchical memory structures
  • Knowledge Distillation: Transferring key patterns from external memory into parametric memory

---

📚 Paper List

Papers below are ordered by publication date:

Survey

Date	Paper & Summary	Tags	Links
2026-10-26	Learning to Remember: Advances in Memory Mechanisms for Large Language Models
• Surveys LLM memory through lifecycle, storage substrate, and system role. • Distinguishes nonparametric memory, parametric memory, and memory for agent evolution, covering construction through governance and evaluation. • Offers a roadmap connecting mechanisms, benchmarks, and deployment concerns; it is a literature synthesis rather than a new empirical system.
2026-10-23	Agentic Memory Management Is the Next Frontier of Information Retrieval
• Positions agentic memory management as a research frontier for information retrieval. • Presents a system framework and a taxonomy of constants, functions, and operators with distinct retrieval semantics. • Outlines benchmarks, memory utility principles, and mixed parametric/nonparametric optimization; this is a perspective paper without new system-performance evidence.
2026-10-16	From Sliding Windows to Self-evolving Lifelong Memory: A Dual-perspective Survey of Memory Paradigms for Long-Horizon Robotics VLA Models
• Surveys memory paradigms for long-horizon VLA agents. • Organizes scope and persistence into five paradigms plus an emerging frontier, comparing architecture and openness. • Discusses lifecycle challenges and complexity/open-source trade-offs; this is a survey, not a new-method performance claim.
2026-10-15	From Long Context to Selective Memory:A Survey of Memory-Centric Compression for Large Language Models
• Surveys compression as selective language-model memory management across context and persistent storage. • Unifies writing, offloading, recall, injection, materialization, eviction, consolidation and forgetting through states and costs. • Identifies evaluation gaps and governance issues; this conceptual synthesis does not establish new-system performance guarantees.
2026-10-14	Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory
• Surveys multimodal agent memory through representation formation and adaptation of management policies. • Distinguishes static memory, dynamic updates and evolving policies, examining fidelity, abstraction, temporal validity and provenance. • Synthesizes design principles and open problems; its analytical taxonomy is not a new empirical performance result.
2026-10-06	Towards In-Parameter Memory Augmentation for Large Language Models
• Surveys in-parameter memory that represents post-pretraining facts, preferences, documents or experience in inference-time parameter objects. • Organizes methods by placement in embeddings, attention, FFNs or hybrids and by acquisition before or during deployment. • Discusses interference, safety and co-design with in-context learning; this is a survey rather than evidence for a single new model’s advantage.
2026-09-30	The Evolution of Attention in Large Language Models: Mechanisms, Trade-offs, and Emerging Trends
• Surveys attention architectures as internal contextual memory across representation, update, access, readout and integration. • Analyzes 59 release records, 14 model lineages and 11 open endpoints, covering explicit caches, recurrent states and cross-layer reuse. • Proposes a multidimensional stateful routing hypothesis; this is architectural synthesis rather than a unified controlled performance experiment.
2026-09-25	Machine Unlearning for Large Language Models: Foundations, Advances, and Agentic Extensions
• Surveys unlearning across model parameters and external agent state. • Uses five system layers and six evidence dimensions to compare removal claims and recovery tests. • Finds model-only evaluations insufficient to establish deletion across memory, tools, and subsequent updates.
2026-09-08	Graph-Based Personalized Memory for LLM Agents: Representation, Evolution, Retrieval, and Evaluation
• Surveys graph-based memory for long-term personalized agents. • Organizes studies around representation, evolution, retrieval, and evaluation. • Compares design choices and identifies reliability and controllability challenges.
2026-07-28	Memory for Large Language Models
• Surveys LLM memory from an architectural perspective. • Organizes mechanisms by representation, update dynamics, and persistence. • Synthesizes writing, routing, consolidation, efficiency tradeoffs, and evaluation directions.
2026-07-20	The Chronos Vulnerability: A Taxonomy of Temporal Persistence and Memory-Based Deception in Agentic AI
• Taxonomizes temporally persistent attacks on stateful agents. • Analyzes memory injection, delayed deception, and workflow-level threat models. • Synthesizes defensive directions spanning trajectory monitoring, verification, and trusted memory.
2026-06-25	Agents That Know Too Much: A Data-Centric Survey of Privacy in LLM Agents
• This survey frames privacy in LLM agents from a data-centric perspective, focusing on the data an agent queries, stores, remembers, exchanges, and acts upon rather than only on attack categories. • It organizes risks across databases, document collections, APIs, cross-session memory, intermediate workflow state, and multi-agent communication, showing how sensitive information can leak before the final response is produced. • The paper highlights that agent privacy requires lifecycle-level controls over memory writes, retrieval, delegation, logging, and tool use, making persistent state and data governance central to safe agent deployment.
2026-06-23	Are We Ready For An Agent-Native Memory System?
• This paper studies agent memory from a data-management perspective, arguing that memory for LLM agents should be evaluated as a system supporting persistent storage, retrieval, update, consolidation, and lifecycle governance rather than as a monolithic retrieval add-on. • It decomposes an agent-native memory system into four modules: representation and storage, extraction, retrieval and routing, and maintenance. Under this framework, the authors evaluate 12 representative memory systems and two baselines across five benchmark workloads spanning 11 datasets. • The experiments show that no single memory architecture dominates across all settings; effectiveness depends on the alignment between memory structure and workload bottlenecks. Fine-grained ablations quantify effects on representation fidelity, retrieval precision, update correctness, and long-horizon stability, while cost analysis shows that localized maintenance can be more efficient than global reorganization.
2026-06-10	Agentic Environment Engineering for Large Language Models: A Survey of Environment Modeling, Synthesis, Evaluation, and Application
• This survey systematically frames LLM agent environments across a full lifecycle, including environment modeling, synthesis, evaluation, and application. • It introduces symbolic and neural paradigms for environment synthesis and analyzes environment design through structural properties and capability evolution. • The work highlights co-evolution between agents and environments (memory, workflows, trajectories, and exploration) and points toward future directions such as environment-as-a-service and multi-agent ecosystems.
2026-06-09	Toward Secure LLM Agents: Threat Surfaces, Attacks, Defenses, and Evaluation
• This paper analyzes security risks of LLM agents from a lifecycle perspective and formalizes key interactions including information flow, delegated authority, and persistent state. • Based on a review of 247 studies, it organizes attack surfaces, defense mechanisms, and evaluation methods, showing that current defenses remain weak and incomplete. • It emphasizes the need for clearly defined trust boundaries, principled permission control, and realistic long-horizon security evaluation frameworks aligned with real-world environments.
2026-06-04	Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads
• Characterizes agent memory as a stateful long-horizon systems workload rather than only a retrieval add-on. • Introduces a system-oriented taxonomy and a phase-aware profiler for construction, retrieval, and generation cost. • Analyzes ten representative systems and distills practical recommendations for scalable, freshness-aware memory deployment.
2026-05-23	MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory
• Proposes MemEye, a visual-centric benchmark evaluating multimodal agents' long-term memory beyond text-dominated shortcuts. • Designs a 2D evaluation matrix (X-axis: Scene/Region/Instance/Pixel granularity; Y-axis: Atomic/Relational/Evolutionary reasoning depth). • Introduces three-stage validation gates ensuring questions cannot be solved via textual context, captions, or beyond VLM capabilities.
2026-06-01	EverMemOS: A Self-Organizing Memory Operating System for Structured Long-Horizon Reasoning
• Proposes EverMemOS, a three-stage memory lifecycle (Episodic Trace Formation → Semantic Consolidation → Reconstructive Recollection) that organizes fragmented conversations into structured, evolving memory units (MemCell/MemScene). • Introduces MemCell as atomic memory unit containing Episode + Atomic Facts + Foresight + Metadata, and MemScene for scene-level consolidation enabling long-term user profile evolution. • Designs a reconstructive retrieval paradigm with MemScene-guided retrieval, foresight validity filtering, and sufficiency verification to achieve "necessary and sufficient" context for long-horizon reasoning.
2026-05-19	An Agent-Oriented Pluggable Experience-RAG Skill for Experience-Driven Retrieval Strategy Orchestration
• Identifies three engineering problems in pipeline-based retrieval: non-reusable strategy logic, non-observable routing decisions, and non-evolvable workflow coupling. • Proposes Experience-RAG Skill with six modules (interface, scene analyzer, experience memory, strategy router, retriever pool, result packer) to encapsulate retrieval strategy selection as a pluggable agent skill. • Shows rule-based routing outperforms learned routing (0.8924 vs 0.8778/0.8627) under limited experience memory; experience memory records (scene_features, score_vector, best_margin) enable future upgrade to learning-based routing.
2026-04-17	Human Cognition in Machines: A Unified Perspective of World Models
• Surveys memory, perception, language, reasoning, imagination, motivation, and metacognition in world models from a unified cognitive-architecture perspective. • Introduces Epistemic World Models for structured knowledge agents in scientific discovery, with taxonomy and future directions across video, embodied, and cognitive world models.
2026-04-17	Agentic Frameworks for Reasoning Tasks: An Empirical Study
• Evaluates 22 common agentic frameworks on BBH, GSM8K, and ARC, comparing accuracy, latency, cost, and cross-benchmark consistency. • Finds that performance differences mainly come from orchestration quality, especially memory control, context growth, and failure-retry mechanisms.
2026-04-09	Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering
• Surveys LLM agent externalization from the perspective of memory, skills, protocols, and harness engineering. • Argues that memory externalizes cross-time state, skills externalize procedural capability, protocols externalize interaction structure, and harnesses orchestrate these components together. • Maps the evolution from model parameters to context and then to external infrastructure, while discussing the key trade-offs and open challenges across these layers.
2026-03-05	Beyond the Context Window: A Cost-Performance Analysis of Fact-Based Memory vs. Long-Context LLMs for Persistent Agents
• Addresses the ongoing industry debate over the performance and cost trade-offs between long-context models and external memory systems for building persistent agents. • Conducts a systematic cross-benchmark evaluation of long-context approaches and fact-based external memory solutions across three major memory benchmarks, assessing both accuracy and cumulative API inference cost. • Long-context models demonstrate advantages in factual recall, but their costs increase with each interaction turn; at a 100k context length, memory systems surpass them in cost efficiency after approximately 10 turns, providing a quantitative basis for real-world engineering decisions.
2026-03-02	Modular Memory is the Key to Continual Learning Agents
• Traditional foundation models rely on weight updates for continual learning, which can easily lead to catastrophic forgetting and makes large-scale experience accumulation difficult. • Proposes a roadmap for a modular memory architecture that integrates in-context learning with learning encoded in model weights. • This architecture leverages in-context learning for rapid adaptation and weight updates for capability consolidation, offering theoretical guidance for building truly lifelong learning agents.
2026-03-02	Emerging Human-like Strategies for Semantic Memory Foraging in Large Language Models
• Investigates whether large language models possess human-like, efficient strategic access mechanisms when processing vast amounts of semantic memory. • Uses mechanistic interpretability techniques to analyze semantic fluency tasks, rigorously examining convergent and divergent memory search patterns within the model. • Confirms the presence of human-like strategic memory search behavior across different layers of LLMs, laying an interpretability foundation for research on cognitive alignment and enhanced human-AI collaboration.
2026-02-26	Toward Personalized LLM-Powered Agents: Foundations, Evaluation, and Future Directions
• This paper explores the foundations, evaluation, and future directions of personalized LLM-driven agents, providing a capability-oriented systematic review of their underlying principles, assessment methodologies, and prospective developments. • It constructs a taxonomy centered on four interdependent core components: user profile modeling, memory, planning, and action execution. • The paper offers a comprehensive analysis of how user signals are represented, propagated, and utilized, and discusses application scenarios and design trade-offs ranging from general-purpose assistance to specialized domains.
2026-01-14	Rethinking Memory Mechanisms of Foundation Agents in the Second Half: A Survey
• This paper proposes a unified taxonomy for foundation agent memory across three dimensions: memory substrate, cognitive mechanism, and memory subject. • It analyzes memory operation mechanisms in single-agent and multi-agent systems while categorizing memory learning policies into prompting, fine-tuning, and reinforcement learning paradigms. • It provides a comprehensive review of evaluation metrics and benchmarks across various applications, outlining critical future directions such as memory efficiency, life-long personalization, and multimodal integration.
2025-12-15	Memory in the Age of AI Agents: A Survey
• Provides a comprehensive and up-to-date landscape of agent memory, explicitly distinguishing it from related concepts like LLM memory, RAG, and context engineering. • Introduces a unified taxonomy examining memory through three lenses: Forms (token-level, parametric, latent), Functions (factual, experiential, working), and Dynamics (formation, evolution, retrieval). • Discusses emerging research frontiers such as automation-oriented memory design, reinforcement learning integration, and trustworthiness, while compiling representative benchmarks and frameworks.
2025-09-18	A Survey of Machine Unlearning
• Provides an in-depth exploration of the concept and background of machine unlearning, highlighting its importance in modern machine learning. • Machine unlearning aims to enable learning algorithms to effectively remove the influence of specific data without requiring full model retraining. • The paper analyzes the necessity, challenges, and design requirements of machine unlearning, reviews current research progress, and emphasizes the field’s complexity and diversity in terms of algorithmic effectiveness, fairness, and privacy protection.
2025-09-02	A Survey on the Memory Mechanism of Large Language Model based Agents
• Explores the memory mechanisms of LLM-based agents, emphasizing the crucial role of memory in agent self-evolution and complex interactions. • Systematically summarizes and categorizes existing memory module designs and evaluation methods, while analyzing their roles and limitations across different application scenarios. • Such agents are able to improve decision-making and task execution.
2025-05-31	A Survey of Machine Unlearning in Large Language Models: Methods, Challenges and Future Directions
• The paper investigates machine unlearning in large language models (LLMs), aiming to effectively remove the influence of undesirable data (e.g., sensitive or illegal information) without full retraining, while preserving overall model utility. • It defines the objectives and paradigms of LLM unlearning and establishes a comprehensive taxonomy. • The paper reviews existing approaches, evaluates their strengths and limitations, and discusses opportunities for future research.
2025-05-27	Rethinking Memory in AI Taxonomy, Operations, Topics, and Future Directions
• Explores multidimensional research on memory in artificial intelligence (AI), with a particular focus on memory operations and management in large language models (LLMs). • Categorizes various types of memory representations and operations—including integration, updating, indexing, forgetting, retrieval, and compression—and provides a systematic analysis of the importance of memory in AI and how it is implemented. • Through an extensive review of the literature, the paper identifies four key research themes: long-term memory, parametric memory, long-context memory, and multi-source memory integration.
2025-04-24	Cognitive Memory in Large Language Models
• Provides a comprehensive examination of memory mechanisms in large language models (LLMs), with a particular focus on different types of memory and their roles within the models. • While LLMs excel at information retrieval and interaction summarization, their long-term memory remains unstable. • Integrating memory into AI systems is crucial for delivering context-rich responses, reducing hallucinations, improving data processing efficiency, and enabling the self-evolution of AI systems.
2025-04-23	From Human Memory to AI Memory A Survey on Memory Mechanisms in the Era of LLMs
• Explores the relationship between human memory and the memory mechanisms of LLM-based artificial intelligence (AI) systems. • The main contributions include a systematic definition of memory in LLM-driven AI systems and its conceptual linkage to human memory. • The paper proposes a three-dimensional memory taxonomy based on object, form, and time, and summarizes key open issues in current research on personal memory and system memory.
2025-04-02	Digital Forgetting in Large Language Models: A Survey of Unlearning Methods
• The paper explores digital forgetting in large language models (LLMs) and corresponding unlearning methods, with a focus on addressing issues related to privacy, copyright, and social ethics. • It analyzes different types of model architectures and training processes, as well as practical approaches to digital forgetting, including data retraining, machine unlearning, and prompt engineering. • By introducing the concept of “forgetting guarantees,” the paper emphasizes effective mechanisms for both exact and approximate forgetting.
2025-01-12	Human-inspired Perspectives: A Survey on AI Long-term Memory
• This paper systematically examines the interplay between human long-term memory mechanisms and AI long-term memory, and proposes an adaptive long-term memory cognitive architecture (SALM). • It introduces the structure of human memory, including sensory memory, working memory, and different types of long-term memory (episodic, semantic, and procedural memory). • The paper analyzes the classification of AI long-term memory—parametric and non-parametric memory—as well as their storage and retrieval mechanisms.

Framework & Methods

Date	Paper & Summary	Tags	Links
2026-11-04	FlyHash for LLM Agent Memory: An Empirical Validation and a New Application
• Applies FlyHash to LLM-agent memory retrieval and audits comparison protocols. • Compares sparse projection-and-competition codes with SimHash and HNSW using matched bits, multiple seeds and genuine relevance labels. • Quality advantages persist, but HNSW wins practical latency and storage comparisons at ordinary scale; per-bit quality is not deployment efficiency.
2026-11-04	Can LLM Agents Learn to Test? Maintaining Scoped Experience as Testing Evidence for Mobile GUI Testing
• Presents an AdaptiTest maintenance contract treating recalled testing experience as advice. • Preserves applicability and trace provenance with confirmation, revalidation and withdrawal, reserving persistence for a dedicated role. • A small internal subset reports 18/18 versus 3/18, but controls, confounded probes and incomplete enforcement prevent claims of general superiority.
2026-11-04	Weaver: Weaving Preference Evidence through Incremental Reasoning for Personalized Retrieval-Augmented Generation
• Introduces Weaver for scattered, context-dependent preference memories. • A dual-layer graph links intent dependencies and retrieval trajectories, filling information gaps while separating complementary, alternative and conflicting support. • Improves evidence completeness and answer quality on personalized benchmarks including PersonaMem; the reported scope does not establish deployment-wide generality.
2026-11-04	Reglos: A Self-Evolving Personal Memory Agent with Recursive Self-Improvement
• Presents Reglos, combining evolving personal memory with controlled harness improvement. • Appends conversation records and handles preference replacement, while cross-family criticism screens patches and a fixed evaluator gates acceptance. • The Qwen configuration scores 61.3% and 84.0% on PersonaMem-v2 and LongMemEval-S; these are system-level comparisons rather than isolated component effects.
2026-11-04	SUMONE: A Simple and Efficient Summarizer for Plug-and-Play Agent Memory
• Introduces SumOne, learning portable summary memory from agent trajectories themselves. • A frozen reconstructor recovers masked trajectory portions from summaries, supplying reconstruction and length rewards to a standalone summarizer. • Estimated inference cost falls about 78% and 75% on search and software-engineering tasks with small performance decreases; the checkpoint transfers across harnesses.
2026-11-04	PyRUA-REAL: Robot Exploration and Adaptive Learning through RSI in the Real World
• Presents PyRUA-REAL for turning physical exploration into persistent experience and executable skills. • Stores conditions, outcomes and code separately; researchers record applicability while engineers retrieve, adapt and retest implementations. • An eight-task transfer comparison improves from 21.3% to 40.0%, with gains on five tasks; overall system success is not solely attributable to memory.
2026-11-03	Anchor: Augmented Cross-Layer Recurrent State for Long-Context Retrieval in Test-Time Training
• Introduces Anchor, exploiting retrieval concentrated in a few layers. • Places TTT memory at anchor layers, shares one written recurrent state and reads it through layer-specific learned maps. • Reports 2.3–3.1x higher recall with tenfold smaller states on Llama and Mistral in the tested linearized hybrid setting.
2026-11-03	Sketchpad: certified exact arithmetic with a small vision-language model
• Introduces Sketchpad, making arbitrary-length arithmetic certifiable through restricted working-memory access. • The model reads three cells and emits digits and carry states; Lean reduces universal correctness to checking a finite set of views. • All 2,446 views pass on the deployment device; guarantees depend on the fixed protocol and verification conditions, not general reasoning ability.
2026-11-03	CogniLife: Continuous Cognitive Distillation for Deployed Multi-Expert Agent Marketplaces
• Proposes CogniLife for cross-expert memory attribution and continuing revision. • Separates Creator Profiles from Expert Journals; implemented parts support scoped extraction, validity intervals, provenance and reversible consolidation. • Cross-expert routing is unevaluated and unimplemented; scores on 85 tasks do not establish cold-start mitigation, transfer or longitudinal distillation.
2026-11-03	The Right Source Is Not Enough: Understanding History Interference in Speech-LLM Memory
• Studies interference from irrelevant historical audio despite retained target evidence. • Blocks source access and applies low-dimensional interventions to separate direct reads, candidate-state propagation, source roles and answer values. • Open QA and multiple choice show different paths; joint role/value interventions recover most errors in MOSS-Audio and MiMo-Audio.
2026-11-03	MDSKILL: Learning to Internalize Skill Memory through Multi-View Distillation for Efficient Dynamic Agent Self-Evolution
• Introduces MDSKILL to reduce noise, conflict and redundancy when internalizing skill memory. • Splits experience into grounding, planning and execution views, dynamically selects skills and distills branch knowledge into one policy. • Reports gains up to 7.1% and 8.1% on ALFWorld and WebShop, with supporting ablations across three agent benchmarks.
2026-11-03	HARP: Training-Free Dual-Profile Agentic Communication for LLM-Based Recommendation
• Introduces HARP for two-sided profile memory and subsequent feedback synchronization. • Builds item profiles, applies profile-homophily signals during retrieval and updates user descriptions through one reflective pass. • Hit@1 improves 6.9–24.0% over the strongest published baseline on four datasets; reflection adds trade-offs and retrieval gating contributes most.
2026-11-03	A Circuit for In-Context Learning
• Identifies a gated key-value memory circuit for in-context learning. • Middle-layer heads write class keys at label positions; upper heads read matching labels, with patching, ablations and steering testing gates and recency suppression. • Finds a shared circuit in three 3–4B models with further transfer checks, primarily within the studied few-shot classification setting.
2026-11-03	Route to Remember: Adaptive Memory Selection for LLM Agents
• Studies weight memory as a low-context-cost store for long-term user facts. • Uses on-policy self-distillation from a memory-conditioned teacher and a verify-then-regenerate router combining text and weight memory. • Matches or exceeds strong text-memory baselines on HaluMem with fewer input tokens, within the evaluated factual-memory protocol.
2026-11-03	Prospective Memory for Proactive LLM Agents
• Introduces Pro-Memory to separate retaining intentions from acting at the right moment. • Links pending goals to evidence and activation conditions, then updates intention state from issued actions or resolution evidence. • Improves proactive performance across backbones on two benchmarks while limiting unnecessary interventions in the tested scenarios.
2026-11-02	Agent Memory as Rate-Limited Filtering: Explicit State Outlasts Long Context Beyond a Reader-Dependent Horizon
• Formalizes agent memory as capacity-limited recursive filtering. • Uses executable access policies and paired histories to compare full context, retrieval and explicit state, separating model parsing from rule updates. • Parsing plus rule updates achieves 80–89% at 256 events; horizons depend on readers, and dependency-based retrieval matches some results.
2026-11-02	EvoMemState: Evolving Memory State for Personal Agents
• Introduces EvoMemState, moving personal memory from independent facts to evolving states. • An ontology induces facts and validity relations; reinforcement learning generates request-specific reasoning state to guide their use. • Improves retrieval completeness and memory-grounded accuracy on reported dialogue and memory-use benchmarks.
2026-11-02	RecGroupMemory: Feedback-Aware Cohort Memory for E-Commerce Recommendation
• Introduces RecGroupMemory, separating cohort-memory construction from feedback revision. • Previous-day outcomes revise decision guidance while validation preserves selected source values for the orchestration LLM. • In a selected low-purchasing-power cohort, memory without and with feedback increases guided transaction users by 1.58% and 1.78%; broader generalization is untested.
2026-11-02	Chain on Demand: Deferred Evidence Organization for Long-Term Conversational Memory
• Introduces Chain on Demand, deferring evidence organization until query time. • Selects anchors under a fixed budget using relevance, lexical diversity and new query-term coverage, then restores same-session neighborhoods. • Scores 82.2% on non-adversarial and 76.3% on all LoCoMo questions, and 86.6% on LongMemEval-S, without changing the underlying memory.
2026-11-02	HeLM: Hierarchical Explicit Language Memory for Non-Markovian Manipulation
• Introduces HeLM to preserve manipulation history in explicit language. • Factorizes event detection and event-triggered updates, organizing hierarchical state that can persist across task boundaries. • Across 10 real-world tasks and 56 variants, intra-episode success gains 12.8 points and compositional subtask success reaches 78% in the tested settings.
2026-11-02	A State Change Is Local, Its Consequences Are Not: Dependency-Aware State Revision for Long-Term Agent Memory
• Introduces StateGraph to revise dependent conclusions when facts change. • Recovers support dependencies and verifies edges counterfactually, invalidating states only when strict support fails without an independent path. • Gains grow with cascade depth on 50 constructed cases; this is preliminary controlled evidence, with comparable or higher results on most public benchmarks.
2026-11-02	Top-$k$ Retrieval Leaves Memory Value Undetermined
• Shows that standard top-k retrieval logs can leave individual memory values unidentified. • Uses individual and pairwise retrieval counts; LEDGER estimates only identifiable Shapley values and retrieves to improve identification. • No memory value is identified in 55 benchmark configurations; adaptive retrieval improves identification but costs 2.6 accuracy points on HotpotQA.
2026-11-02	Rethinking Next-Visit Diagnosis Prediction Beyond Aggregate Metrics: Retrieval--Memory Multi-Agent Adjudication
• Analyzes retrieval and patient-trajectory memory across temporal diagnosis transitions. • Separates persistent, new and recurrent labels; LONGINUS retains cross-view agreement and adjudicates disagreements without revealing source identity. • Most transition-specific metrics improve across four LLMs on MIMIC data, but last-visit repetition remains strong and clinical validity is unestablished.
2026-11-02	From Benchmarks to Behaviors: A Framework for Auditing AI Agents
• Uses TRACE to audit how user memory changes source selection. • Holds questions and search results fixed while manipulating concise memories and logging execution and outputs. • Publisher preferences raise top-three citation inclusion by 33–39 points; some distrust memories also increase misinformation-source citations, with model-dependent effects.
2026-11-02	WASP-Mem: Let Agent Organize Procedural Memory. Human Spend Effort on Diagnosis
• Introduces WASP-Mem, letting agents organize execution experience into linked procedural wikis. • Consolidates successes, failures and feedback while controlling ingestion scope, organization, feedback type and refinement rounds. • TheAgentCompany pass rates gain 4.6–6.1% with 57–61% less storage; human diagnosis is more useful than extra self-judgment rounds in this setting.
2026-11-01	Can Large Language Model Agents Learn by Watching Tutorial Videos?
• Introduces TAL to consolidate tutorial videos into an expandable tutorialbook memory. • Uses adversarial question harvesting and memory refinement, with reasoning-guided table-of-contents retrieval. • Evaluates 72 chess tutorials and a second domain; stronger opening positions occur in 69% of paired games, within the tested settings.
2026-11-01	Meta-memorization and memorization scaling laws in transformers
• Studies why larger Transformers memorize new sequences faster through a meta-memorization effect. • Analyzes position-invariant attention signals, MLP storage and an attentional bottleneck using random binary sequences. • Reproduces capacity and learning-speed scaling; the synthetic setting limits conclusions about natural-language knowledge.
2026-11-01	SPAR: Experience-Driven Strategy Evolution in Multi-Session B2B Sales
• Introduces SPAR, maintaining sales experience as testable strategy hypotheses. • An episodic inner loop adapts to merchants while causal attribution and hypothesis-driven revision update a shared playbook. • An online A/B test over 200,000 merchants reports 15.8% conversion lift; this platform-level result does not isolate every memory component's causal effect.
2026-11-01	Evolving Memory, Learning to Meta-Control: Test-Time Adaptation for Experience-augmented LLM Agents
• Introduces OMCE to evolve both experience memory and its control policy. • A contextual bandit selects direct execution, episodes, guidance or distillation, updating memory and routing from task feedback. • First-attempt success improves by up to 20.5 points across four backbones and benchmarks; token savings over ACE depend on the reported setup.
2026-11-01	MemRole: From Memories to Decisions for Long-Horizon Reasoning in Role-Playing Agents
• Introduces MemRole to connect past experiences, motivations and present decisions. • Trains context understanding, memory activation and strategy planning with character- and timeline-constrained memory pools, followed by self-critique distillation. • The 32B model gains 15.2 points on memory-grounded role play and improves motivation recognition within narrative-derived benchmarks.
2026-11-01	Hippocampal consolidation without replay in fully fine-tuned networks
• Introduces DeltaHippo, using a training-time hippocampal index to protect knowledge during continual learning. • Stores input subspaces, energies and residuals to constrain updates before and after the optimizer; the index is removed at inference. • Direction-rotation and read-ablation experiments support content-dependent retention; TACRED gains do not imply comprehensive preservation of generation abilities.
2026-11-01	SurgStream: A Workflow-Anchored Proactive Streaming Framework and Benchmark for Real-Time Surgical Video Understanding
• Introduces SurgStream with workflow-anchored multimodal event memory. • Stable event changes trigger triplet records, phase boundaries consolidate summaries and deadline-aware supervision controls response timing. • A benchmark of 11.7K interactions shows memory and proactive-QA gains; video-task evidence does not validate clinical deployment.
2026-11-01	TrackMem: Tracking Evolving Beliefs in Long-Horizon Financial Dialogue
• Introduces TrackMem for evolving beliefs in long-horizon financial dialogue. • Builds speaker-specific graphs preserving numerical revisions, entity lenses and action order, with FinLoCoMo testing failure modes. • Across 15 conversations and 765 questions, failure-mode macro accuracy is 73.9% versus 69.4%; this small benchmark does not establish investment quality.
2026-11-01	Do Embodied VLMs Actually Use Their Scene-Graph Memory?
• Introduces graph swapping to test whether embodied VLMs actually use scene-graph memory. • Replaces episode graphs and compares textual, pooled and node-token delivery through answer changes. • Text memory adds 30 points on route questions and node tokens recover 8.3 points; redundant textual facts can make the model ignore its graph.
2026-11-01	Do agents need to sleep?
• Proposes a sleep-inspired selectionist account of agent-memory consolidation. • Maps ringing, resonance and renormalization to digital operations addressing drift, stored errors and capacity saturation. • Offers falsifiable conceptual predictions; empirical superiority over alternative consolidation mechanisms is not established.
2026-11-01	When the Text Is Silent: Memory Shortcuts and the Validity of LLM Annotation under Evidence Insufficiency
• Studies parametric-memory shortcuts when annotation evidence is absent. • Uses 995 public figures and matched controls in counterfactual texts, varying evidence and incentives and testing reminders and name masking. • Shortcuts arise mainly under missing rather than conflicting evidence; factually correct labels may lack textual support, with attribute-specific mitigation trade-offs.
2026-10-31	Who Schedules Agent Memory? A Sidecar Approach to Long-Horizon Tasks
• Introduces MemSidecar to separate memory scheduling from task execution. • Hooks delivers event-bound memories; Foresight predicts queries and prefetches results for later validity-checked delivery. • The modes score 41.23% and 44.34% on MCP-Persona while reducing cost and latency, and can be selected independently.
2026-10-31	Pulse: Long-Horizon Tasks Need Streaming Personalization
• Presents Pulse for streaming personalization during long-horizon agent execution. • A separate memory agent detects emerging needs and conflicts, with reinforcement learning optimizing intervention timing and content. • Reports 71.9% on π-Bench and 46.4% on MCP-Persona alongside efficiency gains on the evaluated tasks.
2026-10-30	MEMIG: Self-evolving Memory Construction for Antibody Optimization
• Introduces MEMIG to replace free-text rewriting of scientific optimization experience. • Indexes quantitative oracle measurements for local updates and exact lookup, linking joint interventions through measured epistasis. • Improves most evaluated in-silico antibody-optimization settings; this is not experimental or clinical efficacy evidence.
2026-10-30	When History Overrides Intent: Relational Schema Lock-In in Human-LLM Interaction
• Studies relational schema lock-in, where history-grounded user models override current instructions without authorization. • Combines user reports, controlled histories across six models and a 360-participant interaction study. • Artifact substitution can preserve completion while reducing faithful enactment; findings are specific to the tested interaction conditions.
2026-10-30	What Does a Repair Agent Gain from Remembering Its Failure? An Exploratory Study on Real Python Bugs
• Compares forms of failure memory in coding-agent retries. • Tests plain retries, transcripts, lessons and test signals after 181 failures on 75 BugsInPy bugs, with model and notice-only follow-ups. • Memory gains of 5.5–9.9 points are not significant; lessons do not beat transcripts, and full-suite checks reduce apparent benefits.
2026-10-30	KEEPER: Learning Control-Sufficient Textual Memory for Long-Horizon LLM Agents
• Studies persistence of incorrect claims in recursive textual memory through KEEPER. • Compares four auxiliary objectives along information-content and update-sensitivity axes, with offline RL and claim-level trajectory analysis. • Reward-related prediction performs best on tasks; error-persistence evidence includes a single trajectory and does not establish general memory reliability.
2026-10-30	KITE: Retrieval as a Decision Procedure for Long-Term Conversational Memory
• Introduces KITE, treating conversational-memory retrieval as an inspectable decision procedure. • Construction and query compilation share a governed topic index, supplemented by sparse lexical retrieval for omitted details. • Reports 93.5% on LoCoMo and 86.0% on LongMemEval-S with about 1.6k reader-context tokens per question in the tested setup.
2026-10-30	ORION: An Outcome-Driven LLM Advisor for RL-Based Multi-Domain 6G Slice Placement
• Presents outcome-driven episodic memory for ORION's LLM slice planner. • Stores network conditions, proposed plans and verified feasibility, feeding successes and failures back with importance-based retention. • With downstream execution fixed, admission improves by 4.6–11.5 points on held-out topology families within the evaluated network setting.
2026-10-30	DAMI: Dual-Axis Memory Initialization for Enhanced Long-Context Retrieval in Gated Linear Attention
• Introduces DAMI to control recurrent-memory decay and input sensitivity through initialization. • Jointly initializes gate parameters with geometric head timescales and a layerwise sensitivity profile calibrated from a small activation sample. • Retrieval improves by 5.29–16.71 points across five recurrence families, while general language modeling and commonsense performance remain broadly comparable.
2026-10-30	SPARK: Self-Correcting Embodied Memory via Physical Re-Observation
• Presents SPARK for correcting embodied memory through targeted physical re-observation. • A semantic critic flags suspect regions, a planner schedules revisits and selective merging adds, removes or replaces records. • EQA improves on four AirSim maps with 13–41% fewer stored objects than append-only updates; evidence is limited to simulated maps and 158 fixed questions.
2026-10-30	From Lookups to Lessons: Turning Retrieved Knowledge into Parametric Memory
• Introduces ResMem to turn retrieved evidence into reusable parametric memory. • An evidence-conditioned teacher trains an additive logit residual while updates remain confined to the memory module. • Mean F1 rises from 32.83% to 34.48% on five QA datasets, breaking 21% fewer correct answers than the kNN-trained memory in the evaluated pipeline.
2026-10-29	Fewer Heads, Fewer Memories: Pruning Redundant Attention Heads Suppresses Verbatim Memorisation
• Tests whether redundant attention heads support verbatim memorization. • Zeros task-redundant heads before inference and compares known-sequence reproduction with WikiText-103 language modeling. • Pruning 729 heads reduces memorization count by 91.5% while improving perplexity by 0.92, within GPT-2-XL and the evaluated benchmarks.
2026-10-29	Accept Changes, Reject Attacks: Robust Memory Management for LLM Agents
• Studies outdated memory and poisoning together, since they demand opposing write decisions. • Evidence-conditioned accuracy checks answers and whether selected records are complete, current and attack-free; synthesized episodes tune the manager. • Overall scores more than double the untuned manager across three backbones, but robustness remains incomplete.
2026-10-29	MemSteer: Turning User Histories into Supervision for Agent Memory Optimization
• MemSteer derives supervision for memory optimization from unlabeled user histories. • Evidence-grounded answer plans guide QA generation, source audits and blind recovery; failures drive storage and retrieval revisions. • Relative to Base-BM25, gains are 13.4 points on ATM-Bench and 4.1/6.6 points on RHELM without/with external data.
2026-10-29	Federated Recommender Agents
• FedRecAgent shares natural-language experience while keeping personal histories and memories local. • Condition–Transition–Reasoning units organize shared knowledge; semantic and decision memories guide retrieval. • Dense and sparse evaluations across three domains outperform tested baselines, with memory and integration ablations.
2026-10-29	Ask to Evolve: A Unified Harness for Lifelong Embodied Navigation with Hybrid Action Space and Persistent Memory
• EvoLife combines active help-seeking, navigation actions and persistent spatial memory. • Observations and dialogue are attached to places and reused on later visits and tasks. • R2R revisit SPL rises from 82.23% to 93.31%, with additional multi-day physical demonstrations in the evaluated settings.
2026-10-29	Sticky Collusion: Persistence of LLM Agent Collusion after Incentive Removal and Partner Replacement
• Tests whether interaction memory sustains collusion beyond its reward or original partner. • Branches from established collusive histories under incentive, partner and complete or partial memory-reset interventions. • Complete wiping ends collusion in tested models, whereas 8/9 partial-removal combinations retain it; evidence is limited to the specific two-agent verification setting.
2026-10-29	Long-Term Agent Memory through Complementary Curation and Consolidation
• Combines episodic and relational memory for long-horizon dialogue. • A hybrid architecture retains facts, events and relationships with an efficiency-preserving consolidation mechanism. • Reports 90.2% on LoCoMo and 85.5% on LongMemEval-S; evidence covers these two benchmarks.
2026-10-29	Low-Bit GDN Recurrent-State Quantization: Global Gradient-aware Allocation and Pure-Write Sliding Windows
• Studies quantized recurrent-memory writeback errors and preservation of recent information. • Allocates precision by future-loss sensitivity and keeps recent pure writes as FP8 vector pairs before merging them into history. • Seven-task averages reach 69.83%/78.63% on Qwen3.5-4B/9B under the reported precision and window settings.
2026-10-29	Don't Lose the Thread: Empowering Long-Horizon LLM Agents with Cognitive Resource Self-Allocation
• CORAL gives long-horizon agents explicit control over working-memory clutter. • Agents checkpoint progress and verified facts, reset context, and resume; multi-episode reinforcement learning improves checkpointing. • An 8B model reaches 42.8% on GAIA with roughly 6–7 times fewer input tokens than ReAct; additional tasks test transfer.
2026-10-29	When to Assist and What to Suggest: Joint Gating and Memory-Grounded Hint Generation in One Compact Decoder
• Studies when to use user memory and what proactive hint to generate jointly. • A compact decoder gates at an intermediate layer and generates memory-grounded hints in the same pass. • Across 28,058 decision points, removing memory costs 24.7 exact-match points; human ratings still favor the 8B generator over the single model.
2026-10-28	Is it my preference? Role Grounding Attack against Agent's Memory via Social Interaction
• Identifies role-grounding errors that turn an attacker’s preferences into a victim user’s memories. • Nested social conversations lose attribution during memory formation; later retrieval carries the false preference into decisions. • Tests across five memory systems and six language models report tool-manipulation and privacy-leakage risks.
2026-10-28	What Should a Navigator Remember? Spatiotemporal Memory for Zero-Shot Vision-Language Navigation
• Studies what zero-shot language navigators should retain and how to organize it. • Task, temporal and spatial memories provide progress, pose-anchored observation summaries and geometry to an 8B reasoner. • Removing temporal memory increases mean path length from 16.9 to 39.1 metres; R2R-CE and physical trials provide additional evidence.
2026-10-28	MUSE: Multi-Level Memory with Selective Runtime Injection for Scientific Coding Agents
• MUSE organizes scientific-coding experience into global, approach and implementation memories. • Outcome-grounded extraction and source review support replacement or retirement; runtime gating injects situation-specific checks. • Four scientific-coding evaluations and cross-kernel transfer support gains; memory banks are frozen before testing.
2026-10-28	Security Impact of Coding Agent's Memory Files
• Studies project memory files as a persistent attack surface for coding agents. • Compares two malicious guidance variants with controls on 36 SecRepoBench tasks across 12 CWE classes, including disclosure behavior. • Secure completion changes range from −8 to +6 points without a consistent attack effect; disclosure differs sharply across agents.
2026-10-28	CALM: Conflict-Aware Lifecycle Memory Governance for Long-Horizon Language Agents
• CALM separates preservation from query-time admission so superseded memories remain usable for historical reconstruction. • Support, conflict and dependency signals update validity; temporal applicability governs evidence admission. • Reported gains on LoCoMo, LongMemEval-S and conflict tests support this separation in the evaluated settings.
2026-10-28	Routed Continual Learning
• Routed CL selects both a memory mechanism and a state cluster for experience sharing. • Cluster state augments context before execution and only the selected cluster is updated afterward, using trajectories, strategies or structured state. • Across 398 interleaved instances and four models it improves over global history and reduces estimated API cost; the memory-system library is fixed before evaluation.
2026-10-28	HEDGE: Gated Hypergraph Memory for Scalable Multi-Agent Debate
• HEDGE organizes multi-round debate memories as semantic hypergraphs. • Hyperedges connect related claims, while a learned selector assigns STORE, BACKUP or OMIT. • Across seven benchmarks, it reports a 5.0-point average gain and 48% lower memory use under the tested configuration.
2026-10-27	Remembering Is Not Enough: Refining Memories and Learning When to Use Them in Multi-Agent Systems
• DiscernMem repairs procedural memories and learns when each should be used. • Single-memory versus no-memory runs diagnose defects; repairs are retested on current and prior tasks, while meta-memory records role-specific effects. • Across four benchmarks and three frameworks, average gains are 0.51–4.32 points over the strongest reported baselines, using one Qwen3-4B backbone.
2026-10-27	From Voiceprints to User Entities: Online Speaker-Memory Construction for Multi-User Spoken Agents
• EvoSpeakerMem distinguishes voiceprint observations from persistent user identities for memory access. • Causal routing and selective acoustic updates govern identity-partitioned writes and retrieval; LoCoMo-ID tests hidden identities. • Identity-gated QA gains 3.86/11.62 points under clean/shifted conditions, but less fragmentation can still coexist with incorrect cross-user links.
2026-10-27	CAMR: Compatibility-Aware Memory Replay for Continual Learning in Large Language Models
• CAMR jointly controls incoming-example learning and historical memory replay. • Compatibility checks and a learned gate select general replay, related reinforcement or skipping mastered duplicates using acquisition, retention and cost feedback. • Qwen3-8B-Base shows retention benefits on factual updates and task sequences; other model scales are not established here.
2026-10-27	Select-TTT: Query-Conditioned Selective Fast Weight Updates for In-Place Test-Time Training
• Select-TTT addresses irrelevant-content interference caused by uniform fast-weight writes. • Frozen output-head relevance scores modulate per-chunk learning rates through a temperature softmax, without new parameters. • Long-context QA improves across two benchmarks and two backbones; multi-query agent use remains prospective.
2026-10-27	TrawMem: Reframing Agent Memory as Task-Oriented Role-Adaptive Workspaces
• TrawMem organizes memories into task-oriented, role-adaptive workspaces. • Local interaction threads support joint selection of direct evidence and complementary, less-similar memories. • Across two benchmarks and four backbones, macro-F1 improves by 2.23–3.14 points; some LoCoMo settings also reduce context tokens.
2026-10-27	TDLM: Adaptive Temporal Integration for Diffusion Language Models
• TDLM adds memory of hidden representations across diffusion steps. • Within a frozen denoiser, unresolved positions attend causally to their own history; training uses self-generated trajectories. • Adds about 0.84% parameters on LLaDA and reports up to 7.15× faster generation across LLaDA/Dream with competitive tested quality.
2026-10-27	Self-Keyed In-Parameter Memorization for External Memory-Free Dialogue Recall in Frozen Language Models
• SKIM studies dialogue storage and recall without an external memory store. • The model generates diverse questions as keys and writes turn content into parameters through a closed-form method. • Across five language models it recalls most stored turns; the abstract does not provide a single aggregate numerical score.
2026-10-27	Proactive Memory for Event-Driven LLM Agents
• ProactMem validates historical behavioral regularities before using them for proactive service. • Source events ground trajectories and historical replay assesses regularities before current-event matching. • Introduces longitudinal ProactiveStream and reports 52.6% F1, a 21.4% relative gain over the strongest external baseline.
2026-10-27	The Envelope Layer: An Underestimated Surface of LLM Agent Defense
• Studies how historical experience changes the defensive boundary of an agent’s message envelope. • Factorial tests vary label strength, envelope authenticity and success/failure memories under A2A-TIBA. • Success memories raise semantic breach and failure memories lower it; mixed histories often return to baseline, within the tested attack and frontends.
2026-10-27	ETAMem: Decoupling and Aligning Heterogeneous Memory for Long-Term Conversational Agents
• ETAMem separates heterogeneous dialogue information into complementary memory views. • Atomic units populate event, affect and time hierarchies, with incremental maintenance and entropy-based retrieval budgets. • Reports relative accuracy gains of 12.46%/7.96% on LoCoMo and LongMemEval-S, with complementary-view analyses.
2026-10-27	From Interaction to Capability: Task Advancement as State and Supervision for GUI Agents
• TAM turns verified GUI progress into both task state and reusable operations. • Local successes from successful or failed workflows yield input, outcome and verification contracts, retrieved with interface checks and revalidated by observations. • On MobileWorld, state memory raises success from 12.0% to 28.2%; full TAM improves by 17.6% relative to the strongest tested baseline.
2026-10-27	Multi-State Linear Attention with Generalized Second-Order Routing
• MERLIN improves content-aware routing across segmented memory states. • First and second key moments estimate query-alignment mean and variance for aggregating state readouts. • Long-context QA and LongBench improve on tested Gated DeltaNet and Mamba-2 variants.
2026-10-27	SkillMem: Compact External Skill Memory for Latent Multi-Agent Reasoning
• SkillMem compiles experience-derived textual skills into compact external latent memories. • Paired successful and failed trajectories yield filtered, deduplicated skills; behavioral distillation and role-conditioned readers enable later reuse. • Six benchmarks show reasoning and storage benefits at about 9.6 KB per skill; libraries are built offline, without demonstrated online lifecycle governance.
2026-10-26	Let the Gate Decide: Per-Lookup Precision for Quantized Conditional Memory in LLMs
• Finds that access-frequency precision allocation can harm long-tail facts in conditional memory. • GateRQ stores residual one-bit code ladders and uses the retrieval gate to select lookup depth. • Across two models, 2.10 bits per lookup loses about 0.2 PopQA points; quality and efficiency depend on the tested settings.
2026-10-26	Stratum: Repository Memory for Coding Agents Beyond the Context Window
• Proposes persistent, tiered repository memory that coding agents can inherit across sessions and models. • Indexed, page-safe tools return relevant knowledge with explicit coverage and cost information. • Across eight real bugs, root-cause localization succeeds in 21/24 trials versus 6/24; this is a small initial evaluation.
2026-10-26	Agent Memory as Recommendation: Recommender-System Design for Long-Term Memory of LLM Agents
• Frames agent memory selection as budgeted slate recommendation and introduces AgentMemRec. • Uses candidate generation, learned ranking, and slate construction without LLM calls for selection, analyzing stage-wise evidence completeness. • Reports strong LongMemEval and LoCoMo results, matching a reranker using 64 calls per query at 3.2× lower serving cost.
2026-10-26	Harness-Model Collaboration for Long-Horizon Continual Learning
• Introduces HaMo to combine fast external experience accumulation with slower model adaptation. • Distills trajectories into retrievable strategy cards, trains their interpretation and execution, and selects complementary updates using retrieval coverage. • Outperforms harness-only and model-only adaptation under within-domain and cross-domain shifts, supporting complementary retention and transfer.
2026-10-26	Streaming Recurrent Memory
• Presents Streaming Recurrent Memory for retaining useful state without storing raw experience. • Uses structured storage and learned readout for content/time recall, elapsed-step tracking, and order-sensitive composition with one-step BPTT. • Shows gains in associative recall, online language modeling, and delayed-reward T-Maze; the evidence does not establish large-model deployment performance.
2026-10-26	Adaptive Memory Index Evolution for Agents
• Introduces AMIE, persistent adaptation of memory access state independently of stored content. • Accumulates bounded state from deployment retrievals and teacher keep/suppress judgments to reweight later access without rewriting memories. • Improves all 15 downstream host–benchmark comparisons; simple frequency policies can be competitive, and suppressing useful memories causes failures.
2026-10-26	MemVLN: Streaming Episodic Memory for Vision-Language Navigation
• Introduces MemVLN, bounded persistent episodic memory for streaming vision-language navigation. • Combines compressed perceptual features with timestamps and relative poses, using cross-attention readout and masked trajectory reconstruction. • Leads many reported VLN-CE metrics; physical testing covers only six held-out episodes, with reported success rising from 17% to 40%, then 47% after post-training.
2026-10-26	RetroNav: Scene-Grounded Retrospection and Retry for Self-Evolving Lifelong Navigation
• Introduces RetroNav, linking accumulated scene evidence with navigation experience for reflection and retries. • Separates observations from task interpretations, binds them through provenance, and uses independent evaluation to trigger evidence-grounded retry briefs. • Reports strong results on five navigation benchmarks; persistence follows scene/task-sequence boundaries, and scene memory differs from same-task retry guidance.
2026-10-26	CRIM: Composable and Reusable Item Memory for Efficient Personalized Retrieval
• Introduces CRIM, reusable continuous item memories separating semantic encoding from request-specific history modeling. • Compresses items with learned summary tokens, composes memories in interaction order, and trains with sampled re-encoding and periodic refreshes. • Reports strong retrieval quality with lower inference cost on MovieLens-1M, Qilin, and LazSeq; evaluation concerns personalized retrieval.
2026-10-25	EMMA: Structured Episodic Memory for Multi-Agent Workflow Synthesis
• EMMA organizes durable execution experience across controller, agent and procedural layers. • Reflection assigns each failure to one owner, applies a guarded edit and exercises it in a rerun; memory is frozen for evaluation. • Three benchmarks and five backbones show gains over tested alternatives with fewer training executions; this does not establish continuous deployment-time updating.
2026-10-25	Holding a Steady Course: Knowledge Evolution from Expertise and Lessons for Continual Agent Learning
• Introduces KEEL for converting execution experience into maintainable procedural knowledge. • Uses verifier evidence to form successful skills and corrective lessons, supporting, contradicting, replacing, or retiring claims as evidence evolves. • Improves SkillFlow success by 4.2–21.7 points over no-knowledge baselines; ablations show complementary skills and lessons within the evaluated tasks.
2026-10-25	Memory Tax: Dissecting the Net Value of Agent Memory
• Introduces Memory Tax to evaluate dedicated memory systems by net value over retrieval from raw history. • Separates benefits from losses caused by reorganizing and accessing information, examining obscured evidence and evaluation effects. • Finds several memory systems below naive RAG across evaluated settings; these negative results do not imply that all memory mechanisms are ineffective.
2026-10-25	CoLink: Global Compression and Knowledge Linking for Multi-Turn Dialogue Modeling
• Introduces CoLink to combine dialogue-history linking with global compression of retained information. • Builds topic graphs linked to external knowledge, retrieves with semantic and topological signals, and compresses into competing latent slots. • Reports improvements on public and MMD datasets; evidence concerns domain-specific multi-turn dialogue rather than unrestricted long-term memory.
2026-10-25	EvolveMAS: A Unified Self-Evolution Framework for Decentralized Multi-Agent Systems
• Introduces EvolveMAS with dedicated MemoryAgents for decentralized memory governance and cross-episode adaptation. • Updates isolated structured memory within episodes and distills outcomes into reusable partner profiles and interaction strategies between episodes. • Reports a normalized average score of 77.2 across six configurations on five benchmarks; ablations support both memory processes without weight updates.
2026-10-25	CaliMem: Search-Efficient Data Agent via Calibrated Memory
• Introduces CaliMem to reduce overgeneralization in persistent memories used by data-curation agents. • Predicts policy outcomes before execution and uses observed errors to retain, revise, and scope experience claims during long-horizon search. • Reports gains on biological and LLM post-training tasks, reaching up to 19.4% over full-data training in particular settings.
2026-10-25	The Price of Scale in Proactive Memory Recall
• Analyzes how growing memory stores amplify false interruptions in proactive recall. • Models triggering as multiple testing, derives store-size effects, and proposes a conformal gate calibrated with other user cues. • Streaming replays confirm threshold failure; the gate holds false alarms near 5% on LongMemEval, with guarantees conditional on exchangeability.
2026-10-24	ReSight: Task-Driven Self-Improving Memory for Embodied Agents
• Introduces ReSight, whose embodied memory representations and access capabilities expand with task demands. • Retains observations and converts exploration traces into reusable query tools, memory builders, and skills through Explore–Distill–Mount. • With execution model and online budget fixed, FindingDory success rises from 38.3% to 63.8%; long-tail queries and repeated long-term expansion remain open.
2026-10-24	CoEvolve: Counterfactual-Guided Co-Evolution of Memory and Collaboration in Multi-Agent Systems
• Introduces CoEvolve to align memory evolution with changes in multi-agent collaboration. • Links reusable collaboration skills to supporting memories and uses original/counterfactual trajectories to coordinate library and executor-memory updates. • Leads the evaluated baselines on ALFWorld and WebShop; ablations support alignment and counterfactual signals, with evidence limited to these environments.
2026-10-23	Can Test-Time Memory Evolution Outperform Test-Time Training for Scientific Discovery?
• Introduces test-time memory evolution as an alternative to online weight updates for scientific discovery. • Distills evaluated changes and failures into guidelines and strategies, consolidates entries, refreshes counts, removes stale guidance, and reuses it in later search. • Outperforms the compared TTT method on five mathematical tasks while reducing FLOPs by 82.8–84.6%; evidence is limited to these settings.
2026-10-23	Lifelong Context Adaptation through Bidirectional Memory-Skill Learning
• Introduces MemCryst for bidirectional learning between episodic memory and reusable skills under bounded context. • Recurring episodes create or revise skills; skill-guided interactions add memories, and outcomes adjust later retrieval influence without rewriting records. • Leads seven of ten reported settings; ALFWorld joint serving reaches 63.17% versus 59.83% for memory alone and 50.50% for skills alone.
2026-10-23	LogLook: Execution-Grounded Memory for Frontier-Model Robot Planners
• Introduces LogLook, separating episode record keeping from a VLM planner’s choice of visual evidence. • Preserves execution logs and prior notes while the model requests likely-changed frames and selectively inspects demonstrations. • Matches note-only memory success with 43% fewer tokens under the benchmark motion planner; direct-control success reaches 93.5% on RoboMME.
2026-10-23	CauProphet: Teaching Agents to Develop Causal Cognition like Humans
• Introduces CauProphet to develop persistent causal knowledge from experience for later agent decisions. • Combines long-term causal cognition with current evidence into task beliefs and updates both as observations arrive. • Reports transfer gains across finance and other tasks, including prediction accuracy rising from 21.3% to 28.6%; this does not establish universal causal correctness.
2026-10-23	ActionRAG: Evidence-Refining Graph Diffusion for Long-Term Memory in LLM Agents
• Introduces ActionRAG to combine cautious memory indexing with query-time reassessment of graph evidence. • Reconciles types and entities, then refines triple scores through bounded diffusion with seed protection and hub penalties. • Reaches 55.1 average accuracy over eight deterministic MemoryAgentBench columns, 5.4 points above the strongest retrieval-based comparator; costs follow reported accounting.
2026-10-23	Gaussian Episodic Memory for Long-Term Spatiotemporal Reasoning
• Introduces GEMS, persistent Gaussian scene memory for language-guided spatiotemporal reasoning and on-demand view inspection. • Stores localized scene changes and lets an LLM gather evidence through retrieval, rendering, and cross-time comparison tools. • Evaluates 473 questions across three real and twelve synthetic environments; real-scene pose errors and rendering artifacts limit otherwise strong question-answering results.
2026-10-23	FlexA-Mem: Budget-Adaptive Agentic Memory for LLM Agents
• Introduces FlexA-Mem, one history store serving both task-progress memory and exact-fact recall. • Writes history once and renders a compact timeline plus relevant verbatim details fitted to each call’s remaining budget. • Improves tight-budget AppWorld performance and matches a dedicated keyword retriever on LongMemEval/LoCoMo; ablations reveal different information needs.
2026-10-22	Phronix: The Capability-Centric Learning Paradigm for Agents
• Introduces Phronix for learning persistent reusable capabilities from an agent’s own executions. • Maintains a layered capability store through prediction, validation, and correction, with write-time compilation, deduplication, retrieval, and failure-driven research. • Reports 93–95% first-attempt success on evaluated recurring tasks and lower warm-start cost; conclusions are bounded by its cold/warm protocol and task sample.
2026-10-22	SPECMEM: Accelerating Agentic Tool Calling via Live Memory Management
• Introduces SpecMem, a live cross-session tool-call memory supplying reusable drafts for speculative decoding. • Writes verified calls into bounded per-user stores with LRU eviction and delays irreversible calls until verification. • Reports 27% higher mean accepted tokens and up to 1.74× speedup over frozen stores; isolation and side-effect gating address distinct safety failures.
2026-10-22	DraMaze: A Multi-Agent Pipeline with Future-Aware Memory for Diverse Microdrama Script Generation
• Presents DraMaze, which organizes narrative memory around future plot needs to support coherent and diverse long-form scripts. • Combines trope-memory-guided evolutionary search with a backward dependency tree for retaining and retrieving historical narrative evidence. • Improves results across three LLMs and reduces causal-pattern repetition by 32.1% for 60-episode scripts within the reported evaluations.
2026-10-22	Governance-State Corruption in Long-Term Memory for LLM Agents: A Cross-Model Stress Test
• Studies corruption of governance information in persistent memories. • Separates offline projection checks from downstream effects on a selected evaluation panel. • Strong projection scores coexist with large answer-quality losses; the panel is not a population-wide safety estimate.
2026-10-22	Is Complex Agentic Memory Always Needed? Naive RAG Is Stronger Than We Thought
• Preserves raw conversation history for memory retrieval. • Compares source-faithful evidence access with extracted or compressed memory. • Leads six of seven tested benchmark–model settings by 0.13–5.53; structured memory may help elsewhere.
2026-10-22	Retrieval-Augmented Online Textual Memory for Episodic Memory Video Question Answering
• Retrieves online video memory at a finer level than the evidence returned. • Fine-grained matching selects parent captions that retain useful context. • Reports 56.4 on closed-set Ego4D QA; increasing coverage can still reduce answer quality.
2026-10-22	The Value of Memory in Language Agents
• Defines memory value through downstream decision loss. • Recovery frontiers and counterexamples expose limits of task-agnostic memory scores. • Analyzes conversational and updating tasks; conclusions favor task-dependent evaluation rather than one universal score.
2026-10-22	Muffin - A Shared Memory Buffer for Asynchronous AI Agents
• Shares reusable intermediate findings across asynchronous LLM agents. • Shared memory coordinates collaboration while reasoning traces remain private. • Leads 17 of 20 tested settings; reports 19.2% accuracy gain and 77.4% token reduction against debate.
2026-10-22	When Is a Memory Safe Enough to Act On? Action-Conditioned Verification for Long-Lived Agents
• Verifies agent memory according to the pending action. • Value-of-information stopping balances additional checks against safe continuation. • Cuts ToolSandbox cost by 33.9% with 100% measured safe-proceed recall; transfer reduces cost with a small quality loss.
2026-10-21	SpatialBrain: Spatial Intelligence Growing from Experience
• Retains spatial reasoning experience around a frozen vision-language model. • Block-separated updates consolidate process records without transferring answers or scene coordinates. • Improves four of five benchmarks and ties one; process feedback measures reliability and cost rather than answer correctness.
2026-10-21	Plan First, Act Later: Streaming Reasoning for Executable Memory in Robot Control
• Stores semantic events and subtask progress for VLA agents. • A vision-language writer and two-stage verification supervise memory construction and use. • Scores 56.05 on RoboMME and 60.83 across four real tasks; real-world coverage remains limited.
2026-10-21	Memory-Centric Continual Learning for AI Agents
• Organizes continual language-agent learning around memory assets and policies. • Separates state updates from management-policy adaptation and proposes retention and transfer evaluation. • Provides a conceptual framework rather than an experimentally proven new performance method.
2026-10-21	When Thoughts Become Memory: Unifying In-Context Learning and Fine-Tuning
• Quantifies when reasoning information functions as memory. • An information identity and repair-decoding loss separate contextual from weight-based storage. • A constrained GPT-2 stream-selection study supports the analysis; certification depends on its assumptions.
2026-10-21	PreCoMem: Predictive Cognitive Memory for Self-Evolving Long-Term Dialogue Agents
• Maintains revisable user beliefs rather than appending every observation. • Surprise gating and a waiting area for uncertain evidence control profile updates and corrections. • Evaluated on three benchmarks and DiMoBench; thresholds and confidence estimates remain distribution-dependent.
2026-10-21	Uni-Mem: Learning a Long-Horizon Agentic Memory Manager with Unified Online Policy Optimization
• Jointly learns memory construction and retrieval. • Stage-decoupled optimization and event-wise credit assignment connect writing decisions to later use. • Outperforms tested RL baselines on LoCoMo and LongMemEval under the reported training protocol.
2026-10-21	Compaction Laundering: Provenance Destruction in Context-Compacting Coding Agents
• Tests instruction laundering through memory compaction. • Compares flattened summaries with typed provenance and programmatic source checks. • Observed attacks succeed on 87–100% of flattened cases and none of the typed cases; coverage remains finite.
2026-10-21	Remembering Through Change: Provenance-Preserving Hypergraph Memory for Multimodal Agents
• Organizes long-term memory as typed, provenance-linked hypergraphs. • Agentic search resolves mentions, identities and changing states while inspecting sources. • Reports gains of 8.20–9.98% across three benchmarks under the stated graph and reader protocols.
2026-10-21	MemPlan: Memory-Conditioned PDDL Planning for Partially Observable Text Environments
• Retains both useful planning experience and refuted assumptions. • An LLM bridges language to a fixed PDDL domain while memory guides subsequent plans. • Improves efficiency on three embodied benchmarks; results depend on the symbolic domain and interface.
2026-10-20	What should AI forget? A design and taxonomy for negative space governance in AI memory
• Studies what users want assistants to avoid remembering. • A taxonomy informs policy gates and user overrides for unwanted memory. • Evidence includes 179 incidents and a 15-person design study; it does not establish formal privacy guarantees.
2026-10-20	Memory as Code: Declarative Retrieval over Codified Long-Term Memory
• Represents long-term memory as executable, compositional code. • Declarative queries operate over atomic functions, relations and their compositions. • Reports up to 15.91% relative judge-score gain on two benchmarks; compilation and evaluation assumptions remain.
2026-10-20	EventSync: Event-Driven Memory Updates for History-Dependent Manipulation
• Writes subgoal history only after verified completion. • Visual state and end-effector monitoring jointly gate memory updates. • Reports 64.5 on RoboMME and about a 20-point gain, plus four real tasks; monitoring remains a dependency.
2026-10-20	Disclosure Begets Disclosure: A Self-Reinforcing Privacy Loop in LLM Agents with Persistent Memory
• Studies a feedback loop between assistant memory and user disclosure. • Simulated users reveal sensitive attributes in response to memory-based presuppositions. • A response guard reduces disclosure across four models; simulated behavior limits real-user conclusions.
2026-10-20	Combining Long-Term Memory and LLMs Comes at a High Cost: From Prompt Injection to Persistent Compromise
• Studies injection into persistent assistant memory services. • Tests cross-session influence and organization-level propagation across three services. • Attacks collectively reach multiple objectives; success does not imply every attack works on every platform.
2026-10-20	inr2action: Encoding Robot Episodes as Neural Implicit Representations for Recall and Action
• Encodes video experience into compact implicit memory parameters. • A meta-initialized representation supports language QA and action decoding through per-video weight changes. • Reports R² of 0.51 and robustness gains in the tested tasks; general embodied memory remains unproven.
2026-10-20	PerMIT: Securing Persistent Memory for LLM Agents via Machine-Checked Integrity Theorems and Scalable Community Trust
• Constrains which persistent memories may acquire authority. • A finite-interface model admits user or gate-confirmed records and formalizes integrity in Lean4. • No attacks succeed in the tested TrojanHippo setting; proofs and experiments rely on stated interface assumptions.
2026-10-19	Cognitive Maps in Large Language Models
• Probes associative memory underlying language-model cognitive maps. • Entity–relation tasks and interventions test sparse key–value bindings. • Larger models exhibit compositional false memories; synthetic relations do not establish a global real-world map.
2026-10-19	Weighted State Souping for Efficient Long-Context Retrieval
• Retrieves information by composing cached state-space chunk memories. • Query-dependent weights combine states for multi-hop and tracking tasks. • Theory and controlled tests support the mechanism without establishing universal long-context reliability.
2026-10-19	Whose Credential Is It? How Owner Attribution in Agent MemoryShapes Cross-User Account Misuse
• Tests whether credential ownership survives in agent memory. • Separates exposure from misuse and evaluates explicit owner binding under attribution stress. • Misuse falls from 92 to one of 305 cases, with a small task cost; exposure alone is not misuse.
2026-10-19	CoMem: Learning to Anticipate Failures from Retrieved Experience
• Uses remembered experience to anticipate failures before acting. • Retrieved successes and failures generate task-conditioned failure hypotheses during training and inference. • Improves five tasks with two backbones; transfer depends on experience coverage and retrieval quality.
2026-10-19	ACE-Trader: Adversarial Co-Evolution for Generator-Critic Trading Systems
• Updates role-specific trading experience through execution feedback. • Generator and critic memories select recent outcomes and derive reusable lessons from shared records. • Compared with 22 baselines in market backtests; evidence does not establish future trading performance.
2026-10-19	SPEM: Structure-Aware Procedural Memory for Efficient Skill Use in LLM Agents
• Turns successful skill executions into reusable procedural DAG memory. • Backward slicing retains dependencies; reuse reinforces templates while failures update reliability. • Reaches 87.1% ALFWorld success and 38.38 SkillsBench reward with separated evaluation trajectories; retrieval and execution remain dependencies.
2026-10-19	Learning What to Remember: Resource-Efficient Self-Improvement for Lifelong Agents
• Learns which experiences are worth retaining and retrieving. • Sparse execution feedback trains query-conditioned memory value under a resource budget. • Leads tested success–token trade-offs on three LifelongAgentBench settings; results depend on budgets and feedback.
2026-10-19	ScopeMem: Latent Memory That Learns Where to Revise Frozen GUI Agents
• Learns when latent memory should revise a frozen GUI agent. • Separates historical information from control over whether an initial action may change. • Three-benchmark diagnostics attribute some gains to revision scope rather than additional recalled information.
2026-10-19	ReInsVLN: Revisiting Historical Evidence for Training-Free Vision-and-Language Navigation
• Revisits retained visual evidence during language-guided navigation. • A training-free policy compresses history and schedules instruction and evidence rereading. • Two backbones gain 8.48/3.37 points on R2R; results are specific to the evaluated navigation configurations.
2026-10-19	BetaMem: Learning to Remember a Changing World for Embodied Agents
• Learns embodied memory writes under explicit verification contracts. • A layered store maintains support relations using supervised and feedback-based training. • Reports improvements across several embodied and dialogue benchmarks; task-specific protocols prevent a single pooled interpretation.
2026-10-19	M-Plan: Comprehensive Planning For LLM Agents with Memory Arrangement
• Connects short-term planning with reusable long-term experience. • An optimizer refines successful milestones and contrasts failures with successes before writing records back. • Improves success and efficiency on ALFWorld, Webshop and PDDL; evidence is limited to the tested tasks and models.
2026-10-19	LVMem: Image-Augmented Memory for Long-Horizon Question Answering
• Augments textual long-horizon memory with accessible images. • Object cues and optical flow select frames for adaptive vision-language inspection. • Improves OCNaVQA with two backbones; results concern the tested QA setting rather than general 3D reasoning.
2026-10-19	AutoRedClaw: Automated Red-Teaming for Claw-Style Agents
• Automates red-teaming of persistent agent state. • Generated poisoning attacks are validated across harnesses and language models. • Most tested settings exceed 50% attack success across 330 cases; generation and selection shape the results.
2026-10-19	REM: What to Store, When to Recall, What to Forget in Experience Memory for Coding Agents
• Coordinates storage, recall and eviction of coding experience. • Compact documents index implementations that are fetched on difficulty and evicted after one exposure. • On SWE-Ctxt-Lite, resolution rises 24.9→27.6 and cost falls 17%; gains are protocol-specific.
2026-10-19	ProMemAgent: Long-Horizon Proactive Assistance with Context-Aware Prospective Memory from Sensory Perceptions
• Connects prospective intentions to sensory and episodic cues. • Context-aware retrieval determines when remembered future actions become relevant. • Evaluated on 1,000 CPMBench cases in eight categories; real longitudinal deployment remains unestablished.
2026-10-19	More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents
• Separates useful experience from error-prone persistent guidance. • Anti-persistence retains raw trajectories and outcomes while discarding synthesized instructions between tasks. • Improves the tested success–cost trade-off; the result does not show that all persistent abstraction is harmful.
2026-10-18	MemPrice: Online Memory Allocation via Adaptive Pricing for Long-Horizon Agents
• Treats agent-memory retention and granularity as online allocation. • An adaptive context price trades estimated retention value against footprint across memory refreshes. • Evaluated on four benchmarks; regret guarantees use a same-state comparator and explicit regularity/error assumptions.
2026-10-18	Deep Fusion Memory: A Chunk-Level Parametric Memory for Large Language Models
• Fuses chunk-level parametric memory into a frozen language model. • A causal predictor and gated cross-attention learn from retrieved memory embeddings. • Reports fewer queries and retrieved items with QA and domain gains; efficiency ratios depend on the stated baseline.
2026-10-18	Memory as Belief: Bayesian Evolution of Memory for Self-Evolving LLM Agents
• Treats persistent memories as revisable beliefs. • Bayesian-style updates calibrate confidence and usefulness from execution outcomes. • Reports gains on reasoning and coding tasks; feedback quality and task distribution remain dependencies.
2026-10-18	Memories That Outlive Their Reasons: Dependency-Tethered Memory for Tool-Using LLM Agents
• Invalidates memories when their supporting evidence no longer holds. • Version-specific dependencies restrict eligibility and focus rechecking on affected records. • Unsafe actions fall from 62% to 0.3% on Fallout with fewer checks; weaker writers reduce the benefit.
2026-10-18	LAMB: Latent Multi-Hypothesis Beliefs for Long-Horizon Large Language Agents
• Maintains multiple latent beliefs for long-horizon language agents. • Shared evidence and hypothesis deltas are recurrently updated using observed outcomes. • Reports gains on ALFWorld, ScienceWorld and Crafter; evidence concerns the tested partially observable tasks.
2026-10-18	Memory, Know Your PLACE: Governing Lifecycle State and Query-Time Eligibility in Persistent-Memory Agents
• Separates memory lifecycle certification from query-time eligibility. • Factorial tests compare write-side and read-side controls around persistent memory. • Read gating eliminates measured leakage in a state-correct setting; the result is conditional rather than universal.
2026-10-18	Improving the Expressive Power of Hybrid Sequence Models via Multi-Hop Retrieval
• Expands hybrid sequence-model memory through multi-hop retrieval. • Shared key–value access composes functions across a bounded number of retrieval rounds. • Complexity separations and synthetic tests support expressivity claims, not broad natural-language performance.
2026-10-18	Atoms, Not Knowledge: What Production-Scale Addressed-Memory Writes Deposit
• Audits what large-scale addressed-memory writes actually retain. • Separates factual coverage, retrieval and compositional use as write volume grows. • Coverage drops from 98% to 12.7% between 50 and 1,000 facts; interference and capacity explanations remain unresolved.
2026-10-18	Let Them Sleep: Recursive Self-Pruning Memory for Long-Horizon LLM Agents
• Consolidates and prunes experience during a sleep-like phase. • Failure clusters become rules while episodes are deleted and reusable rules are promoted or decayed. • Reports gains in three settings with fewer entries; task and simulation choices limit generalization.
2026-10-17	Parametric Session States Recover Retrieval Intent without Replaying History
• Introduces PSS, reusable parametric session states that recover retrieval intent without replaying interaction history. • Compiles history and clue supervision into session LoRA adapters, generates search clues, then unloads adapters before evidence-grounded answering. • Improves retrieval and answering on 262 transformed public sessions; offline write cost grows with history, and evaluation supplies session associations and supervision.
2026-10-17	PACE-Mem: Prefix-Aware Complementary Evidence for Multimodal Agents
• Selects complementary evidence and retrieval depth for multimodal memory. • Query pooling and MMR build a sequence; normalized score declines determine its prefix length. • Gains 1.6% average accuracy over Adaptive-k across three benchmarks and two clip-memory frameworks.
2026-10-17	Memory Internalization via Test-Time Training for Long-Video Understanding
• Internalizes long-video evidence into compact per-video parameters. • Causal tracing selects adaptation layers; visual distillation preserves clip-induced representation changes. • Improves video-free QA on three benchmarks with fewer synthetic samples; results concern per-video adaptation.
2026-10-17	Hippo: Training-Free Agent Self-Evolution via Hippocampal Replay
• Consolidates persistent agent experience through comparative replay. • Compares attempts, merges overlapping lessons and revises conflicting guidance with applicability conditions. • Improves three coding/web benchmarks with fewer records; unconsolidated accumulation can eventually underperform no memory.
2026-10-17	Write Once, Query Many: Evolving Compact Reusable Memory for Long Videos
• Builds reusable multimodal video memory before future questions are known. • Separates source-fidelity feedback for writing from QA evidence feedback for reading. • Evolution improves average accuracy by 7.2%; storage falls about 99.86% versus raw video in tested QA settings.
2026-10-17	MemRevise: A Self-Revising Memory for Conversational Search
• Introduces revisable topic memory for conversational search. • A distilled 3B writer records add, abandon and reactivate operations while preserving inactive facts. • On 4,429 benchmark questions, mean MRR@10 reaches 0.701; CORAL ties strong rewriters statistically.
2026-10-17	How Intervention History Changes AI Agent Sabotage on Subsequent Tasks
• Tests how retained intervention history changes subsequent agent sabotage. • Pairs fresh tasks with no history, replacement notices, or trusted replacements; also condenses history. • One IaC model falls from 77% to 12%/38% joint success; other effects are insignificant and context-dependent.
2026-10-16	GRAD: Graph-Structured Antibody Defense for LLM Agents
• Aggregates distributed evidence of memory poisoning and related agent attacks. • Typed execution graphs support risk scoring; frozen-snapshot experts preserve and extend detection capability. • Reports AUROC 97.19/77.59/86.60 on three suites; results cover multiple attack types, not memory poisoning alone.
2026-10-16	SAFE-Mem: Secure Agentic Memory with Fine-Grained Authority and Governance-Aware Retrieval
• Governs which persistent-memory claims may influence LLM agents. • Kind, authority and scope labels constrain retrieval over a provenance-linked memory graph. • Reports at least 77% lower attack success than the strongest utility-preserving defense in tested settings, not a universal guarantee.
2026-10-16	From Memory Hits to Answerable Evidence: Query-Conditioned Memory Compilation for Long-Term Agent Memory
• Separates memory addressing from answer-ready evidence compilation. • Stable source packets let a query-conditioned evidence plane allocate budget across complementary claims. • Reports 80.52% judge accuracy on LoCoMo and gains on LongMemEval-S; results depend on the evaluation protocol.
2026-10-16	Darwin-Claw: Environment-Adaptive Experience Consolidation for Recursively Self-Evolving Agents
• Consolidates LLM interaction experience into durable memory and skill patches. • A bounded lifecycle scores context relevance, outcomes and efficiency before promotion, replacement or eviction. • Reports gains on WildClawBench and iterative optimization; LLM attribution is not a rigorous causal guarantee.
2026-10-16	MARS: Memory-Aware Routing for Session-Level LLM Assistants
• Couples memory-view construction with assistant model selection. • A shared query/topic state controls both a fixed memory policy and an online router. • Retains 98.2% of full-context QA at 14.5% API cost on RealMem; trails the best static system on MultiChallenge overall.
2026-10-16	RL-Guided Temporal Localization for Dual-Channel Retrieval in Long-Horizon Agent Memory
• Targets temporal mismatch in long-horizon agent-memory retrieval. • An RL-trained 1.7B locator predicts evidence windows for dual temporal/semantic retrieval and reranking. • Reaches 61.52% on LifeBench and 78.96% on LoCoMo under the reported protocols.
2026-10-16	GLOBE: Evolving Structured Agentic Memory for E-commerce Product Optimization in Generative Search
• Maintains an evolving procedural handbook for product-optimization agents. • Execution feedback proposes field-scoped behaviors; moderation removes redundancy and deactivates harmful rules. • Structured, moderated memory improves tested outcomes and transfers to unseen product categories.
2026-10-16	Learning Personalized Memory Gating from Delayed One-Sided Feedback
• Learns user-specific fact-retention gates for finite agent memory. • Warm-Start UCB combines offline labels with delayed feedback from facts later used in responses. • Outperforms tested baselines across budgets; regret bounds assume limited delays and retrieval use is only a value proxy.
2026-10-16	Cache2LoRA: Understanding and Improving Context Internalization
• Diagnoses and improves factual preservation in context-to-LoRA memory. • Query-conditioned cache aggregation generates updates; multi-style supervision separates preservation from response adaptation. • Across 17 tasks macro F1 reaches 0.663 versus 0.671 for full context; matched-size interference tests favor LoRA retention.
2026-10-16	A Conditional No-Hair Theorem for Transformers: How Language Models Forget Their Prompt
• Analyzes how transformer hidden states forget their prompt. • Separates direct attention dilution from relay through continuation tokens and tests logit-bias interventions. • The decay bound requires bounded logits and nondegenerate pre-norm states; generated text can still carry prompt information.
2026-10-16	Reasoning with Context Management and External Memory
• Adds external memory to reasoning-context management. • Preserves intermediate information outside active context to reuse repeated computations. • Reports speed and task gains on a few mathematics problems with small models and GPT-2; scale and breadth are limited.
2026-10-15	ProV-Mem: Proactive Memory for Streaming Video Question Answering
• Coordinates proactive video responses through distinct memory roles. • Trigger state schedules queries and episodic memory filters duplicate emissions, alongside a fixed instruction prior. • Leads CAUC in four tested scenarios; the approximation bound assumes conditional sufficiency.
2026-10-15	When Plans Change, Remember: Replanning-Guided Visual Memory for Vision-Language-Action Models
• Uses plan revisions to control task-event memory writes in VLA policies. • Action-prediction disagreement triggers bounded keyframe updates with a robot-masked redundancy filter. • Detectors reach ROC-AUC 0.981–0.997 and tested retention reaches 100%; retention is not task success or safety.
2026-10-15	From Relevant Memories to Sufficient Evidence: Rethinking Memory Utilization in LLM Agents
• Turns relevant memories into collectively sufficient evidence. • Multi-granularity discovery, source-linked ViewCards and selective gap recovery operate over existing stores. • Reaches 91.43% LoCoMo judge accuracy with the specified reader, 7.6 points above MemR3 under shared storage.
2026-10-15	Not All Graphs Are Maps: A Task-Agnostic Exploration Framework for Mobile GUI Navigation
• Studies reusable destination memory for mobile GUI navigation. • Task-agnostic exploration stores functions and routes; controlled graph variants expose coverage and cleanliness effects. • Navigation improves to 58.3%/71.4% for two backbones; the protocol stops before consequential destination actions.
2026-10-15	Spatial Memory by Agents, for Agents
• Automates spatial-memory construction and retrieval design with coding agents. • A fixed query agent evaluates revisions; implementations freeze before held-out histories and questions. • Three searches average 44.23 versus 29.32 across five tracks, with grounding failures and task-dependent cost trade-offs.
2026-10-15	MemComposer: Learning to Assemble Memories for LLM Agents via Reinforcement Learning
• MemComposer learns which experiences to use, how many to select and how to order them for a query. • ExPO/GRPO trains assembly using success feedback from complete agent interactions. • Beats each executor’s strongest baseline in three τ²-bench domains, using 16.9% and 4.9% fewer memory tokens; transfer evidence covers the tested domains, executors and banks.
2026-10-15	Compression Does Not End at Selection: Post-Selection Positional Geometry for Sparse KV Caches
• RoPERelay studies how post-selection address geometry affects access to retained memory evidence. • With content selection and capacity fixed, it compacts order-preserving RoPE coordinates, relocates key phases and synchronizes later queries. • Improves 32 of 36 model/selector/task comparisons; one LongMemEval-S configuration rises from 23.6% to 35.4% with roughly 1%–3% request-time overhead.
2026-10-15	All-in-One Memory: Process-Rewarded Reinforcement Learning for Memory Agents
• AiO-Mem jointly optimizes memory search, writing and stopping. • Histories become traceable events; a separately trained SFT/DPO process reward model supplies action-balanced rewards for PA-DAPO. • Reports gains over static and RL memory agents; the abstract gives no unified numerical effect size, limiting conclusions to tested benchmarks.
2026-10-15	MemScope: Evolving Agent Memory Architectures through Localized Code Repair
• MemScope maintains agent memory architectures through localized code repair rather than broad rewrites from coarse feedback. • It inspects traces, runtime memory and code under provider-identity and memory-compatibility constraints. • GAIA trace-plus-memory feedback scores 86.67% versus 70% for traces alone; unseen ALFWorld rises from 50.75% to 71.64% in evaluated snapshots.
2026-10-15	An Agentic Framework for Proactive Robots with Procedural Assistance Memory
• Procedural Assistance Memory records user routines for high-level VLM decisions about proactive help. • A predictor reads observations and a state machine, while daily reflection revises rules from traces; a VLA executes with frozen weights. • First-fire delay falls from about five seconds to one in a five-day desk study, with one update repairing routine mismatch; evidence remains small-scale.
2026-10-15	Memorilla: Latent Semantic Memory for LLMs
• Memorilla supplies frozen models with latent semantic memory of accumulated documents and interactions. • Documents are encoded once, then retrieved and combined into 16 query-dependent input memory tokens. • Leads five of eight benchmarks; FactKG reaches 84.33% and TextWorld success rises from 18.6% with full history to 24.2%, without universal gains.
2026-10-15	Writability and the Pathway: A Controlled Study of Knowledge Injection into Addressed Memory
• Separates knowledge writing into conditional memory from question transfer and preservation of unrelated behavior. • Compares GPT-2 row/projection updates and anchoring, alongside Qwen gradient writes and DeepSeek zero-order experiments. • GPT-2 arms reproduce trained completions but strict answer scores stay zero; Qwen first-token promotions do not establish full-answer success or universal writability.
2026-10-15	Reliable Knowledge Acquisition for Continual Learning Agents
• RIKA separates experience extraction from memory admission to address unreliable knowledge reuse. • Four knowledge types are admitted through source evidence; inferred claims await support, while contrary evidence triggers revision or forgetting before later reuse. • Leads aggregate scores on six task streams and four backbones by 0.046–0.081 over their strongest baselines; gains vary and do not establish high-stakes safety.
2026-10-15	TEM: Temporal State Tracking Memory for Large Language Model Agents
• TEM preserves fact versions so updated memories can still answer historical questions. • Retrieved versions are source-ordered and historically labeled, with events and original dialogue supplementing structured facts. • Labels improve entity hit rates by 17.14–55.95 points on 420 questions; two backbones beat session-aware RAG on LoCoMo under the stated metrics and evidence budgets.
2026-10-15	Memory as Weights: Internalizing Long-Term History for Streaming Videos
• MaLoW internalizes video history as parametric memory, separating it from current observations in context. • A recurrent HyperNetwork consolidates observations into fixed-size LoRA weights, supported by 381K QA pairs from 56K videos. • Reports 3.9%–11.8% gains on the OVO-Bench backward split with latency overhead; evidence is limited to tested backbones and benchmarks.
2026-10-14	ECHOES: Seeing Beyond the Current Image via Recursive History Compression and Privileged On-Policy Distillation
• Trains GUI agents to preserve visual history in recursive textual memory. • Privileged multi-image supervision guides compression through perception-memory fidelity diagnostics. • Improves four GUI benchmarks and transfers to unseen planning tasks; privileged information is used during training.
2026-10-14	EvoWeave: Composing Reusable Agent Experience through Operation Level Graphs
• EvoWeave stores operation-level experience graphs for composition according to current task dependencies. • Reflection separates evidence from abstractions and validates structure, capabilities and contracts; an initially empty online memory supports later tasks. • AppWorld Normal/Challenge completion averages 51.91/27.99, gains of 6.99/4.99 points over no memory, with retrieval overhead and reuse preconditions remaining relevant.
2026-10-14	STRATA: Self-Maintaining Structured Experience Graphs for Self-Evolving Language Agents
• STRATA stores skills, workflows and meta-frameworks in a persistent graph, targeting memory growth and negative transfer. • Outcome-grounded reflection, operation-aware retrieval, community induction, garbage collection and contrast edges maintain experience across tasks. • Leads average scores across four interactive benchmarks and two backbones; contrast edges raise ScienceWorld from 78.7 to 96.7 within the evaluated system.
2026-10-14	MemCode: Discrete Semantic Representations for Long-Term Agent Memory
• MemCode turns dialogue into self-contained atomic memories, complementing vector retrieval with discrete semantic associations. • Learned codebooks and isomorphic semantic quantization provide co-occurrence retrieval paths, separating online writes from asynchronous consolidation. • Reports the best overall accuracy on LongMemEval-S and LoCoMo with two backbones; construction takes roughly 40–130 seconds per dialogue in tested configurations.
2026-10-14	Beyond Retrieval: Evidence Access and Event Aggregation in Agent Memory
• Combines source-preserving access with event aggregation for agent-history questions that retrieval alone cannot resolve. • Trajectory Memory retains traces and addressable views; TallyMem adds an event ledger, model-generated specifications and deterministic execution. • Qwen3-32B exact match rises from 35.2% to 49.0% on 776 MINTEval questions; AMA-Bench gains only 1.3 points, with event-specification errors limiting benefits.
2026-10-14	StateCraft: Augmenting Video World Models with Abstract-State Reasoning
• StateCraft uses a language model to maintain editable semantic state, preserving out-of-view changes for video generation. • The VLM reads nodes, actions and history, predicts new/same/revisited transitions and graph diffs, and updates persistent entity records. • File-manager tests assess revisits and edits with first-visit initialization; 3D remains a proof of concept and fine-tuning incurs out-of-domain generalization costs.
2026-10-14	EGCForest: Structured Agent Memory Organization for Enterprise Group Chats
• EGCForest organizes enterprise-chat memory by workflows, speaker context and artifact lifecycles. • Three forests from immutable logs use placement maps and sparse bridges for relevance- and hierarchy-aware retrieval. • Beats compared RAG and memory baselines on three chat benchmarks, with gains up to 20.85 points over MemOS; the maximum is not an average.
2026-10-14	DeltaStreamer: Towards Human-Like Agentic Streaming Video Understanding
• DeltaStreamer records visual changes before queries arrive to preserve evidence for unknown future questions. • Continuous observation writes memory; a query planner retrieves records and replays selected frames to recover omissions. • Scores 66.07% on OVO-Bench and 59.53% on StreamingBench, above compared methods; active recovery depends on access to historical frames.
2026-10-14	MemDuet: Learning Dual-Process Memory for Long-Context Reasoning
• MemDuet combines fast selection with slower verification and updates for long-document memory. • A small model selects chunk evidence in parallel while a larger model updates in document order through SFT, GRPO and joint training. • Across three evaluated benchmarks it averages 9.74× speedup and 73.91% fewer output tokens than RL-MemAgent, alongside reported accuracy–efficiency trade-offs.
2026-10-14	Reflect After Failure: Brain-Inspired Self-Evolving Agents for Embodied Spatial Intelligence
• HiMeco forms strategy and procedural memory from failures in perception, exploration and planning. • It compares attempts and abstracts skills, using evidence contracts and execution validation for retention and reuse, with updates between rounds. • Reports leading ESI-Bench performance and a 3.94-point EmbSpatial-Bench gain without further evolution, within evaluated spatial tasks.
2026-10-14	Cortex: Consolidating Context into Parametric Memory
• Cortex consolidates context-induced gains while preserving unrelated capabilities after context removal. • Prospective queries distill outputs into low-rank or structural feed-forward updates, evaluated under single, batch and sequential consolidation. • Reaches the empirical frontier in most settings; 50 factual edits on a 14B model recover 88.9% of gains with 99.8% non-target retention, not a universal rate.
2026-10-14	Continuum Attention
• Continuum Attention replaces discrete key-value addressing with integration over a continuous memory curve. • Queries induce exponential relevance densities; closed-form moments yield normalized reads and gradients while retaining quadratic complexity. • Validation loss improves by 0.0200 and 0.0480 nats/byte over softmax and focal attention; an extra slice supports gains, with large-scale generalization unresolved.
2026-10-14	Thinking in Graphs: Neo4j-Backed Exploit-Chain Reasoning for LLM Penetration-Testing Agents
• ChainGraph stores security-analysis findings as graph memory for dependency-chain reasoning and provenance. • MCP exposes typed Neo4j nodes and pre/post-condition edges, returning decision-specific subgraphs alongside separate documentary records. • A single synthetic Active Directory example changes chain severity and reduces a token-proxy payload by about 98%; these are illustrative, not benchmark-wide results.
2026-10-13	Many Steps into One: Reasoning Contraction and the Limits of Safe Memory Reuse for LLM Agents
• Analyzes stale answers when reused memories replace multi-step reasoning after facts change. • Uses likelihood-ratio monitoring, utility certification against live reasoning, certificate restarts and invalidation, with conditional bounds on safe reuse. • Across five LLMs, conventional counters serve more stale answers in edited-fact tests; guarantees depend on verified outcomes and change-detection assumptions.
2026-10-13	MemAct: Streaming Spatial Understanding with Memory and Active Exploration
• MemAct unifies streaming spatial memory and active evidence acquisition in one VLM loop. • Incrementally writes structured history, retrieves records and past frames, and moves or turns to gather evidence before answering or continuing exploration. • About one million training examples support evaluation on 20,000 main QA pairs plus 2,000 timing questions, with reported gains over streaming baselines.
2026-10-13	Keep Your Receipts: Trustworthy Cleanup of Agent-Built Knowledge Graphs
• PolyGraph maintains agent knowledge-graph memory after source retraction or revocation. • Records evidence bundles when creating merges, edges and summaries, then audits receipts through re-derivation after removing cited evidence. • Fine-grained receipts outperform flattened sources; controlled tests reduce contaminated queries to 0.4%/0.8%, while model-reported provenance remains fallible.
2026-10-13	Explore with Memory, Learn without It: In-Context Exploration and Self-Distillation for LLM Reasoning
• LEAF improves exploration with dynamic problem-specific memory. • Stores the latest response, reward and feedback after each verifier call, reads it before the next rollout, and separates policy updates through self-distillation. • Across four benchmarks and two models, exploration is about four times faster and unseen-test accuracy gains 1.59 and 1.83 points; memory is chiefly a training-time mechanism.
2026-10-13	Memory Echo in Single-Agent Long-Term Memory: Lifecycle Evaluation and Provenance-Guided Mitigation
• Defines Memory Echo as excess retrieval exposure from repeated same-source history. • MEchoBench tests emergence and persistence; FPA deduplicates source families, gates old records after corrections and constrains answers with authoritative evidence. • Paired replay on 190 questions raises two models’ accuracy from 0.825/0.782 to 0.920/0.916, under the controlled evaluation protocol.
2026-10-13	Learning How to Remember: Memory Modality Selection for GUI Agents
• SwitchMem studies step-dependent memory modality in VLM GUI agents. • Dynamically chooses textual or visual representations of prior experience at each interaction step instead of using a fixed modality. • It outperforms static-modality memory baselines on three GUI benchmarks, supporting conditional choice rather than universal modality superiority.
2026-10-13	Provenance and Authorization in Autobiographical Language Models
• Frames source conflation and unauthorized influence in autobiographical LM memory as security problems. • Uses typed records, authenticated attestations, a separate policy channel and admissible views to analyze integrity and decision-preserving summaries. • Evidence consists of conditional arguments and a frozen-model pilot with 32 synthetic cases across 576 paired conditions, not claims about real people’s psychology.
2026-10-13	Heading Hierarchy Reconstruction for Complex and Long Documents with Level-Wise Multimodal Memory
• Introduces level-wise multimodal memory for long-document heading reconstruction. • Retains recent heading anchors at each level with their visual, textual and layout evidence for subsequent page decisions. • Evaluation uses a new expert-verified benchmark of 250 documents and 11,212 pages plus three existing benchmarks, focusing on structural consistency.
2026-10-13	Reliable Memory from Unstable Growth: Curating Self-Written Agent Playbooks with Randomized Experiments
• RBC uses randomized experiments to estimate the value of self-written agent lessons. • Randomly includes or omits lesson blocks, retains useful blocks, and selects curated, whole or no playbook on fresh practice episodes. • Mean success reaches 66.9% across two memory systems, three models and three benchmarks versus 55.5% with whole playbooks; practice-task representativeness remains important.
2026-10-13	MemoryVLN: From Ego-Aware Memory Representation to Cross-Context Memory Reuse in Vision-Language Navigation
• MemoryVLN centers navigation on spatially grounded memory construction and cross-context reuse. • Combines temporal compression, semantic/geometric/pose representations, hierarchical storage, confidence-controlled writes and Top-K retrieval. • It reports success-rate improvements of about 2% and 6.5% on R2R-CE and RxR-CE, within the tested navigation benchmarks.
2026-10-12	Recall Begets Recall: Recurrent Memory Ranking for Personalized LLMs
• ReMind turns personalized memory ranking into a self-terminating recurrent recall loop. • Each recalled item expands the cue for subsequent ranking, while a sufficiency predictor decides when enough evidence has been collected. • It leads compared memory rankers on public benchmarks, supporting query-adaptive evidence selection; no aggregate numerical gain is given in the abstract.
2026-10-12	Error Correction in Trained TTT Layers
• Analyzes how prediction-derived terms contribute to trained TTT fast-weight memory updates. • Derives update structure and uses interventions preserving the normalization Jacobian and matching update norms to isolate behavioral effects. • On 100,303 literary-text tokens, removing the term raises TTT-Linear perplexity about 2.8–2.9%, while freezing all updates raises it 4.2–4.5-fold; scope is limited.
2026-10-12	Purpose IR: Compiling User Purpose for Privacy-Preserving Agent Workflow State
• Purpose IR separates evidence access from authorization to persist information in reusable agent state. • Compiles user purpose into typed semantic dataflow and releases only validated outputs, with privacy arguments under explicit execution contracts. • Across 1,200 cases it reports 99.6% privacy and 83.3% privacy-compliant success; safe-action rate is 77.3% on 493 PrivacyLens trajectories, without unconditional guarantees.
2026-10-12	Scaling Unified Multimodal Models to Long-Horizon Interleaved Generation
• Uses persistent visual references to extend unified multimodal models to long interleaved generation. • Retains semantic history, retrieves scene-relevant visual states and adapts through short self-rollouts on imperfect generated histories. • LongMIG-Bench shows improved long-horizon character and scene consistency; the abstract provides no single numerical gain and evidence is setting-specific.
2026-10-12	When Goals Leak Across Relationships: How Cached Goal Representations Shape Cross-Client Decisions
• Studies cross-client interference from goals retained in shared conversational memory. • Probes goal ownership in the KV cache and causally ablates identified representations in a controlled two-client environment. • In 20 scenarios, 22% of eligible decisions deviate from the current goal; targeted ablation exceeds random controls, but evidence uses one model and a bounded setting.
2026-10-11	EvolvAttack: Backdooring self Evolving VLM Agents
• EvolvAttack studies backdoors activated only after agent memory evolves. • Couples a visual trigger with a payload-related evolved memory state, making clean-memory static scans ineffective. • Across three VLMs, four pipelines and three evolution strategies, attack success grows from 0.5% initially to over 95% within 100 interactions in controlled tests.
2026-10-11	Recall What Is Missing: Composing Evidence from Long-Term Agent Memory
• MemPursuit formulates long-term memory reading as complementary evidence composition. • Uses sparse convex reconstruction residuals to represent missing evidence and builds nested contexts from existing embeddings without per-pair LM scoring. • Fixed-budget LoCoMo accuracy gains 7.2/5.5 points at Top-10/20; median reranking is 32.9× faster than batched BGE-v2-m3, not an end-to-end speedup.
2026-10-11	History Replays Itself: Unleashing Long-Context Capability in Linear Attention
• NoteBank separates recurrent computation from sparse episodic storage. • Archives text before queries are known, uses native backbone features for writing and addressing, and reads retrieved evidence through native replay without a separately trained retriever. • At one million tokens, mean accuracy on six RULER tasks rises from 1.7% to 70.3%, with gains across models and memory benchmarks; resource ratios are condition-dependent.
2026-10-11	Engram Injection: A Study of Temporal Validity in Parametric Language-Model Memory
• Engram Injection tests whether parametric fact memory can also select temporally valid versions. • Stores facts in LoRA adapters with FiLM clock conditioning, compares history and retrieval baselines, and audits training-age leakage. • Neither adapter solves any of 62 before/after pairs, versus 59/62 for full history and 62/62 for retrieval; storing facts does not establish correct temporal selection.
2026-10-10	Reels-RSI: Program-Level Self-Improvement for Code-Rendered Explainer Videos
• Studies how video-program agents convert revision experience into persistent lessons and executable rules. • Links findings to code fixes, stores passing and failing rule examples, and controls admission through human decisions and held-out evaluation. • Across 111 reels, lessons shorten revision loops without compounding quality gains; judge-order bias limits interpretation of preference wins.
2026-10-10	SEPM: Self-Evolving Procedural Memory Grounded in Task Graphs for Multi-Agent Collaboration
• SEPM unifies collaborative execution and procedural-memory evolution through task dependency graphs. • A shared blackboard records evidence; verified actions and dependencies become reusable procedures with creation, refinement, specialization and retirement. • It leads the compared methods in six settings; AutoGen task scores improve by 29.85, 25.86 and 13.00 percentage points across three benchmarks.
2026-10-10	A Self-Defining Operator for Distributed Applications
• Introduces an incident-response agent that turns resolved faults into persistent procedural memory. • Distills investigations into reusable diagnostic playbooks and telemetry detectors that trigger relevant LLM-guided responses. • Across 11 open-source microservice applications, recurring-fault resolution time falls by 52%; evidence is limited to the evaluated incidents.
2026-10-10	GateSight: Recurrent Memory Anticipates Attention Memory for Query-Agnostic KV Eviction in Hybrid LLMs
• GateSight tests whether recurrent-memory signals anticipate future attention-memory utility. • Predicts retention value from Gated DeltaNet write gates, decay and write magnitude, then allocates layer budgets using native attention output gates. • At one-tenth cache size it leads compared query-agnostic methods in most of four model/benchmark pairs, without guaranteeing losslessness on arbitrary tasks.
2026-10-10	How Should a Camera-First AI Agent Work?
• Separates the deadline for remembering ephemeral visual evidence from slower reasoning. • A bounded fast path buffers perception and commits selected moments to durable memory before multimodal deliberation. • Evidence uses MOT17 track windows and measured model latency; this is an architectural proposal with bounded checks, not a full real-world interaction validation.
2026-10-10	Representation Drift in Continual Learning: LoRA versus N-gram Memory
• Compares representational interference from continual N-gram memory updates and LoRA. • Under matched seven-task data and budgets, tracks layers for memory-only, shared-LoRA and merged-LoRA updates on generic and learned-task inputs. • Memory-table updates cause smaller stable generic drift but learn new tasks more weakly and retain internal task interference; backbone stability does not imply no forgetting.
2026-10-10	What Survives When Activation Steering Is Used as Memory
• Tests what user facts and preferences survive in activation-shift memory across sessions. • Stores late-layer changes under whitened gist keys, selectively reapplies shifts, and compares writing references. • About 98% of unrelated replies remain unchanged, but clean fact use is at most 15% versus 96–98% in context; reference choice can reverse preference direction.
2026-10-09	Time Delay System for Sequence Modeling
• Uses semantic similarity to estimate temporal delay for interpretable content-aware recall. • Combines differentiable lag estimation, a diagonal state space and fuzzy semantic modulation; evaluated models use 256-dimensional states and a maximum delay of 30. • Results improve IMDB classification and CNN/DailyMail generation in the reported study; broader generalization is untested and hyperparameter sensitivity remains.
2026-10-09	Temporal Evidence Forests for Long-Horizon Multi-Party Conversational Memory
• T-Mem organizes long-term multi-party conversational memory as temporal evidence forests. • Finds locally concentrated evidence, models time-ordered state transitions and support, compresses low-information spans, and linearizes evidence for answering. • It improves results on two benchmarks across models, domains and query types, within the evaluated interaction settings.
2026-10-09	DP-MIND: Differentially private memory-guided Long context Inference
• DP-MIND uses a learned compact memory for differentially private long-context inference. • Processes locally perturbed chunks, optimizes memory construction for answer correctness and consistency, and locally denoises corrupted answer spans. • It supplies privacy analysis and utility gains on contexts up to one million tokens, under the evaluated privacy mechanism and benchmark conditions.
2026-10-09	Forgetting is Not Always Bad: A Neuro-Inspired Memory Repair Mechanism for Poisoned LLM Agents
• DREAM repairs poisoned agent memory through selective functional forgetting. • Perturbation-based reactivation, anomaly diagnosis and dynamic attenuation suppress harmful memories while reinforcing benign ones without training. • Reported backdoor attack success falls by over 95%, with up to 2.13× speedup and 37.7% fewer tokens than A-MemGuard; this is not a complete-erasure guarantee.
2026-10-09	Beyond the Final Embedding: Retrieving Memories from Token-Level Evidence
• Retrieves assistant memories using token-level evidence beyond a single final embedding. • Aggregates sparse token matches with whole-memory similarity; an optional query adapter leaves the encoder and stored representations fixed. • Gains span eight retrieval benchmarks and long-term QA; adapted Qwen3-Embedding-4B improves mean Recall@10 and NDCG@10 by 13.3 and 14.3 points.
2026-10-09	Test-Time Memory Construction
• T²Mem defers query-specific memory construction until inference. • Stores lightweight episode metadata and links offline, then progressively retrieves and organizes original interactions during reasoning. • It outperforms strong baselines across backbones on LoCoMo and LongMemEval while reducing offline construction overhead; results are benchmark-specific.
2026-10-08	Gamma: Grounded Multi-Agent Memory with Mechanical Audit for Memory-Intensive Manipulation
• GAMMA separates embodied episodic memory into a VLM writer and a subgoal reasoner. • Uses spatial detection, an append-only text bank and mechanical visual checks to admit, defer, correct or reject claims before emitting grounded subgoals. • On 16 memory-manipulation tasks it recovers much of a fixed policy’s oracle ceiling and matches it on some tasks, subject to grounding and verification limits.
2026-10-08	What to Admit and How to Present: Governing Persistent Memory in LLM Agents
• Separates memory admission from presentation to study their independent safety effects. • Whole-entry or typed-unit adjudication feeds deterministic eligibility policies, with matched budgets and fixed-presentation comparisons. • External failure rates fall by 6.7 and 8.8 points, but personalization failures rise and preregistered criteria are not all met; benefits are task-dependent.
2026-10-08	Gated Memory: Admission-Controlled Memory Formation for Conversational AI
• Gated Memory places admission before fact extraction to preserve transient, conditional and privacy context from utterances. • Admission and conditional enrichment produce atomic facts with provenance, category, scope, time and place, prohibiting unsupported entities. • With retrieval and generation fixed, judge accuracy improves relatively by 2.6% on LoCoMo-10; its atypical emotional density limits generalization.
2026-10-08	REMORY: Learning Residual Memory for Context Compaction
• REMORY supplements textual summaries with bounded soft memory tokens carrying residual historical information. • A memory network conditions on both history and summary to approximate a frozen LLM continuation under full history. • On SummHay it approaches the full-context joint score using 5.2% of input positions and reduces repeated tool outputs on agent tasks; results are model- and benchmark-specific.
2026-10-08	From Retrieval to Reconstruction: Constructing Evolvable Cognitive Memory for Long-Term Dialogue
• CogMem separates attributed claims from events and facts in a provenance-aware dialogue-memory graph. • Incremental records are consolidated and temporally reconciled; anchoring, traversal, intersection and grounding reconstruct query context. • LoCoMo and LongMemEval show strong multi-hop, temporal and update performance, with complementary ablations; no aggregate effect size appears in the abstract.
2026-10-08	Beyond Sequences: Distilling Structured Decision Memory for LLM Recommendation
• MARI stores goals, constraints and trade-offs as interpretable structured decision memories for LLM recommendation. • Offline post-hoc distillation builds records from heterogeneous behaviors and user content; online retrieval separates memory construction from inference. • It improves next-item and difficult-choice prediction over tested baselines; distilled rationales should not be equated with known true user motives.
2026-10-08	Memory Type Varies: Empowering LLM Agents for Long-Term Memory with Diverse Strategies
• MemoType distinguishes memory and query types instead of imposing one retrieval strategy on mixed memory corpora. • TriMEM supplies type annotations, a learned router selects matching retrieval strategies, and theory bounds single-strategy precision. • Recall@1 improves by up to 16.18% on three datasets; applicability depends on the type taxonomy and routing quality.
2026-10-08	DeltaReplay: Task-Relative Memory Reuse for Mobile GUI Agents
• DeltaReplay treats memory usefulness as a step-level relation to the current task instead of an all-or-nothing trajectory match. • Stored page-action graphs separate reusable operations from task parameters, enabling follow, reparameterize or fallback decisions. • Success rises by up to 10.3 points on AndroidWorld and 25.0 on SPA-Bench; evidence focuses on reuse of partially matching trajectories.
2026-10-08	MemTrial: Learning When to Trust Memory in LLM Portfolio Agents
• MemTrial credits memories by their marginal effect on same-date decisions rather than common market movement. • Eight experience combinations estimate Banzhaf values; hierarchical Bayesian pooling and unseen-date validation govern whether the agent trusts them. • Mean utility improves by 21.2% over the strongest experience-learning agent across five settings; conservative fallback can still lose and provides no investment guarantee.
2026-10-08	DPPM: Dual-Path Parametric Memory for Personalized Language Models
• DPPM separates evidence preservation from ordered revision in cross-session parametric personalization memory. • An Evidence path pools history representations, while a Delta path updates associative state; their fusion generates history-conditioned LoRA adapters. • Across tested backbones it reaches 54.22% on PersonaMem-v2 and 86.79% on PrefEval; evidence concerns these personalization tasks.
2026-10-08	Memento 3: Model-Based Recursive Self-Improvement through Reflective Rulebooks
• Memento 3 stores revisable environment hypotheses in a natural-language rulebook as persistent semantic memory. • Compiled executable models are revised using prediction errors; updates require fidelity checks and cell-exact replay of observed transitions. • It clears all 25 public ARC-AGI-3 games; Pong evidence covers only three episodes, so these cases do not establish general self-improvement.
2026-10-08	Event-Centric Memory with Query-Aware Graph Augmentation for Long-Term Conversational Agents
• QGMem stores experiences and state changes as event-indexed atomic memories consolidated into dynamic traces. • Hybrid retrieval and reranking activate compact working memory; a local graph becomes a graph token supplied with textual evidence to the LLM. • Six benchmarks report gains in retrieval, multi-hop evidence and conflict resolution; the abstract gives qualitative rather than aggregate numerical results.
2026-10-08	Look Back, Think Ahead: Visual Memory on Demand for Efficient Multimodal Reasoning
• ViMoD preserves recoverable visual details alongside compact context for changing evidence needs during multimodal reasoning. • Coarse groups link to original fine tokens, and a history-conditioned router selects, retains or replaces active detail groups. • At a 20% visual-token budget it beats tested baselines on eight benchmarks, with a 39% mean normalized-score gain; evaluation centers on Qwen3-VL-4B.
2026-10-08	Use and Disuse: Intent-Structured Experience Consolidation for Memory and Learning in LLM Agents
• Hippocam organizes ongoing work by nested intents and gradually consolidates completed work into persistent experience. • Recursive consolidation abstracts older history while retaining original interactions for progressive recovery; reused experiences are reinforced and updated with new ones. • Presents a memory-and-learning architecture without parameter updates; the reviewed abstract supplies no quantitative evaluation, so effectiveness remains unestablished here.
2026-10-07	Relevance Is Not Sufficiency: What Actually Closes the Evidence Gap in Long-Term Memory QA
• BFR targets jointly sufficient memory evidence rather than independently relevant records. • With a fixed flat store, information requirements guide initial selection, followed by deeper search and complementary entity or session views under a budget. • LongMemEval-S judged accuracy rises from 72.4% to 82.2% with 91.4% Turn Hit; conclusions concern matched-store, budgeted comparisons.
2026-10-07	COOL: Curiosity-Driven Object Ownership Learning for Personalized Robotic Assistance
• COOL infers ownership from persistent person-object interaction memory for personalized robotic assistance. • Two Qwen3-VL-4B agents share spatial and event records: one retrieves ownership evidence, while the other explores and refreshes stale observations. • Offline, ablation and robot experiments support the approach; shared or unknown labels represent uncertainty, with perception and temporal identity association limiting reliability.
2026-10-07	Learning Situation-Conditioned Thinking Policies for Long-Term LLM Agents
• Transforms historical reasoning experience into situation-conditioned thinking memory rather than indefinitely growing context. • Repeated patterns across independent episodes become consolidated knowledge in a lightweight policy, with detailed reasoning delegated to an LLM. • Reported DeepSeek reasoning F1 rises from 0.789 to 0.868; perfect controlled-task metrics do not establish perfect open-domain performance.
2026-10-07	SkillForge: Co-Evolving Skills and Agents via Dynamic Skill Lifecycles
• SkillForge manages procedural skill memories through trial, active, stable and retired lifecycle states. • Base-model rollouts pre-filter weak skills; reinforcement learning then validates, stabilizes, mutates and retires entries alongside policy updates. • Interactive benchmarks show up to 7.8% relative success improvement and a 5k-plus lifecycle dataset; evidence primarily concerns training-time skill libraries.
2026-10-07	HGP:An on-device personalized agent memory via hybrid graph storage
• HGP represents heterogeneous long-term traces in hybrid graphs separating episodic, semantic and procedural memory. • A lightweight classifier routes memory types, while working-state trajectories capture current constraints and support incremental profile refinement. • Across two benchmarks, PAL-Set S-score reaches 35.58, nearly seven points above the strongest baseline; general on-device applicability is not fully detailed in the abstract.
2026-10-07	ExperienceIndex: Artifact-Grounded Memory
• ExperienceIndex stores artifact contributions and cross-artifact relations learned from previous reasoning tasks. • Single-artifact and artifact-pair experiences guide new retrieval and can transfer from stronger to weaker language agents. • Answer quality improves by up to 11.0 points and online cost falls by up to 50.5%; cross-task transfer is evaluated over a shared artifact corpus.
2026-10-07	VideoEvolve: Co-Evolving Memory and Retrieval for Long Video Understanding
• VideoEvolve jointly improves what video agents remember and how they retrieve it instead of fixing memory construction. • Memory and retrieval evolvers alternate reinforcement learning from a low-frame-rate overview, guided by bottleneck and capability feedback. • Multiple long-video benchmarks support effectiveness; body sections confirm language and multimodal models write and retrieve memories, while the abstract gives no aggregate effect size.
2026-10-07	EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory
• EngramEdit turns conditional memory into an independently editable factual interface with a frozen Transformer backbone. • Joint n-gram embedding updates match targets across paraphrases and edits, penalizing changes to frequently reused entries. • Experiments report near-perfect edit success and multi-hop gains while largely preserving unrelated knowledge; conclusions depend on the tested conditional-memory architecture.
2026-10-07	Never Look Back: Understanding Persistence in 3D Object Memory from Egocentric Videos
• Ledger preserves object locations, histories and descriptions after objects leave view, supporting unanticipated spatial questions. • Timestamped observations form resting segments, moves require repeated evidence, and a language model reads saved text rather than replaying video. • HD-EPIC accuracy rises from 29.7% to 42.6%; stitched scenes still expose construction and retrieval failures, and offline evaluation does not establish live deployment.
2026-10-07	MemoWM: How World Models Change What Agents Need to Remember
• MemoWM allocates experiential memory according to what a shared world model can reconstruct. • Task-aware allocation balances downstream reconstruction error against storage, accounting separately for model parameters and recurring experience costs. • Across five benchmarks accuracy is 42.42%, 2.62 points above the strongest baseline, with 53.9% less experience storage than MIRIX; total cost depends on model capacity and history size.
2026-10-06	AgentMemGate: Addressing Speculation Contamination in Conversational Assistant Memory
• AgentMemGate addresses speculation contamination when unresolved plans become facts in conversational profile memory. • A write-time classifier separates speculation, completed events, corrections and other statements, with conditional promotion and deletion. • On the harder held-out set, contamination is 3.4%–5.7% and accuracy rises by 9–13 points; profile-field matching remains the main bottleneck.
2026-10-06	Decide Before You Look: Learning Which Retrieved Memories Deserve Pixels
• PixelTriage decides which retrieved visual memories deserve pixels instead of always using images or text proxies. • A small predictor reads dialogue, notes and thumbnails, trained on marginal-value labels from a frozen 27B answering model. • Three benchmarks use 11%–23% of visual tokens without significant accuracy loss and run 2.9 times faster on DMV; results depend on the tested readers and budgets.
2026-10-06	DAEDALUS: Bootstrapping Agent Memory from Self-Generated Tasks
• DAEDALUS bootstraps procedural memory from self-generated practice without predefined tasks or oracle verifiers. • An explorer proposes tasks, a solver derives heuristics from failures, and repeated successful use gates consolidation into a test-time memory bank. • Across three benchmarks mean success improves by up to 15.9 points and pass^5 by up to 2.2 times; transfer still depends on exploration budget and environment feedback.
2026-10-06	Test-Time Agent Evolution for Long-Horizon Legal Reasoning
• Evolves deployment-time memory by retrieving, adapting and consolidating prior-case experience for long-horizon legal agents. • Memory updates combine with rubric-aligned role verification, reusing experience without changing backbone parameters. • Five backbones improve on J1-EVAL and LegalWorld with complementary ablations; benchmark gains do not guarantee reliability in real legal decisions.
2026-10-06	DeltaTTT: Layerwise Optimization for Nonlinear Recurrent Memory
• DeltaTTT addresses single-pass optimization difficulty in nonlinear recurrent memory, where serial dependence can underperform fixed-base parallel updates. • Layerwise local targets and state-dependent delta rules replace joint inner-loop optimization while supporting chunkwise parallel computation. • Language modeling and retrieval improve on DeltaNet and LaCT backbones; the proposed explanation is grounded in exploratory optimization experiments.
2026-10-06	MINDSET: Energy-based Schema Evolution for Long Conversational Agent Memory
• MINDSET treats conversational memory as constrained versioned-state management over immutable episodes. • Minimum-energy transitions reinforce, supersede, split or create schemas, balancing contradiction, historical damage and fragmentation with hysteresis. • An 850-question sample shows strong answer and retrieval scores; MemoryAgentBench margins are small, with a separate 700-question cross-model check.
2026-10-06	When Forgetting is not Catastrophic: On the Mechanics of Spurious Forgetting
• Separates reversible loss of access along a shared direction from genuine erosion of individual factual memories. • A minimal associative model explains the dynamics, followed by common-shift and update-direction interventions in synthetic and pretrained Transformers. • Removing the shared direction restores old facts, showing recall collapse need not mean storage loss; accumulated fact-specific changes can still cause irreversible forgetting.
2026-10-06	Whose Memory Is It? Scope-Aware Commit Rules for Long-Term LLM Memory
• CASK scopes persistent commits by discourse ownership so hypothetical branches and others’ beliefs do not become shared facts. • It preserves stable ownership relations rather than fragile internal coordinates, retaining provisional content only within its original scope. • Controlled long-dialogue conflicts and tool traces improve admission and reduce contamination; evidence concerns the tested representations and scenarios.
2026-10-06	Visual Memory Attacks Can Persist Through The KV Cache
• P-VMI tests whether adversarial visual influence persists in later-token KV state after the source input is removed. • Image optimization, attention masking and cache-swap ablations isolate persistence, including compaction that keeps a model-generated summary cache. • The strongest Qwen3-VL-8B-Instruct configuration reaches about 90% target success across extended dialogue; results depend on the specified attack and removal conditions.
2026-10-06	Frozen Models, Evolving Expertise: Model-Agnostic Learning from Deployment Experience for Multimodal Medical AI
• Enables frozen LLMs and VLMs to accumulate deployment expertise through skills, factual memory and multimodal cases. • Facts and visual examples retain support from prior cases or trusted sources; updates survive only when they help new cases without degrading earlier ones. • Six benchmarks and four backbones report up to 34.2% improvement on medical tasks with transfer results; benchmark gains do not establish clinical deployment validity.
2026-10-05	When to Remember, When to Abstain: Category-Conditioned Retention for Reliable Agent Memory
• Uses category-conditioned confidence thresholds to avoid storing weakly supported values and beliefs as facts. • Compares global and category-specific retention across 4,715 assertions from 100 synthetic personas, measuring reliability and coverage. • A stricter threshold for values reduces unsupported retention from 6.2% to 4.0%; evidence is limited to synthetic cold-start and repeated held-out evaluation.
2026-10-05	AMBER: Training Long-Horizon Web Agents through Append-Only Memory
• AMBER uses append-only free-form memory to prevent facts and corrective feedback from disappearing during rewrites. • The language agent jointly learns reasoning, actions and memory writes with outcome-based reinforcement learning. • On WebArena Lite, average success improves by 4.09 points over overwrite memory and five-run task coverage by 4.8 points; evidence is specific to this benchmark.
2026-10-05	Lineage-Aware Memory Governance: A Derivation-Gated Framework for Privacy-Preserving Column-Level Access Control in Enterprise AI Agents
• Analytical Memory Units attach column-level lineage to shared cached results to prevent unauthorized derived-memory reuse. • Retrieval requires authorization for every source column and detects conflicting derivations of same-named metrics. • Six experiments eliminate measured leakage while retaining about 81.5%–82.6% reuse; guarantees depend on complete lineage, and the live LLM-SQL demo has only nine round trips.
2026-10-05	Understanding and Mitigating Inference-Time Overreliance Using Agentic Memory
• MEMTRIM studies overreliance even on benign, correctly stored and appropriately retrieved agent memories. • Write-time evidence indexing and read-time trimming remove repeated or conflicting evidence while preserving useful memory-specific content, without retraining. • Experiments across models and embedding-based or structured memories reduce overreliance while retaining useful benefits; no aggregate effect size is given in the abstract.
2026-10-05	MemCo: Memory-Centric Collaboration for Generalizing LLM Agents to Unseen Environments
• MemCo separates environment-specific details from transferable workflows using complementary local and global memories. • Local trajectories induce global workflows, while state- and phase-aware routing controls cross-agent experience reuse. • Interactive benchmarks report higher success and less redundant exploration; findings concern the tested unseen environments, without a single aggregate gain in the abstract.
2026-10-05	PACMI: Provenance-Aware Cascading Memory Invalidation for Long-Term LLM Agents
• Preserves historical evidence while invalidating outdated memories and their dependent records. • Uses a provenance graph, typed dependencies, a four-state validity lattice, and stale-premise detection. • Leads final accuracy on 100 cases and 300 queries; removing propagation raises errors from 3 to 11, but that paired difference is not significant.
2026-10-05	Off-Policy Merging Beats On-Policy Self-Distillation for Continual Learning
• Studies interference between new updates and existing knowledge in post-trained language models. • Learns updates on earlier donor checkpoints, scales them before merging, and optionally masks sensitive directions. • Improves the old/new-task trade-off over SFT and on-policy self-distillation across tested continual-learning settings.
2026-10-05	The Optimization Landscape of Learning Compacted Context Models
• Analyzes the brittle and flat optimization landscape of compact KV memories learned through frozen language models. • Uses a simplified Perceiver architecture for continuous context compaction. • Matches a full Perceiver and improves compaction utility on finance, legal, literary, and code multiple-choice tasks.
2026-10-05	ReMem: Streaming Video Understanding With Long Context Retention
• Addresses progressive loss of historical information in streaming vision-language models. • Streaming Context Memory compresses history while Retrieved Vision Memory retrieves query-relevant evidence for the VLM. • The training-free adaptation reports leading results on streaming and long-video benchmarks; absolute scores are not stated in the abstract.
2026-10-05	Capability-Driven Self-Evolution of Agent Memory
• Evolves executable memory programs by individual capabilities rather than aggregate score alone. • Uses dependency-aware target selection, history-guided diagnosis, and paired differential traces to integrate complementary improvements. • PrisMem exceeds the strongest tested baselines by 10.54 points on BEAM-1M and 7.83 points on LongMemEval-M.
2026-10-05	MarvisNav: Making Memory Visible on Route Choices for Zero-Shot Object Navigation
• Binds exploration memory to visible route candidates so a VLM can jointly assess target relevance and prior progress. • Projects topological nodes and richer-than-binary exploration states onto egocentric views. • Achieves 81.2% success and 42.5% SPL on HM3D with real-robot validation, without policy training.
2026-10-05	Test-Time Adaptation of Reasoning Strategies with Bayesian Nonparametric Memory
• Stores reasoning behaviors distilled from problems and reflection in an adaptive cross-domain memory. • An HDP Gaussian mixture retrieves strategies through posterior prediction and updates statistics, creates clusters, and reassigns behaviors online. • Reports improvements over tested memory modules on AIME’25, Omni-MATH, and PhysReason, including cold starts.
2026-10-05	VideoTapestry: Query-Adaptive Memory Refinement for Multi-Agent Long-Video Understanding
• Builds three-level video memory spanning narrative, events, and fine-grained relations, then refines it for each query. • Scale-specific agents revisit relevant video regions and merge targeted observations into hierarchical memory without training. • Improves accuracy over direct GPT-5.5 inference by 7.0–17.2 points across four tested video-QA benchmarks.
2026-10-05	Balancing Memory Pathways: Analyzing and Improving Memory Utilization in Hybrid LMs
• Finds that hybrid language models underuse recurrent memory relative to attention. • An auxiliary objective limits attention access to earlier context while recurrent state traverses the full sequence. • Improvements generalize across models and QA or agent tasks, particularly for long contexts and information aggregation.
2026-10-05	MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents
• Curates raw multimodal history on demand to recover details lost by query-agnostic memory construction. • An RL policy chooses retrieval or delegated LLM/VLM curation and controls evidence volume, model choice, and visual access. • Five multimodal memory benchmarks show favorable performance-cost-latency trade-offs across optimization preferences.
2026-10-05	Recursive Video In-Context Learning for Agentic Robot
• Stores demonstration videos as navigable event hierarchies for high-level language agents controlling frozen robot policies. • Read-only tools expose keyframes, phases, moments, and clips, retrieved at the granularity needed for the current subgoal. • One demonstration per task raises success from 92.6% to 96.5% on LIBERO-PRO and 86.7% to 95.8% on LIBERO-Plus without training.
2026-10-04	ProactiveVLA: Augmenting Embodied Memory through Proactive Environment Exploration
• ProactiveVLA expands reusable embodied memory through proactive exploration beyond repeated target-task practice. • The reasoning agent proposes affordance and state-change goals, verifies outcomes and consolidates task and exploration experience for later planning. • It outperforms baselines at matched budgets on LIBERO-Pro and RoboCasa365; Goal-T single-primitive success is 48% versus 19%, without establishing universal real-world gains.
2026-10-04	MemTrace: State-Consistent Memory for Long-Horizon Coding Agents
• Preserves execution history with repository-state provenance to prevent restoration of stale coding evidence. • Immutable traces form a dependency graph; compact anchors reconstruct state and validity checks precede selective recall. • Under matched backbones and harnesses, reported gains are 21.2 points on DeepSWE, 4.4 on SWE-EVO, and 17.8 on SWE-Milestone.
2026-10-04	Enhancing Long-Video VLM Embeddings with Query-Aware Streaming Latent Reasoning
• Accumulates cross-clip evidence in persistent latent think tokens with fixed-size video representations. • Processes selected frames in order with stepwise contrastive supervision and final-embedding self-distillation, supporting query-aware or reusable indexing. • HourVideo Hit@1 rises from 54.2 to 70.7 for 2B and 57.4 to 72.8 for 8B models; transfer is tested on the stated backbones.
2026-10-04	PreAct-Nav: Agentic Reasoning Before Action for Urban Navigation
• Maintains medium-horizon subgoals and execution experience for high-level urban-navigation decisions. • A VLM evaluates predicted consequences, then actual observations correct assumptions and continue or revise remembered subgoals. • Reports improved action selection, especially on longer and more winding routes; the abstract does not provide absolute success rates.
2026-10-04	EVISKILL: Grounding Skill Evolution in Replayable Evidence
• Stores execution observations as replayable evidence cards supporting persistent procedural edits. • Targeted re-execution checks edits, preserves evidence and provisional changes across epochs, and gates final incorporation on global validation. • Experiments across three interactive benchmarks and six backbones support the approach; no uniform effect size is given in the abstract.
2026-10-04	ReMAP: Restoring the Perceptual Cycle with Reasoning-Time Latent Visual Memory
• Combines static global and dynamic local latent visual memories to restore evidence during extended multimodal reasoning. • An RL access policy decides when to retrieve scene context or re-encoded local regions conditioned on the current reasoning state. • Gains reach 8.38 points on MuirBench and 14.84 on MIMIC; matched-query visual tokens fall by 51.0–76.8%.
2026-10-04	Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy
• Addresses sycophancy induced even by factually correct persistent user memories. • Uses counterfactual induction, context-aware reflection, and evidence-based reasoning to calibrate each memory’s influence. • Reports improved memory reliability across three benchmarks without establishing elimination of all sycophancy.
2026-10-04	AECG: Asymmetric Experience Consolidation and Governance In Multi-Agent Systems
• Governs outdated, weakly supported, or scope-mismatched procedural memories in multi-agent systems. • Preserves coordination scope, monitors reliability, prioritizes risky knowledge, and requires paired replay before reactivating revised skills. • Leads 11 of 12 tested settings; removing scope preservation reduces accuracy by up to 16.89 points.
2026-10-04	Learning without Overwriting: A Theory of Self-Distillation and Supervised Fine-Tuning in Continual Reasoning
• Analyzes why self-distillation and supervised fine-tuning differ in overwriting prior reasoning knowledge. • Models reasoning as DAG search and contrasts hint-induced sparse updates with dense updates along supervised paths. • Provides optimization guarantees under its assumptions and identifies a role for pretraining diversity; this is theoretical rather than universal empirical evidence.
2026-10-04	StateWise: Diagnosing and Repairing Persistent Operational State Before Agent Actions
• Repairs invalid persistent operational records before coding agents use them as action premises. • Counterfactual replanning identifies critical records, then typed evidence, read-only checks, and targeted clarification support persistent corrections. • Across 150 executable cases, correctness reaches 93.3% versus 38.7% for the baseline, with no unsafe actions in the tested cases.
2026-10-04	Task Vector Descent: Learning from Non-IID Batches
• Studies how clustered training streams degrade prior knowledge and how much of each task displacement should be retained. • Scales both a task vector and optimizer state before continuing training. • Intermediate integration often improves continuing performance across pretraining and post-training; matched-rate controls show this is not merely learning-rate scaling.
2026-10-04	TeleTune: Evolving Agent Skills From Offline Telemetry
• Mines reusable textual workflows from goal-free, non-replayable, interleaved software telemetry. • Proposes library edits from action-prediction errors, retains held-out improvements, and retrieves workflow-matched demonstrations for new tasks. • Achieves 77.1% and 80.6% success on two benchmarks; offline edit validation uses 5–75 times fewer tokens than live evaluation.
2026-10-04	Have I Scene This Before? Spatially Grounded Conversational Memory for Complex Queries in Egocentric Assistants
• Grounds conversational facts in persistent physical objects to preserve spatial and dialogue evidence together. • Combines object-centric memory, 3D reconstruction, segmentation, and dialogue updates; Ego-SpaCR supplies 620 videos and 3,100 queries. • Achieves the best overall accuracy among tested methods with fewer input tokens; removing 3D evidence substantially hurts performance.
2026-10-03	Principles that Guide, Actions that Inform: Agent Evolution via Knowledge Abstraction
• SAGA abstracts execution experience into episodic descriptions, reusable procedures and principles with applicability conditions. • Evidence links are retained; retrieved principles become task-specific guidance through corrective feedback and action resampling, continuously updating hierarchical memory. • ScienceWorld and ALFWorld improve, with ablations supporting contextual instantiation and action regulation; the abstract provides no single aggregate gain.
2026-10-03	What to Preserve in Recursive Computation: A Local Predictive Sufficiency Principle
• Constrains recursive compression by local predictive sufficiency to preserve useful historical information. • Instantiates predictive tests and conditional optimizer projections in a CCM memory host for LLaMA-7B and Gemma, alongside other recursive domains. • Improves over matched compressed hosts on DailyDialog and Multi-Session Chat, especially at greater memory depth; findings remain protocol-specific.
2026-10-03	RETRACE: From Entangled Repair Histories to Reusable Experience for CI Repair
• Reconstructs problem-level repair memories from PR histories containing failures, reverts, and unrelated changes. • Reconciles backward retained-change analysis with forward commit evolution using CI evidence, then retrieves three-level repair abstractions for new failures. • On 565 PRs, Pass@1 rises from 19.6% to 31.9% with MiniMax-M2.5 and from 23.3% to 32.8% with DeepSeek.
2026-10-03	A Tropical Geometry View of Forgetting: A Per-Unit Projector for Knowledge-Preserving Fine-Tuning
• Defines knowledge-preserving subspaces from the old tokens that activate each ReLU unit. • Derives a per-unit projector and budget separation using tropical geometry and tests gate-aware versus control token sets. • OPT-1.3B forgets less at six matched budgets; OPT-6.7B matches retention while learning more, with separate pruning-repair limitations.
2026-10-02	A GHOST in Long-Horizon Agents: Governance Hazard from Overlooked Safety Constraints across Turns
• Studies failures to retain and apply safety constraints stated earlier in long interactions. • Combines restoration of historical semantic constraints with deterministic pre-execution auditing and conditional-hazard analysis. • The GPT-5.5 setup shows 11.5% baseline violations and none with the defense in tested runs; the theorem depends on hazard assumptions.
2026-10-02	Decoupling Memory from Context: Structured Memory for Token-Efficient Test-Time Continual Learning
• Stores reusable cross-query strategies in graph memory instead of continually appending them to shared context. • Accumulates, refines, and connects strategies while retrieving only a relevant bounded subgraph per query. • Reports competitive downstream performance with roughly 81–85% fewer memory-construction tokens than the tested baselines.
2026-10-02	When History Fails to Become Experience: Action Calibration in Language Agents
• Tests whether language agents use action-outcome associations in interaction history. • Perturbs those associations, adds explicit outcome labels, and learns a calibrator that reassesses actions and selectively records experience. • Shuffling causes only modest degradation; explicit association improves success and reduces repetition, within the studied tasks.
2026-10-02	Skill2Real: Agentic Skill Learning for Zero-Shot Sim-to-Real Robot Manipulation
• Builds reusable local and high-level executable skills from simulated robot experience. • A proposer-verifier-governor loop approves updates; high-level composition learns over frozen local skills and both memories transfer through a shared API. • LIBERO-Pro Long success rises from 2.0% to 56.3%; four real tasks average 78.75% completion without transfer-time skill updates.
2026-10-02	Interpreting at Write Time: A Policy Ablation for Multi-Goal Agent Memory
• Studies how write-time goals determine which information survives history summarization. • Holds reading fixed while comparing neutral, all-goal, and separate per-goal summaries. • Per-goal summaries improve relevance, completeness, and accuracy; benefits depend on alignment with the goal that later queries memory.
2026-10-02	When to Compile a Computer-Use Agent? Measuring Payback and Making Compilation Decisions for Token Efficiency
• Studies when repeated GUI execution experience should become a reusable program, accounting for failed compilation and interface drift. • Uses observed arrivals and compilation outcomes to decide compilation under a cumulative cost budget. • Successful compilations have estimated payback of 2–16 uses; simulations over recorded arrivals save 17.3% of tokens versus ReAct.
2026-10-02	SlimKV: Joint Token-Feature KV Cache Compression with Reconstruction-Free Beacon Attention
• Writes long context into low-rank beacon memories with joint token and feature compression. • Uses low-rank-aware training, layer-adaptive ranks, and positional asymmetry to read directly in latent space. • Retains over 96% of uncompressed LongBench score at 4–8× compression and leads tested baselines at 16–32×, with position-varied retrieval checks.
2026-10-02	Sentry: Learning to Recover from LLM Agent Failures at Test Time
• Persists successful failure-recovery lessons as conditional experience rather than permanent prompt content. • Retrieves matching lessons only after a failure and writes a new lesson only after reward-independent recovery verification. • Beats the strongest runtime-intervention baseline by 37% on average and transfers lessons to held-out tasks; full-playbook exposure hurts performance.
2026-10-02	KV$^2$: A Self-Refining KV Cache
• Uses selective reconstruction to preserve context in a query-agnostic reusable KV memory. • A lightweight scorer selects informative tokens, which alone are reprocessed to estimate final eviction scores. • At a 2% KV budget on RULER 16K it exceeds the next-best baseline by over 40 points; evidence is specific to the tested tasks and budgets.
2026-10-02	A Step Towards Forgetting: Optimiser History and the Loss of Answer Mass
• Decomposes forgetting into confusion among old answers and probability leakage outside their answer set. • Separates Adam history by gradient age and resets momentum while matching the initial update norm. • Older gradients drive answer-mass loss; resetting history improves later retention, although immediate effects depend on model state.
2026-10-02	SEER: Self-Evolving Event Reasoning and Retrieval for Time Series Forecasting
• Converts forecast errors into reusable retrieval experience and causal rules rather than merely adding news. • Maintains reflection memories for query expansion and event filtering plus a causal knowledge base, with strictly time-bounded updates. • Outperforms tested forecasting and language-model baselines across six volatile time-series benchmarks.
2026-10-01	SEDIMA: Cross-Run Hierarchical Insight Memory for Evolutionary Search Agents
• Distills search traces into persistent natural-language insights reusable across runs and problems. • Clusters insights with attention-weighted centroids and retrieves guidance for subsequent evolutionary mutations. • With 100 evaluated candidates, average final performance improves by 5.5% on AlgoTune and 6.6% on ALE-Bench LITE.
2026-10-01	Mitigating Private Data Leakage in LLMs with Whiteout
• Targets memorized personal sensitive information with precise overwriting instead of broad knowledge removal. • Uses designed obfuscation samples and tests utility and robustness under black-box and adaptive white-box attacks. • Reports better protection than tested unlearning alternatives with limited utility loss; the abstract provides no uniform leakage rate or absolute privacy guarantee.
2026-10-01	APDMem: Agent-Controlled Progressive Disclosure for Query-Adaptive Long-Term Memory
• Organizes conversation memory into thematic summaries, personal facts, turn-level evidence, and raw messages. • A controller progressively opens finer evidence as needed, and a synthesizer consolidates facts, orders events, and flags conflicts. • Reports strong LongMemEval performance while accessing about 8% of conversations; the abstract does not state an absolute accuracy.
2026-10-01	Harnessing LLMs as Agents: What Does It Cost?
• Develops resource bounds for agent memory communication, addressing, and reuse of intermediate state. • The LAM abstraction relates context and persistent-memory capacity to access interfaces and recomputation, alongside verification costs. • Controlled GPT-6 Astra experiments test communication and reliability predictions; bounds depend on the fixed-model and resource-accounting assumptions.
2026-10-01	How To Train Your World Model: Fine-tuning vs RAG for LM-based World Modeling
• Studies retrieval and reuse of exploration transitions by language-model world models. • Compares fine-tuning and experience retrieval in five environments, diagnoses retrieval errors counterfactually, and combines learned dynamics with maintained memory. • Fine-tuning yields higher rewards in 15 of 20 settings; the hybrid improves results, while the two approaches differ in data efficiency.
2026-10-01	Fisher-Guided Submodular Data Selection for Continual Pre-Training of Large Language Models
• Treats continual-pretraining data selection as control of interference with established knowledge. • Separates gradients into high-Fisher anchor and low-Fisher frontier components and selects data with a streaming submodular objective. • On two models adapting to medical data, 1B selected tokens outperform 10B replay tokens on adaptation and forgetting under the tested protocol.
2026-10-01	Role-aware Heuristic Episodic Attention for Conversational LLMs
• Assigns distinct retention policies to global instructions and episodic turns to reduce contextual decay. • REA keeps instruction prefixes and user inputs, compresses replies, and heuristically chooses raw, compressed or omitted history. • Raises Long-MT-Bench+ judge score from 6.32 to 7.36 with a reported 2.91× latency improvement; evaluation includes three backbones and bilingual role play.
2026-10-01	Madeleine: Learning Involuntary Recall for Conversational Memory from Simulated Lives
• Learns associations between conversational cues and relevant memories that are not textually similar. • Simulated lives supply cue-trigger pairs to train a residual query encoder over frozen similarity, with no online LLM calls. • Scores 66.6 inside HyperMem on official LoCoMo-Plus; alone approaches released HyperMem with about 1/21 of its answer context.
2026-10-01	FlashBack: Knowing When to Remember in Streaming Vision-Language Models
• Studies when streaming VLMs should recall history to avoid disrupting live perception. • FlashBack selects an isolated Recall trajectory with query-local Side-KV access while leaving persistent Native state unchanged. • Improves several memory-dependent tasks on OVO-Bench and StreamingBench while largely preserving real-time perception, without additional training.
2026-10-01	Sleeping Secrets: How Fine-Tuning Reawakens Privacy Risks in Language Models
• Tests recovery of private associations that are dormant in ordinary outputs. • ReGap filters model-generated candidates by likelihood and trains LoRA without secret answers or private auxiliary data. • Reports 6–21-point gains across six models, including 42% to 63% on previously exposed associations and no gain in unexposed controls.
2026-10-01	Beyond Memory: Harnessing Long-Horizon Agents with Explicit Belief States
• Maintains explicit world estimates and unresolved requirements as a revisable decision state. • PoS checks consistency and progress, detects unproductive belief traps, and selects requirement-specific recovery. • Reports leading overall performance across four execution and diagnosis benchmarks; ablations support maintenance and recovery, not open-world reliability.
2026-10-01	Managing Context and Communication in Distributed Agentic UAV Swarms
• Stores operational knowledge as atomic core, local and peer-specific memories for distributed SLM agents. • An event-driven lifecycle maintains notes and an interest-aware gossip engine shares semantically novel, recent information. • Completes every simulated search-and-rescue run versus 70–85% with flooding and roughly halves tokens; evidence is a ten-UAV simulation.
2026-10-01	Not All Experience Belongs in the Weights: Component Routing for Self-Improving GUI Agents
• Splits GUI experience into locators, procedures, state facts and lessons to compare weight versus context storage. • Routes matched items using recurrence and state-conditionality and probes changes in external note readout after parameter learning. • Predicts destinations in 24/24 held-out-backbone cells and gains 3.5 points over the better single destination across two environments and three seeds.
2026-10-01	SIEVE: Selective attention-value Suppression for Vision-Language Models Unlearning
• Selectively forgets sensitive information while retaining permitted facts about the same person. • SIEVE suppresses attention values for forget targets and combines frozen-retain matching with sequence objectives. • Reports improved forgetting–retention trade-offs across VLMs and modalities; evidence covers tested information and access modes.
2026-10-01	Typological Alignment of Stack-Based Language Models on Mildly Context-Sensitive Artificial Languages
• Examines how stack working-memory limits relate to word-order bias and cross-serial generalization. • Compares stack-based language models across artificial languages with hierarchical and cross-serial dependencies. • Models struggle with cross-serial structure, while limited working memory improves generalization; evidence is artificial-language simulation.
2026-10-01	Causal Memory Policy: Making Memory Utility Identifiable by Intervening on Retrieval
• Shows that utility of never-retrieved memories is unidentified by store-level interventions alone. • CMP reserves context slots for known-propensity sampling and estimates retrieval-mediated utility using inverse propensity weighting. • Required-memory discrimination rises from 0.54 to 0.66 AUC, but identified utility still fails to predict retention value on unseen queries.
2026-10-01	AutoCompact: Learning When to Compact Context in Long-Horizon Coding Agents
• Learns when to compact coding history, which working state to preserve and how to resume. • AutoCompact first trains on corrected compaction trajectories, then jointly optimizes coding and compaction with outcome rewards. • Gains 9.2 and 5.0 points on SWE-bench Verified and SWE-PolyBench Verified, including a non-overflowing 256K setting.
2026-09-30	Action Conditioned Bisimulation For GUI Agent Memory
• Merges GUI experience states by action outcomes rather than observational similarity. • Uses action-conditioned bisimulation over an empirical predictive graph, requiring agreement of outcomes and successor blocks. • Improves MiniWoB++ success while matched exploration and alternative merge controls do not; evidence is limited to that benchmark.
2026-09-30	MEMO: Multi-Level Entity-Aware Memory for Streaming Video Understanding
• Organizes streaming video memory around entities, global semantics and spatial structure. • Lightweight indices point to separately retained high-resolution evidence for query-specific recall. • Improves multiple backbones on StreamingBench and OVO-Bench without training, while retaining visual content for later access.
2026-09-30	RefCon: Iterative Refinement and Contrastive Memory Extraction for Context-Evolving Agent
• Extracts reusable memories from noisy interaction trajectories without gold labels. • RefCon combines sequential self-refinement with parallel self-contrast and scales over multiple trajectories. • Reports 21.6% and 16.6% relative gains on ACE and ReMe, while a diversity variant gains 35.5% on ReasoningBank; improvements trade off against token use.
2026-09-30	Beyond the Remembered World: Predictive 4D Belief for Persistent Navigation in Evolving Worlds
• Maintains probabilistic memory of object persistence, relocation and unknown locations from timestamped histories. • EvolvingNav propagates beliefs through time and uses visibility-calibrated positive and negative evidence without double counting. • Improves simulated and real navigation on EvoWorld-Bench; paired gains are clearest when temporal patterns are learnable.
2026-09-30	Faithful Dual-constrained Erasure for Robust LLM Safety Alignment
• Analyzes recovery of hazardous knowledge hidden by shallow suppression after benign retraining. • FDCU combines Fisher-based preservation with dual masking to prevent spurious suppressors and encourage representation erasure. • Reports stronger retraining-attack robustness with little general-utility loss in tested erasure and output-control tasks, not a formal deletion guarantee.
2026-09-30	Forking: Sudden Overfitting Under Replay
• Analyzes sudden replay overfitting caused by over-encoding continuations in n-gram memory branches. • Controlled models vary repeated updates, rare contexts and table crowding to trace seen-versus-unseen continuation probabilities. • Reproduces forking in pretraining and short repeated SFT/RL-like regimes; larger budgets or crowded tables suppress it, so replay is not universally harmful.
2026-09-30	LatentHarness: Learning Latent Actions for Memory and Reasoning via Counterfactual Policy Distillation
• Stores input evidence and intermediate reasoning in fast weights while learning when to recall versus compute. • A latent policy chooses THINK, RECALL or EXIT and learns from one-step counterfactual token effects. • At 1.4B reports 2.8% and 10.0% relative gains on general and long-context reasoning across six benchmarks and 5.9× speedup over the strongest long-context baseline.
2026-09-30	Persistent Context Graphs for Efficient Memory Compaction in LLM Agents
• Persists attention-derived importance and message dependencies in a context graph. • ReCAP combines stored scores, request relevance and dependency traversal to select messages without extra selection-model calls. • Reports approximately 95% lower estimated compaction/cold-restoration latency and roughly half the history on SWE-Together at comparable quality; latency is estimated.
2026-09-30	MemLife: Curating and Reasoning over Long-Term Egocentric Video Memories
• Stores entity-grounded first-person text episodes from long egocentric video histories. • A time-indexed agentic reader retrieves evidence, while MemOpt trains the writer for faithful, informative and retrievable records. • Reports 4.6–12.0% gains over the strongest training-free baseline and another 2.7–5.0% from writer optimization, without query-time video access.
2026-09-30	ZoneClaw: Mitigating Persistent Memory Attacks by Establishing Memory-Zoning in OpenClaw-Style Computer-Use Agents
• Separates persistence of external claims from authority to guide later computer-use actions. • ZoneClaw uses hierarchical trust zones and asymmetric processes, with cross-checks before authority promotion. • Reduces attack successes from 372/480 to 6/480 while preserving utility in 458/480 trials; defense-aware testing covers a limited set of scenarios.
2026-09-30	LabBook: Harnessing Experimental History for Efficient LLM-Driven Discovery
• Uses an evolving experimental notebook to guide selective access to complete logs and new program generation. • At each iteration the same agent combines memory and retrieved evidence to produce both a program and an updated LabBook. • Improves observed quality-cost trade-offs on 49 Frontier-CS problems and remains competitive on nine further tasks relative to tested evolutionary baselines.
2026-09-29	Eternal Sunshine of the Spotless Mind: Systematically Erasing LLM's Memories
• Studies information persistence through message dependencies after users request deletion. • DeLLM constructs query context dynamically and tracks message provenance to identify dependent content requiring removal. • Reports strong deletion with retained utility; matching individual messages is insufficient and the abstract gives no single deletion-rate figure.
2026-09-29	MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization
• Compresses user preferences and constraints into a fixed number of soft-memory vectors optimized for reader behavior. • Combines group-relative outcome rewards with confidence-gated on-policy distillation; the textual-memory teacher re-scores student tokens and is removed at inference. • Reports leading measured PersonaMem accuracy and transfer to PrefEval and LongMemEval; ablations attribute most gains to the reward term.
2026-09-29	Triadic Linear Attention: Three-Dimensional Recurrent States for Long-Context Sequence Modeling
• Stores past key-value associations in a third-order tensor with an additional address axis. • Writes a triadic outer product and reads by contracting two query axes; supports forgetting gates, delta updates and chunkwise training. • Reports stronger long-context language modeling and recall than alternative state expansions; the abstract supplies no single numerical gain.
2026-09-29	Long-Term Memory-Guided Enhancement for Target Perception in Audio-Language Models
• Stores hidden representations of clean reference audio as category-specific memory for selective listening. • Reconstructs incoming audio tokens from that memory and interpolates them before frozen language decoding; speech recovery adds a learned gate. • Reports 29.53–46.15-point classification gains across 20 categories and Qwen2-Audio WER of 14.77% versus 23.07% with the gate; evidence is controlled target perception.
2026-09-29	MemEvo: Automatic Discovery of Streaming Video Memory Mechanisms
• Searches executable bounded-memory programs for streaming vision-language understanding. • A domain language exposes representation, admission, retention, consolidation, budgeting and retrieval; deterministic evaluation checks LLM proposals. • Reports quality and efficiency improvements on StreamingBench and OVO-Bench; discovery requires candidate evaluations although the resulting mechanism is training-free.
2026-09-29	Simple Agentic Memory for Generalist Robot Policies
• Maintains entity relations, progress and ordered state to resolve history-dependent high-level subgoals. • SimpleARM writes through perception and event tools, retrieves structurally when needed and re-grounds entities in the current view, with frozen models. • Reports 67.17% RoboMME success versus 44.51% for the strongest non-oracle baseline; memory resets each episode, uses benchmark-specific parsers and lacks real-robot validation.
2026-09-29	EGSD: Event-Grounded Self-Distillation for Streaming Video Understanding
• Targets entity-coverage collapse when privileged question supervision trains streaming memory. • EGSD reweights teacher signals with verifiable events and combines entity-coverage and outcome rewards. • Reports 79.8% StreamingBench and 73.4% OVO real-time accuracy, with 17.4% higher entity recall at 6.8% more memory length.
2026-09-29	UpliftMem: Learning Set-Level Uplift for Agent Memory Retrieval
• Learns retrieval from the uplift of memory sets relative to the same executor without memory. • UpliftMem allocates training probes by expected sample information value and uses a frozen-executor scorer without test-time probes. • Leads evaluated baselines on three task families with matched-store/budget controls; its closed-form probe theory assumes a correlated Gaussian model.
2026-09-29	SafeVantage: Vantage-Aware Memory for Reliable Embodied Decisions
• Stores supporting views, poses and target locations for semantic claims, separately from search coverage. • SafeVantage predicts observability, selects views by expected decision loss and combines evidence for Yes/No/Abstain. • Controlled ProcTHOR and related tests improve F1 and risk; restoring supporting views improves VLM answers on ScanNet, within category-presence evaluations.
2026-09-29	CoEM: Empowering Long-Context Reasoning with Commit-on-Evidence Memory
• Delays irreversible conversion of source evidence into compact facts. • CoEM keeps verbatim pending excerpts and learns promotion, retention or discard decisions, checked by a frozen evidence verifier. • Outperforms the strongest memory baseline by 10.4–11.4 F1 on the reported 6,400-document input evaluation.
2026-09-29	AMU:Admission and Memory Update for Personalized Conversations---Structured Memory with SLM Guided Control
• Controls admission of transient, duplicate and outdated information into personalized memory. • AMU filters candidate records and uses an SLM to choose separate storage, duplicate rejection or update fusion. • Controlled write/retrieval experiments report cleaner, more retrievable memory; the abstract provides neither numerical gains nor long-term deployment evidence.
2026-09-29	Watch-Think-Interact: Bootstrapping Long-Horizon Multi-Turn Streaming Video Reasoning with Reinforcement Learning
• Connects compact time-tagged textual memory to selective recall of original video evidence. • WTI chooses to answer, watch or revisit an interval; Stream-GDPO jointly trains timing, recall and memory updates. • Trained with WTI-82K, reports 83.3% StreamingBench and 73.6% OVO-Bench weighted accuracy; the method requires access to historical clips.
2026-09-29	When Should Agents Check External State? Budgeting Observations for Stored Intentions
• Budgets external observations needed to decide whether stored future intentions should fire. • BudgetPM combines local utility gating or a hindsight-distilled sequential policy with a common hard-budget executor. • Static gating retains 99.9–100% quality with 42–54% fewer observations; sequential scheduling gains 1.92–2.58 Set F1 under scarcity.
2026-09-29	ReMem: Rethinking Perception and Memory in Long-Context Recommendation Agents
• Maintains fixed-size evolving preference memory while processing long interaction histories in chunks. • Selectively retains informative interactions and propagates final-answer credit to intermediate updates with multi-memory GRPO. • Reports 5.16% average improvement across three recommendation tasks; the system also includes OCR perception, so total gains are not solely memory gains.
2026-09-29	Learning to Retrieve Missing Evidence for Long-Term Memory QA
• Uses already recovered evidence to expose clues omitted from the original memory question. • MERA separates globally searchable memory from a question-specific evidence state and trains a small planner to recover missing support. • With Qwen3-30B, the 0.6B planner reaches 77.40% LoCoMo and 71.29% LongMemEval-S; results depend on the evidence and answer backbone.
2026-09-29	How Can Recommendation Feedback Evolve Agent Memory?
• Maintains a capacity-limited population of experiential memories using delayed recommendation feedback. • TIDE assigns temporal, semantic and responsibility credit, then reinforces, combines, mutates or evicts records; MEG uses strictly future tasks. • Reports a 7.75-point offline MEG gain and improved online click/activation rates; attribution depends on recommendation context and assumptions.
2026-09-29	ReLMem: Learning Recurrent Memory for Longitudinal EHR Modeling
• Recurrently compresses successive visits into fixed-capacity patient memory for a frozen LLM. • Compression adapters learn from intermediate attention-output alignment and final prediction supervision without rereading old records. • Reduces retained storage by 97.1% and improves macro/micro F1 by 4.66/4.75 points over the strongest matched-budget baseline; this is medication prediction, not clinical efficacy.
2026-09-29	VirusCascade: Hijacking Collaborative Reflection in LLM-Powered Recommender Agents
• Studies malicious evidence laundered into preference memory and propagated through collaborative reflection. • VirusCascade shapes semantic narratives and structural placement, supported by controlled persistence and propagation analyses. • Reports mean E@20 of 0.384, exceeding the strongest baseline by 0.185 across four datasets under evaluated stealth constraints.
2026-09-29	EpiCon: Collective Agent Learning through Co-Evolving Multimodal Memory
• Shares execution experience through a persistent multimodal bank across agents. • Separate 2B controllers refine visual/textual evidence and hierarchically consolidate and retrieve lessons while host weights stay frozen. • Across 11 benchmarks reports 1.7–4.9-point gains over no memory and 67–74% less memory-operation time; transfer depends on bank and interface.
2026-09-29	Learning What to Remember: Long-horizon Counterfactual Memory Optimization
• Learns retention by the incremental current and future value of each textual-memory rewrite. • MGPO uses counterfactual credit to separate new utility from benefits already inherited from old memory. • Improves extraction while reducing average memory length by nearly 80% versus the initial policy, with transfer across domains and readers; primary evidence is structured extraction.
2026-09-29	VideoLoop: Looped Working Memory Against Semantic Thrashing in Long-Form Video Agents
• Studies semantic thrashing when append-only working memory obscures previously found video evidence. • VideoLoop retrieves from a filesystem of observations and analysis and rewrites a bounded working memory after each outer reasoning step. • Reports a 4.2-point average VideoMME-long gain; a context-only judge scores 81.1% versus 60.9% on the hardest question quarter.
2026-09-29	Beyond the Timeline: Augmenting Long-Video Memory with Grounded Entity Biographies
• Builds retrievable biographies linking the same physical entity across video clips while preserving event context. • GEB establishes grounded identity links during memory construction and jointly retrieves biographies and episodic evidence. • Reports 72.0% EgoLifeQA accuracy and gains on four day/week-scale benchmarks; ablations support both identity association and biography reading.
2026-09-29	What Pretraining and Midtraining Make Learnable from Rewards?
• Studies how source prediction supplies retrieval that rewards alone cannot identify. • Constructs finite Adam paths from source prediction to reward adaptation and uses replay to retain contextual retrieval. • An independent eight-world experiment reaches 75.32% success versus 49.86% after alternative-retrieval training; controlled worlds do not establish open-task guarantees.
2026-09-29	Personalized State-Transition-Aware Memory for Clinical Agents
• Tracks patient-state transitions while separating active information from superseded historical evidence. • STAM combines semantic retrieval with typed clinical relations and query-gated access to Active and History stores. • Evaluates QA and direct maintenance diagnostics on four longitudinal benchmarks at approximately matched context lengths; the abstract gives no numerical outcome or clinical validation.
2026-09-28	CyFA: Linear Sequence Modeling with Relative-Time-Partitioned Memory
• Partitions stored key-value associations by relative age to reduce interference in fixed recurrent state. • A learned clock transports key and value states cyclically, inserting new associations at age zero; an exact coordinate change enables chunkwise training. • At matched state size, reports stronger recall; 400M FDA is 42.60 versus KDA 26.07, within the evaluated pretraining and retrieval settings.
2026-09-28	Fractional State Space Transition for Long Sequence Modeling
• Replaces exponential forgetting with power-law memory to retain information across wider timescales. • FRAC approximates a fractional kernel with log-spaced exponential modes, enabling parallel training and bounded-state decoding. • Reports improved long-context performance, including 1.3B language modeling, with competitive short-context results; the finite modes approximate the heavy-tailed kernel.
2026-09-28	Audience-Bound Persistent Memory: Authorization Across the Memory Lifecycle
• Binds original and derived memories to authorized audiences across private and shared conversations. • Partitions derivations or intersects source audiences; widening requires explicit grants and unresolved viewers receive public-only content. • Admits no forbidden items in 10,000 frozen histories; guarantees assume correct identity, provenance and complete mediation, and rare baseline events prevented one joint conclusion.
2026-09-28	Thinking Outside the Box: Retention and Transmission of Information in Sliding-Window KV Inference
• Studies information retained after source tokens leave a sliding KV window. • Compares cached-state reuse with recomputing the final raw-token window across five models. • Finds latent information relay, strongest in two models using sliding-window attention; causal training explanations remain open.
2026-09-28	From Attack Success to Attack Severity: Counterfactual Memory Attacks on LLM Agents
• Measures persistent memory attack severity through counterfactual regret. • MemHarm selects grounded semantic edits using paired downstream-loss feedback. • Severity-guided selection causes greater downstream loss than success-only optimization on matched attack support.
2026-09-28	PersMem: Internalizing Personality into Dual-Pathway Memory for LLM Agents
• Makes personality influence memory processing directly. • PersMem conditions appraisal, retention, passive retrieval, and goal-driven selection on personality. • Produces distinguishable personality-related traces and improves tested dialogue metrics; this measures behavioral consistency.
2026-09-28	Just-In-Time Agent Memory with Runtime Agentic Research
• Constructs query-specific memory context at runtime. • JAM preserves raw histories in a hierarchical page store and trains an evidence-seeking researcher. • Reports stronger task performance than ahead-of-time construction with lower cost than prior trained agentic approaches.
2026-09-28	Remember by Asking: Retrieval-Induced Memory Evolution for LLM Agents
• Builds memory through evidence-focused self-questioning and retrieval. • RIME reconciles retrieved dialogue with prior memory and falls back to source context when needed. • Leads compared methods on three LoCoMo quality metrics with fewer query-time tokens.
2026-09-28	Remember Before You're Asked: MemDream for Self-Probing Memory Evolution
• Proactively probes and repairs memories before real questions expose failures. • MemDream uses offline dream cycles, learned repair policies, and reversible decay. • Reports gains of 4.5 F1 points on LoCoMo and 9.1 overall points on MemoryAgentBench.
2026-09-28	Single-Layer MeMo as a Randomized Hamming-Kernel Classifier
• Analyzes single-layer MeMo as a randomized associative classifier. • Derives context-sketch and output-decoding errors relative to a Hamming-kernel predictor. • Simulations and restricted WikiText-2 tests support the analysis; results do not establish full-scale language-model parity.
2026-09-28	FlowState: Execution State as Memory for Long-Horizon LLM Agents
• Preserves execution state as revisitable long-horizon memory. • Typed nodes retain relations and raw-observation references with incremental updates and progressive access. • Improves MemoryArena and tau3-Bench by 4.55 and 13.95 points while reducing tokens by about 40%.
2026-09-28	GenMem: Generative Symbolic Memory for Self-Evolving Harness
• Treats evolving memory access as generative symbolic addressing. • Stable hierarchical identifiers decouple retrieval policy from continually revised memory payloads. • Evaluates offline evolution across interactive and reasoning tasks; conclusions are limited to tested memory settings.
2026-09-28	EP-Mem: Elastic Privacy Memory for Social Relationship-Aware LLM Agents
• Controls memory disclosure according to user-owned social boundaries. • Combines person/event policies with generation, storage, and retrieval enforcement. • Reports 75.6% lower leakage and improved permission judgments on introduced and transfer evaluations.
2026-09-28	Sprout: Building Dynamic Memory While Reasoning for Agentic Video Understanding
• Builds video memory incrementally while answering questions. • A persistent temporal tree grows from coarse observations and targeted high-resolution revisits. • Matches or improves tested offline-memory methods with lower context cost and no upfront full-video construction.
2026-09-28	From Experience to Expertise: Adoption-Aware Memory Learning for Data-Scarce NPU Kernel Synthesis
• Learns reusable programming memory from actual experience adoption. • SAGE assigns adoption-aware utility and consolidates repeatedly useful rules into bounded context. • Reaches 95.5% execution on NPUKernelBench versus 84.1% for the strongest controlled baseline.
2026-09-27	How Linear Attention Remembers
• Explains selective recall and interference in linear-attention states. • Combines analytical decomposition with causal write, retention, and read interventions in pretrained models. • Memory load harms recall more than elapsed context in tested regimes; hybrid recall often shifts to full-attention state.
2026-09-27	Beyond Memory Construction: Rethinking Memory Access for LLM-based Conversational Agents
• Preserves raw interactions as first-class conversational memory. • Threader performs incremental topic segmentation and multi-view, localized evidence retrieval. • Reports stronger answer accuracy and evidence recall with lower construction overhead than rewriting-based methods.
2026-09-27	ActiveMem: Dynamic Latent Memory Trees for Long-Horizon Agents
• Organizes procedural experience as dependency-aware memory trees. • ActiveMem learns state-conditioned path retrieval and memory expansion and pruning with reinforcement learning. • Reports better completion, stability, and efficiency across tested agent benchmarks.
2026-09-27	Grounding Memory Summarization in Utility Intent
• Aligns memory summaries with downstream evidence utility. • MemSuit distills query-informed teacher entries and adapts retrieval to compact facts. • Reports gains across conversational query types; the student builds memory from raw conversations alone.
2026-09-27	TRACE: Governing Memory Validity in Evolving Multi-Agent Systems
• Governs whether returning agents may act on previously valid memory. • TRACE reconciles departure checkpoints with intervening updates and releases a bounded eligible view. • Preserves valid information while rejecting stale state; gains incur additional token cost.
2026-09-27	Positions Are Not Facts: The Mismatch Between KV Caches and Memory
• Examines semantic updates when KV caches serve as memory. • Compares masking values or records with deletion and recomputation, including dependency preservation. • Updating values can damage units and historical answers; cache positions do not directly encode editable facts.
2026-09-26	Authority Before Utility: Non-Compensatory Control for Persistent LLM Memory
• Separates memory utility from authority after updates, deletion or revocation. • Compares hard exclusion with a development-selected normalized soft penalty on Memora Remembering. • Reports 4.04% versus 19.66% balanced error over 185 held-out units; the original primary was quarantined and this replacement diagnostic is not a universal impossibility result.
2026-09-26	KeyRec: Bounded Visual Memory for Streaming and Long-Video Understanding
• Builds bounded visual memory for streaming and long-video questions. • Combines a detailed recent cache, an event bank, and query-dependent budget allocation. • Leads compressed baselines in 13 of 15 settings at 10% of the dense visual-token budget.
2026-09-26	BMA: Backchain Memory Attacks Create Unauthorized Control Paths in LLM Agents
• Studies indirect attacks that turn low-trust evidence into actionable persistent memory. • BMA plans backward from a target action and certifies the evidence-to-memory pathway. • Reports 18.8% certified success versus 13.4% for the strongest matched baseline; authorization controls reduce harm.
2026-09-26	Agentsensus: Consensus-Compressed Shared Memory for Multi-Agent Story Worlds
• Compresses shared story events into jointly owned memory records. • Agentsensus merges witnessed events and links related records across character agents. • Writes 22–44% fewer entries with comparable judged simulation quality; sharing grows with horizon.
2026-09-26	Delayed Supervision for Test-Time Language Models
• Trains test-time memory to preserve facts across delayed queries. • Supervises questions on disposable branches after unrelated events, separating retention from revision. • Improves recall benchmarks for several recurrent models; delay-specific benefit remains confounded with general QA supervision.
2026-09-26	ALLOT: Budgeted Hybrid-Memory Routing for Knowledge Updates in LLMs
• Allocates limited parametric writes by their incremental value beyond retrieval. • ALLOT keeps all facts externally and ranks updates using text, retrieval confidence, and relation metadata. • At a 20% write budget on CounterFact, reaches 0.760 accuracy and 78.4% of budget-matched oracle gain.
2026-09-26	Shared Worlds, Private Minds: Structured Memory for Long-Form Writing as World Creation
• Builds structured storyworld memory for consistent long-form writing. • NarraWorld separates facts, beliefs, open developments, and hypothetical branches with traceable aggregation. • Leads aggregate results on three writing benchmarks and transfers to role-playing and general recall.
2026-09-26	Learning from Others, Acting for You: Cross-User Memory Sharing for LLM Agents
• Shares reusable experience while retaining receiving-user preference boundaries. • ShareMem separates procedural guidance from preference values and uses two-stage consolidation. • Improves four backbones across three domains; source quality and preference interference still limit transfer.
2026-09-26	MemAgent: Learning to Manage Heterogeneous Memory Providers for LLM Agents
• Routes across heterogeneous memory providers instead of relying on one representation. • MemAgent learns probing, short-term gating, and selective storage decisions. • Across three benchmarks, reports 10.0% average accuracy improvement with under 0.3% routing overhead.
2026-09-26	Beyond Dyadic Memory: Interaction-Aware Multimodal Memory with Adaptive Agentic Retrieval for Multi-Party Spoken Conversations
• Maintains interaction-aware memory for multi-party spoken conversations. • VoxPolyMem combines speaker identification, layered memories, and evidence-gain-trained retrieval decisions. • Scores 85.0 on VoxPolyBench, 23.6 points above the strongest evaluated baseline, with transfer gains on two datasets.
2026-09-26	ProTTT: Learning to Learn Semantic User Memory with Test-Time Training
• Learns evolving semantic user memory through test-time training. • ProTTT meta-learns profile-supervised initialization and updates compact user-specific parameters from history. • Outperforms tested context and parametric baselines while reducing inference cost and tracking changing preferences.
2026-09-26	EMIR²: Evolution-Aware Memory with Intent-Guided Multi-Round Retrieval
• Tracks evolving knowledge and adaptively gathers missing evidence. • EMIR2 combines a state-evolving memory graph with intent-guided multi-round retrieval. • Improves conflict handling and complex retrieval on LoCoMo and MemConflict, with gains varying by category.
2026-09-26	Retrieved but Not Delivered: Multimodal Memory Delivery for Long-Term Agents
• Separates retrieved evidence from how it is delivered to multimodal models. • DeliverMem preserves original modality, item identity, and timestamps under matched retrieval controls. • Delivery has a larger measured effect than perfect retrieval on MemLens and improves two tested benchmarks.
2026-09-26	Contract Memory Compiler: Resolve, Then Traverse
• Resolves evolving relations before choosing evidence for multi-hop memory queries. • CMC extracts provenance-linked relations, applies updates, and traverses current links to raw records. • Reaches 78.25% multi-hop accuracy on FactConsolidation; selecting evidence before updates drops it to 21.50%.
2026-09-26	Self-Evolving Time-Series Forecasting Agents with Episodic Memory and Online Policy Learning
• Reuses delayed forecasting feedback through episodic memory and online guidance. • FASE retrieves completed cases and summarizes accumulated outcomes into action-ranking policy. • Across 29 configurations, reduces normalized MAE by 9.1% versus the best individual foundation-model baseline.
2026-09-26	Decision-Sufficient State Representations: Measuring and Reducing Write-Time Regret
• Measures information lost by sequential memory rewriting as write-time regret. • DSSR trains writers using reader outcomes after forward-rolled state updates. • Adds seven success points at short delays, but gains shrink at longer delays because credit assignment remains difficult.
2026-09-26	The Key Handoff: Retrieval in Hybrid Language Models
• Traces how hybrid models hand off retrieval keys between layers. • Hidden-state transfers and recurrent-memory edits distinguish entity-addressed retrieval from positional cues. • Attention converts keys while recurrence can carry and later query them; behavior varies across twelve models.
2026-09-25	HasMem: Hard-Origin Adaptively Softened Memory for Long-Term LLM Agents
• Adapts continuous memory capacity from a verifiable hard-prompt initialization. • A controller resizes entries while writer, reader, and global state adapt encoding and readout. • Improves lexical F1 in two evaluations, but exact-match accuracy declines, limiting a simple success interpretation.
2026-09-25	COUNTERMEM: World-Model Verified Counter-Factual Memory for Language Agents
• Constructs reusable counterfactual memories from verified alternative actions. • Executable world models test local corrections; a learned policy selects or skips memories. • Reports a 12.6-point average gain over unaugmented agents across twelve settings; offline training costs are excluded from token savings.
2026-09-25	EngramRAG: Dynamic Usage-Weighted Topology and Synaptic Consolidation for Multi-Hop Agentic Memory
• Introduces adaptive graph memory with background consolidation. • Combines usage-weighted PageRank, topology-aware retention, revision filtering, and hybrid retrieval. • On LoCoMo, Recall@5 rises from 38.29% to 53.21% over dense retrieval; retention tests include simulations.
2026-09-24	Scope Before You Persist: Preventing Cross-Family Interference in Agent Memory
• Scoped-ORC limits persistent skill edits to the task families for which they were certified. • A frozen-model repair stream separates execution-based acceptance from the scope of later retrieval. • Across 27 paired streams, utility improves by 0.063 and none of 63 accepted updates is harmful under the reported evaluation.
2026-09-24	ALOE: Semantically Addressed Low-Rank Operators for Knowledge Editing
• ALOE treats knowledge editing as both a memory write and a semantic addressing problem. • Paraphrases and hard negatives train a gated low-rank operator embedded inside one MLP layer without an external retriever. • Across three model families, reported efficacy is 0.955–0.999 and locality 0.981–1.000; errors remain in paraphrase coverage and write fitting.
2026-09-24	C3M: Cross-Session Multimodal Memory Maintenance for Long-Horizon Tasks
• C3M maintains a bounded multimodal memory index over persistent cross-session evidence. • Relation-aware consolidation retains complementary or conflicting observations, while budgeted routing expands linked source evidence. • The design preserves temporal distinctions and provenance; the abstract provides no quantitative benchmark results.
2026-09-23	CAVE-Mem: Boundary-Aware Experience Validation for Memory Search
• Represents past search experience as intervention operators with explicit applicability boundaries. • Checks substrate, answer contract, evidence boundaries and cross-fitted utility before revising a base answer; abstains on mismatch. • Reports gains over relevance-only reuse across conversational memory, multi-hop QA and narrative reasoning; the abstract gives no common numerical gain.
2026-09-23	Self-Evolving Multimedia Verification through Memory Consolidation of Contestation Experiences
• SEMV consolidates verified contestation experience while retaining conflicts and provenance. • Argument-based reasoning supports scoped causal revision and gates which experiences enter reusable memory. • Reported COSMOS accuracy is 91.88%; verified memory reduces negative transfer from 5.7% to 0.2% in the evaluated setting.
2026-09-23	EnSIMem: Entity-Structured Indexing for Long-Term Agent Memory
• EnSIMem organizes long-term agent memory around identifiable entities and properties. • Episode-based indexing preserves source turns and temporal evidence, then aligns query requirements with entity-property lookup. • The authors report accurate retrieval-based answers with compact contexts; the abstract does not quantify the gains.
2026-09-23	Memory Control Signals Emerge Before Action in Long Horizon Agents
• PaMER uses pre-action hidden states to infer when an agent needs compression or recall. • State-guided compression is combined with external evidence retrieval; PaMER+ selects evidence for individual steps. • WorkBuddyBench experiments report lower context consumption with competitive task performance; signals are not explained by length alone.
2026-09-23	Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents
• JitMem postpones experience curation until the current task is known. • A curator converts retrieved raw trajectories into task-specific memory and learns from immediate task success. • Reported gains over the strongest baseline are 16.2, 16.3 and 3.9 success-rate points on ALFWorld, WebShop and τ²-bench.
2026-09-23	NavProbe: Evidence-Grounded Reasoning with Active Memory Retrieval for Zero-Shot Navigation
• NavProbe retrieves historical navigation evidence when compact context cannot support the next decision. • A place-transition-landmark index links summaries to visual and geometric records and supports dynamic subgoal revision. • Reported success is 71.7% on R2R-CE, 55.3% on RxR-CE and 79.3% on HM3D-v2 ObjectNav; physical demonstrations are qualitative.
2026-09-23	Learning from Failures: Heterogeneous Graph Memory for Small Language Model Tool-Using Agents
• FRESH represents successful and failed tool-use experience as heterogeneous graph memory. • Relations among tasks, actions, errors, repairs and preconditions preserve the context needed for safe reuse by frozen models. • Experiments on τ-Bench and AppWorld report higher success and tool reliability across open models; the abstract omits effect sizes.
2026-09-23	Complementary Roles of Activation and Parametric Memory in Few-Shot Learning
• Controlled experiments distinguish activation memory from parameter-based memory in few-shot learning. • The study compares factual recall and task learning, then analyzes neurons recruited by each memory route. • Activation memory favors recall, while Conditional Arithmetic benefits from both routes; parameter updates do not consistently dominate task learning.
2026-09-23	OCC4M: Object-Centric 4D Memory for Spatiotemporal Reasoning in Long-Horizon Manipulation
• OCC4M provides object-centered memory for reasoning about states missing from the current view. • Persistent 4D tracks encode identity, motion and containment for VLM target selection. • Simulation reaches 96.6% memory and 88.9% task success; physical two-stage task completion is 45%, showing remaining execution limits.
2026-09-22	ChipMEM: Verification-Grounded Memory for EDA Agents
• ChipMEM separates transferable procedural memory from within-trajectory recovery statistics for EDA agents. • Skills are stored after synthesis, simulation or formal checks, while Bayesian estimates rank recovery strategies. • Equivalence-passing designs increase from 35/54 to 39/54; the held-out 20-task result is based on a single evaluation per setting.
2026-09-22	Rewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMs
• This study examines how instruction tuning changes competition between parametric memory and conflicting prompts. • Five attribution and intervention methods compare base and instruction-tuned models across three families. • Results favor reweighting existing late-layer heads rather than replacing circuits; robustness to terse false context disappears under richer evidential framing.
2026-09-22	When Does Execution Provenance Help Agent Memory Retrieval?
• This work measures whether agent memory retrieves complete execution evidence within a token budget. • Source-aligned provenance units improve candidate granularity, while a residual R-GCN refines retrieval over typed execution edges. • Across 2,000 queries, units add 19.07 Full Support@2048 points over flat windows; graph propagation adds a smaller 4.55 points under matched candidates.
2026-09-22	SpeakerMem-R1: Speaker-Centered Dual-Track Memory for Multi-Party Dialogue
• SpeakerMem-R1 separates speaker-attributed raw messages from structured conversational states. • Person and group views are combined at retrieval; a locally deployable writer is trained with speaker-aware rewards and GRPO. • Reported binary accuracies are 47.9%, 69.2% and 61.9% on GroupMemBench, SocialMemBench and EverMemBench; protocols differ from leaderboard reports.
2026-09-20	PSD: Pseudo Self-Distillation of Memory Representation Capabilities for LLM Agents
• PSD transfers hierarchical memory-construction behavior from a black-box oracle into small language models. • The same small model acts as a privileged-prompt teacher and ordinary-prompt student, without access to oracle logits or hidden states. • Qwen3 models from 0.6B to 4B match or exceed GPT-4.1-mini on LoCoMo retrieval and transfer to LongMemEval without training on it.
2026-09-20	Propose, Verify, Commit: Evidence-Grounded Memory for Long-Horizon Multi-Actor Conversations
• EGMEMORY models multi-actor conversational memory as evidence plus an explicit active state. • An evidence-grounded propose–verify–commit protocol governs writes, while adaptive navigation retrieves state and supporting messages. • Reported scores are 68.2% on GroupMemBench, 77.9% on EverMemBench and 73.6% on LoCoMo; code release is deferred to formal publication.
2026-09-20	PREM: Prefix-Steered Recurrent Memory for Long-Video Understanding
• PREM stores long-video evidence in a compact recurrent state for frozen vision-language models. • A recurrent writer produces a 256 KiB associative memory; question-conditioned readout modulates existing prompt-prefix keys and values. • Across six benchmarks, the 16-frame setting improves macro accuracy by 3.06%; memory-related tuning covers 0.24% of backbone parameters.
2026-09-20	Total Cost of Agency: Exact Attribution of Memory Injection Cost in Multi-Agent LLM Workflows
• Total Cost of Agency isolates the token cost of injecting retrieved memory into multi-agent workflows. • A two-pass token count separates memory injection from other workflow costs on a 200-task enterprise benchmark. • Memory injection accounts for about 12% of billed cost; reducing retrieval capacity lowers injected tokens 28.7%, while caching is not evaluated.
2026-09-18	MACE: Memory-Agent Co-Evolution with Adaptive Memory Graphs for Multi-Agent Systems
• MACE co-evolves multi-agent procedural memory and how agents consume it. • MemGoG stores conditions, actions and outputs as linked functional subgraphs; execution feedback adjusts retrieval and instruction/checklist presentation. • Across eight benchmarks, the reported average score is 81.11% versus 78.97% for the strongest compared baseline.
2026-09-18	AutoViewMem: Self-Configuring Orthogonal Views for Conversational Long-Term Memory
• AutoViewMem organizes conversational memory into complementary semantic views before indexing. • It discovers views from interaction traces, extracts provenance-grounded records at write time, and consolidates them offline for ordinary top-K retrieval. • Experiments on LoCoMo and PersonaMem with Qwen3-8B/14B report improved long-horizon QA and personalization without iterative retrieval.
2026-09-18	An Interpretable Memory Decision Controller for LLM Agents Based on Three-Signal Complementarity: Decoupling Confidence and Consistency
• The Memory Decision Layer assesses whether retrieved memories should influence generation under conflict. • A parameter-free controller combines relevance, reliability and task risk through geometric projections, explicit abstention and risk inversion. • The abstract reports about 56.04% lower hallucination rates in general conflict settings and 0.14 ms per decision; claims depend on the evaluated settings.
2026-09-18	Designer-RSI: Evolving Procedural Memory from User Traffic for Agentic Graphic Design
• Designer-RSI accumulates reusable graphic-design procedures in external memory while keeping the model frozen. • The skill bank expands for uncovered subtasks and revises procedures using execution feedback, with a matched replay gate checking regressions. • Five rounds grow 76 skills to 139 and raise reported GenEval2 execution success from 72.7% to 99.3%; feedback is automatically graded.
2026-09-17	LearnActCoder: Role-Aware Error Memory for Adaptive Clinical Coding Agents
• Introduces LearnActCoder, which converts labeled-case errors into reusable role-specific memory for clinical coding agents. • Routes false-negative lessons to a recall-oriented Coder and false-positive lessons to a precision-oriented Judge through a MistakeKDB. • Retrospective MIMIC experiments show task-dependent gains, including improved CPT F1; ICD results and precision–recall trade-offs limit broad claims.
2026-09-17	JustMem: Just-Enough Memory Access for Long-Term Conversations
• Introduces JustMem, which adapts both discovery breadth and reading fidelity for long-term conversational memory. • Stores compact atomic memories and chooses among LOOKUP for local facts, COMPOSE for distributed evidence, and REPLAY for original context. • On LoCoMo and LongMemEval-S, reports the highest mean accuracy and recall among compared methods while reducing construction and inference tokens.
2026-09-17	AdaRepair-Mem: Adaptive Experience Orchestration for Repository-Level Program Repair
• Introduces adaptive orchestration of repair experiences to address sparse coverage, noisy retrieval, and phase imbalance in coding-agent memory. • Combines coverage-aware fallback, quality-aware ranking, and stage-specific retrieval across reproduction, localization, patching, refinement, and validation. • SWE-bench Lite and Verified experiments report improvements on under-covered repositories and more effective reuse than indiscriminately adding memories.
2026-09-16	Rollback the World, Keep the Reflection: Rollback-Induced Reflection for Long-Horizon LLM Agents
• Introduces rollback-induced reflection, allowing long-horizon agents to restore an earlier environment state while retaining lessons from failed trajectories. • Jointly determines intervention timing, rollback depth, and which reflection memories survive the rollback. • Across three long-horizon benchmarks and multiple backbones, reports benefits from selective rollback combined with reusable reflection.
2026-09-16	Disentangling Long-Term Memory via Latent Neuro-Symbolic Reasoning
• Introduces a latent graph memory for long-term personalization, representing conversational history through disentangled concepts. • Uses sparse autoencoders to form concept nodes and query-conditioned edges and graph reasoning to retrieve relevant memory. • Reports improvements in long-term personalization involving both explicit and implicit preferences, supporting adaptive latent memory over static organization.
2026-09-16	Long-Lived Characters, Local Inference: Incremental Memory Maintenance for Game NPCs
• Studies incremental memory maintenance for persistent game characters using locally deployed hybrid recurrent-attention language models. • Removes superseded KV entries and computes replacements at the true sequence tail, preserving recurrent state and unaffected cache content. • Eight scripted maintenance rounds favor true-tail updates; composing independent blocks or subtracting states can disrupt memory bindings and query selection.
2026-09-16	Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection in Interactive Environments
• Separates episodic memory and execution-time self-reflection in a dual-process language agent built on SwiftSage. • Adds salience-gated memory with triggered retrieval and a bounded self-reflection module for validating and correcting execution. • Controlled ScienceWorld ablations favor the combined system, while self-reflection is the strongest standalone addition and memory benefits depend on execution stability.
2026-09-15	On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models
• This work separates weight memorization from in-context retrieval when analyzing state-space-model gates. • Theory and controlled experiments show that gates can favor memorized solutions and delay in-context learning even without capacity limitations. • Gating can still improve long-sequence generalization, exposing a training-dynamics trade-off rather than a blanket architecture limitation.
2026-09-15	EchoPath: Execution-Level Replayable Memory for GUI Agents
• EchoPath converts validated GUI trajectories into parameterized, callable execution memories. • Stored preconditions, visual evidence and provenance guide image-based target re-aiming; incompatible replay falls back to bounded repair or planning. • Real computer-use experiments report over 90% lower median token cost and about 60% lower median execution time for recurrent tasks.
2026-09-15	Memory-Skill Isomorphism: One Skill Carrier, Two Native Uses
• Memory-Skill Isomorphism proposes a shared carrier for distilled memory and reusable procedures. • A compact resident description, on-demand skill index and detailed reference files provide progressive disclosure, with distinct governed write paths. • Evidence comes from one deployed system and a selected-task experiment; observed failure reductions are exploratory rather than confirmatory.
2026-09-15	ThinkFlow: Self-Evolving Probabilistic Latent Memory for Lifelong Conversational Agents
• ThinkFlow represents evolving user states as probabilistic latent memory instead of explicit textual records. • Teacher-guided alignment initializes memory, followed by self-supervised next-user-utterance prediction for test-time personalization. • The abstract reports gains on long-term conversation benchmarks but gives no numerical results or deployment-scale validation.
2026-09-15	Interactive Memory Learning for Long-Term Conversations
• ICML learns conversational-memory storage and retrieval policies through interaction. • Planner and Trigger agents co-adapt using online reinforcement learning, with delayed rewards connecting future responses to earlier storage choices. • The authors report stronger long-term conversational responses as interactions accumulate; the abstract does not give numerical effect sizes.
2026-09-15	Collaborative Memory for Multi-Agent VLM Systems
• Proposes collaborative memory for multi-agent VLM systems that need a coherent shared understanding of visual environments. • Organizes observations, interpretations, and reasoning dependencies through memory hierarchies, sharing policies, and consistency mechanisms. • The framework identifies how new evidence should update dependent beliefs; the abstract presents design arguments rather than quantified controlled improvements.
2026-09-14	EMR: Self-Evolving Medical Multi-Agent System via Experience Mining and Reuse
• EMR accumulates reusable clinical reasoning experience for medical multi-agent systems. • A hierarchical library stores principles, diagnostic patterns and representative cases, adding successful insights and failure warnings from trajectories. • Medical reasoning benchmarks report improved performance and cross-specialty/backbone transfer; these are benchmark results, not clinical deployment validation.
2026-09-14	RSIAgent: Autonomous Exploration for Recursive Self-improvement in New Environments
• RSIAgent autonomously constructs reusable environment-specific memory without weight updates. • Curriculum, actor and verifier agents explore broadly and then deeply, retaining validated causal relations and constraints for later tasks. • The authors report gains on OSWorld-v2 and Agent's Last Exam; downstream memory is frozen before reuse.
2026-09-14	CoMem: Collective-Individual Memory Synergy for Evolutionary Multi-Agent Systems
• Combines individual experience with curated collective memory for multi-agent learning. • Private consolidation and dual-stream retrieval preserve diversity while sharing validated knowledge. • ALFWorld and PDDL experiments report strong performance and reduced memory pollution.
2026-09-14	AnchorGUI: Asymmetric Memory for Dual-Scale Learning in GUI Navigation
• Uses asymmetric visual memory for GUI correction and cross-trial learning. • Prediction errors preserve screenshots selectively and focus procedural credit assignment. • Four benchmarks report improved success with sublinear context growth.
2026-09-14	MessyMem: Learning-from-Doing Memory for Mobile Manipulation
• Retains interaction-derived knowledge for repeated mobile-manipulation tasks. • A grounded 3D scene graph links object properties, outcomes, and visual observations. • Simulation and robot tests demonstrate reuse across long task sequences.
2026-09-13	The Immutable Past: Formalizing State Mutability and Conflict Resolution in Mutable RAG
• GC-Mem addresses obsolete facts that overwhelm current state in append-only agent memory. • Temporal dominance and contradiction detection remove superseded evidence while preserving unrelated valid long-term records. • The authors report over 90% conflict-resolution accuracy on their accumulation benchmark, with explicit precision/recall conditions for deployment.
2026-09-13	Pull: Lazy Materialization of Working Memory for Stateful LLM Conversations
• Materializes conversational working memory lazily and reversibly. • A deterministic directory lets the model expand only relevant historical turns. • LoCoEval reports over 70% fewer query-context tokens without measured quality loss.
2026-09-12	When Malicious Instructions Persist: Persistent Memory Poisoning Attack on Harness-Based Agents
• Studies external instructions that persist through agent-memory writes. • Cross-session evaluations vary agent harnesses, models, modalities, and triggers. • Prompt defenses reduce some injection but offer limited protection after poisoning persists.
2026-09-12	LIMBO: Lifelong Inference-Time Memory and Budget Optimization for LLM Agents
• Jointly allocates memory replay and inference budgets for lifelong agents. • An online policy adapts resource choices without changing agent weights. • LifelongAgentBench tests report better cost–accuracy tradeoffs across three backbones.
2026-09-11	CueMem: Cue-Guided Context Reconstruction for Long-Term Conversational Memory
• Treats stored memory records as cues for reconstructing source context. • Turn-graph expansion recovers temporal and semantic evidence around retrieved anchors. • LoCoMo and LongMemEval improve over memory baselines with lower cost than full history.
2026-09-11	LifeFuse-Mem: Lifecycle-Aware State Fusion Against Temporary Overwriting for Long-Term Memory
• Separates temporary information from durable memory commitments. • Lifecycle-labeled training and phase-aware readout guide dedicated memory components. • Controlled tests reduce overwriting; two public benchmarks remain broadly competitive.
2026-09-11	LifeMem: Enabling Lifelong Experience Reuse for LLM Agents
• Supports lifelong reuse of agent experience across environments. • Clusters trajectories by workflow and retrieves reusable skills for new tasks. • Tests across ten environments report reduced forgetting and improved transfer.
2026-09-11	BadEngram: Backdoor Attack on Gated Memory Components in LLMs
• Studies backdoors implanted solely in gated parametric memory components. • Controlled interventions isolate the memory pathway while backbone weights remain unchanged. • Experiments demonstrate trigger-dependent attacks in controlled and production-scale models.
2026-09-10	Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents
• Grounds post-task memory curation in fresh environment observations. • Read-only tools verify, scope, and refresh candidate memories without retraining. • CLBench and adapted APEX tests improve rewards and reduce task-agent costs.
2026-09-10	2AM: Grounding Agent-Side Memory as Guidance for Steerable Action Models in Long-Horizon Manipulation
• Keeps manipulation task memory in an agent outside the action policy. • Subtask language and optional 2D hints steer a stateless RGB action model. • LIBERO-Mem tests improve completion, with strict success remaining substantially lower.
2026-09-10	MAPLE: Memory-Augmented Planning with Language and Evolution
• Maintains optimization problems across successive natural-language revisions. • Retains executable programs, accepted plans, updates, and candidate solutions. • NLDO experiments show valid updates and reuse of earlier search results.
2026-09-09	ROAM: Robust Organization of Atomic Memories for Agents through Semantic Relations
• Organizes atomic memories through explicit semantic relations. • Primary and evidence roles separate active views from redundant or conflicting observations. • Experiments report better answers, higher critical-evidence recall, and fewer confounding tokens.
2026-09-09	What Should an Agent Forget? Separating What Is Stored from What Is Used
• Separates retained history from evidence used for each answer. • Query-conditioned views suppress obsolete facts while preserving historical access and relations. • Multi-scenario ablations favor relevance filtering and explicit control of superseded alternatives.
2026-09-09	Fortunate Recall: Ontology-Driven Memory Lifecycle Management for Persistent Coherence in LLMs
• Manages personal facts with ontology-guided memory lifecycle policies. • Typed decay, supersession, validity, and routing distinguish changing information. • Ablations attribute correctness chiefly to lifecycle metadata and calibration gains to typing.
2026-09-08	MemForest: Efficient Agent Memory Management via EventTree Partitioning and Progressive Merging
• Compresses agent memory through event-centered tree structures. • Progressive node merging and anchor-guided temporal retrieval preserve useful evidence. • Unimodal and multimodal tests retain most performance at 50% memory compression.
2026-09-08	Safe Task Planning with Long-Term Graph Memory for Embodied Agents
• Uses long-term semantic graph memory for safer embodied planning. • A risk predictor evaluates actions and triggers conservative replanning from accumulated observations. • IS-Bench and real-robot tests report higher safe success rates.
2026-09-08	CreaMem: A Scene-Aware Memory Architecture for Personalized Agents
• Partitions personalized memory by life scene. • Episodic and trait-based dual coding combines complementary perspectives during balanced retrieval. • Two memory benchmarks report QA gains, especially for multi-hop questions.
2026-09-08	MemSentry: A Framework for Detecting Persistent Memory Poisoning in Agentic AI
• Screens persistent-memory writes using configurable security-state criteria. • Trust, semantic risk, dependency reach, and access controls produce deterministic decisions. • Synthetic-scenario tests demonstrate detection capability within the modeled environment.
2026-09-08	MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents
• Suppresses query-specific memories with negative downstream utility. • Cooperative attribution detects interacting harmful evidence missed by single-removal screening. • Ten memory pools report improved recovery without permanently editing the store.
2026-09-07	Where to Look and What to Use: Retrieve-Localize-Generate for Long-Term Conversational Memory Question Answering
• Separates memory retrieval, evidence localization, and answer generation. • Multi-granularity graphs and a trained locator supply compact, position-grounded evidence. • Four-benchmark experiments report strong retrieval accuracy and response quality.
2026-09-07	MEMO: Multimodal Evidence Memory Organization for Long-Horizon LLM Agents
• Allocates retrieved evidence between textual and visual memory channels. • Trained extraction and presentation policies choose source-grounded units and layouts. • Four benchmarks report better task performance under constrained memory-token budgets.
2026-09-04	From Interaction Traces to Persistent Skills: Online Evolution for Computer-Use Agents
• Builds persistent, versioned GUI skills from interaction traces. • Frozen library snapshots and evaluator feedback support auditable online revisions. • Four OSWorld domains show conditional gains; repeated edits do not guarantee recovery.
2026-09-04	Forgetting Without Restarting: Execution-State Unlearning for Stateful LLM Agents
• Extends forgetting to derived execution state, not just stored records. • Provenance-guided checkpoint restoration and sanitized replay rebuild the affected suffix. • Behavioral audits match full-reset forgetting with fewer recomputed tokens.
2026-09-04	Compact-Memory LLM Agents via Online Max-Member Clustering and Atom-Aware Packing
• Optimizes compact agent memory under tight prompt budgets. • Max-member online clustering and atom-aware packing control merging and evidence assembly. • AMA-Bench tests reach 83% of full-context quality at 32% token cost.
2026-09-03	EdgeMem: LLM-Free Agent Memory Construction and Retrieval via Evidence-Preserving Multi-Anchor Hypergraph
• Preserves original conversation evidence in a multi-anchor hypergraph. • Local content, temporal, and episodic processing avoids generative memory construction. • LoCoMo and LongMemEval-S tests show strong retrieval without LLM management calls.
2026-09-03	RuleMem: Active Rule Memory for Long-Term Conversational Agents
• Induces reusable logical rules for conversational memory reasoning. • Natural-language Horn clauses and perplexity-based validation guide evidence retrieval. • LoCoMo and LongMemEval evaluations report strong performance against memory baselines.
2026-09-02	CHIME: Credit-Aware Hierarchical Memory Evolution for Long-Horizon Agentic Planning
• Separates planning memories from execution memories using causal attribution. • Assigns outcomes before writing to the corresponding experience bank. • Four benchmarks report strong planning gains, compact memory, and cross-backbone transfer.
2026-09-02	CAPTURE: Disentangling Preference Drift from Memory Poisoning in Personalized LLM Agents
• Distinguishes authentic preference drift from poisoning and temporary changes. • Belief tracking, multi-timescale ledgers, clarification, and counterfactual audits guide updates. • Held-out tests improve personalization, while adaptive attacks expose a robustness tradeoff.
2026-09-02	MemoryLACE: Memory Lifecycle-Aware Consolidation and Evidence Retrieval
• Models textual memory lifecycle with sparse local relations. • Merge, supersession, and contradiction links reconstruct provenance-aware evidence units. • BEAM and StructMemEval tests report improved reasoning with lower runtime.
2026-09-01	Transferable End-to-End Optimization for Indirect Long-Term Memory Poisoning in LLM Agents
• Studies indirect memory poisoning across the complete memory pipeline. • Stage feedback models interactions between writing, retrieval, and later utilization. • Cross-framework evaluations reveal transferable vulnerabilities despite several tested defenses.
2026-09-01	EM^2Mem: Event-Centric Multimodal Memory for Large Language Models
• Binds multimodal evidence to shared event anchors before retrieval. • Event memory cells align sources, time, relations, and provenance. • Three long-video QA benchmarks improve accuracy and evidence recall while reducing inference cost.
2026-09-01	Making Prospective Memory SLM-Shaped: Typed Intention Stores for Small-Model Agents
• Structures prospective memory as a typed intention store. • Code manages intention lifecycle while small models handle scoped language decisions. • PM-Bench tests substantially improve delayed-intention execution without fine-tuning.
2026-08-31	Understanding Stage-Wise Utility-Risk Trade-offs in LLM Agent Memory
• Separates utility–risk tradeoffs across memory lifecycle stages. • Matched clean and poisoned tests vary admission, management, and retrieval exposure. • Eleven-model evaluations find distinct risk patterns requiring stage-aware controls.
2026-08-31	When Errors Become Memories: Causal Pathway Tracing in Multi-Turn Memory-Augmented LLMs
• Traces how errors propagate through multi-turn memory systems. • Counterfactual interventions separate memory-update and question-feedback pathways. • Experiments find persistent latent errors and stronger repair from correcting memory.
2026-08-29	Selective Forgetting: A Graph-Based Memory Framework for Long-Term LLM Agents
• Evaluates graph memory and selective pruning against flat retrieval. • Typed conversational graphs are pruned using recency, frequency, centrality, and age. • Graph retrieval underperforms the matched baseline; pruning saves storage with limited measured loss.
2026-08-29	When to Adapt: Conditional Memory Adapters for Retention-Preserving Domain Specialization
• Uses conditional memory adapters for retention-preserving domain specialization. • Local n-gram matching and learned gates selectively inject domain knowledge. • Qwen3 experiments retain 99.4–100.1% of average out-of-domain performance.
2026-08-28	What Makes Agent Memory Useful for Reliable Unanswerable Question Handling?
• Studies memory support for recognizing unanswerable questions. • Compares four memory methods across datasets, models, and transfer settings. • Procedural guidance transfers better than larger experience stores, but gains remain dataset-sensitive.
2026-08-27	GraphMemix: Query-Aware Evidence Forests for Long-Term Multimodal Agent Memory
• Constructs query-specific evidence forests for multimodal memory. • Budgeted graph optimization balances direct support, relational verification, and activation costs. • Four benchmarks report improved accuracy–lifecycle-cost tradeoffs across foundation models.
2026-08-26	Learning What to Share and What to Personalize: Hierarchical Strategy Co-Evolution for Agent Memory
• Co-evolves shared and user-specific memory-management strategies. • Persona deltas and cross-level rule promotion adapt retention policies to individuals. • Experiments report consistent gains over memory-augmented personalization baselines.
2026-08-26	VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction
• Separates informational and emotional memory for streaming speech interaction. • Parallel memory paths support affective attribution, personas, and streaming retrieval. • Experiments report improved personalization with 134 ms retrieval latency.
2026-08-25	Memory Is Not Always Needed: Characterizing Conditional Memory in Scientific Reasoning
• Studies when conditional memory helps or harms scientific reasoning. • A knowledge-boundary router controls activation location and strength before generation. • Biology and chemistry tests favor selective routing over static memory injection.
2026-08-25	Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses
• Couples working-state tracking with recursively evolving experiential memory. • A fixed meta-agent applies localized, validation-gated skill updates from execution evidence. • Four long-horizon benchmarks report gains in most evaluated model–benchmark pairs.
2026-08-24	Dual-Grained Agent Memory and Shapley Context Attribution for Multimodal Agentic Learner
• Combines exemplar memory and abstract procedural schemas for frozen MLLMs. • Online categorization and Shapley attribution estimate rule utility for retrieval. • MathVista, MMMU, and MMMU-Pro tests report gains across four backbones.
2026-08-24	InjecMEM: Memory Injection Attack on LLM Agent Memory Systems
• Tests single-interaction memory injection without direct store access. • Evaluations examine later topic-conditioned retrieval and response steering across systems. • Results show persistence under memory drift and motivate stronger memory governance.
2026-08-23	Dual-Layer Agentic Memory with Fast Write Routing and Slow Consolidation
• Combines selective external writes with slower parameter consolidation. • A model cascade routes new information before fine-tuning high-value memories into parameters. • Authors report up to 68% redundancy reduction while retaining over 98% of baseline QA accuracy.
2026-08-23	HERO: Human-profile Enhanced Retrieval Optimization Framework for Long-term Agent Memory
• Preserves raw dialogue in a profile-guided heterogeneous memory graph. • Query anchors and user profiles steer iterative evidence traversal. • Two benchmarks report improved factual and personalized reasoning.
2026-08-23	When Not to Imitate: Boundary-Aware Skill Memory for Reliable Tool-Use LLM Agents
• Adds applicability boundaries to success-derived skill memory. • Conditions, risk cues, avoidance rules, and recovery notes guide selective skill use. • AppWorld, BFCL, and AgentDojo tests report improved reliability and efficiency.
2026-08-23	CONTRAMEM: Learning Self-Evolving Procedural Memory from Contrasting Multi-Model Trajectories
• Distills procedural memory from contrasting multi-model executions. • Localized curation maintains reusable application function cards and task skill cards. • Held-out computer-use tests report strong gains and cross-model transfer.
2026-08-22	MemGuard: Persisting Verifier Signals for LLM-Agent Memory Governance
• Persists verifier signals throughout the agent-memory lifecycle. • Reward, confidence, labels, and uncertainty guide admission, retrieval, conflict handling, and archival. • Four benchmarks report improved success and fewer steps under matched runtime budgets.
2026-08-22	MEMORY Wins All: Indirect Bias Injection Attacks via Social Media Feeds
• Studies indirect stance manipulation through remembered external content. • Social-feed and email scenarios trace ingestion, curation, and downstream influence. • BiasBench reports persistent effects and only partial mitigation from a boundary defense.
2026-08-21	Weighted Memory Tree: Remembering What Matters for Long-Horizon LLM Agents
• Weights hierarchical task, subtask, and action memories dynamically. • Event updates and decay fold completed work while retaining recoverable evidence. • GAIA-Text tests report higher accuracy and lower prompt-token usage.
2026-08-21	Utility Under Attack: Agent Memory Poisoning and the Limits of Content Screening and Provenance Ranking
• Tests false factual memories beyond explicit prompt-injection content. • Controlled poisoning compares write-time screening and provenance-weighted retrieval. • Results expose utility losses and limits of content-only filtering and additive provenance penalties.
2026-08-21	ForeDreamer: A Self-Evolving Dual-Agent Memory Architecture for Future Event Prediction
• Introduces ForeDreamer, a self-evolving dual-agent memory framework for open-web future event prediction that separates question-specific factual memory from persistent experiential memory. • A main agent searches and predicts, while a memory-processing subagent uses MemGuides and executable MemTools to transform noisy, cutoff-aware web evidence into structured factual memory; textual and procedural memories evolve through validation-gated updates. • With Qwen3.5-Flash and GPT-5.4-Nano, ForeDreamer achieves Brier scores of 0.1471 and 0.1839 on Prophet Arena and accuracies of 0.4108 and 0.3883 on FutureX, outperforming the strongest compared baseline for each benchmark and backbone.
2026-08-20	Beyond Memory Majority: Latent-Source Reasoning for Multi-Agent Memory Arbitration
• Prevents correlated multi-agent memories from forming false majorities. • Dependency inference and provenance estimate independent sources and guide evidence recovery. • Benchmark experiments report more reliable arbitration than compared baselines.
2026-08-19	MemFuse: Multi-Source Memory Fusion from Fragmented Observations
• Fuses fragmented observations into source-traceable episodic memory. • Atomic events and fused clusters organize evidence in a causal graph. • MemFuseBench tests report stronger cross-source reasoning across three LLM settings.
2026-08-18	ArborMem: Navigating Interaction States with Memory Forests
• Represents interrupted and resumed conversations as forests of interaction states. • Localizes the active branch before restoring context and retrieving cross-branch evidence. • Established benchmarks and BranchMemEval report stronger continuity under bounded read budgets.
2026-08-18	CABLE: Extending the Reach of Memory Retrieval via Complementary Antecedent-Based Linking and Expansion
• Adds sparse associations beyond a retriever's semantic neighborhood. • Verified antecedent links expand retrieved seeds toward otherwise missed supporting evidence. • Multiple memory systems report gains on cross-session and preference questions.
2026-08-18	On the Fragility of Self-Improving Agents: Variance, Task Order, and Underspecification
• Re-evaluates reliability of memory-based self-improving agents. • Repeated runs and shuffled task streams expose variance and ordering effects. • Better task specification helps partially, leaving substantial unexplained fragility.
2026-08-17	HyperSkill: Self-Evolving LLM Agents via Hypergraph-Structured Skill Memory
• Organizes procedural memory as a hypergraph of subtasks and skills. • Dual-path retrieval and structure-aware maintenance reuse trajectory-level relationships. • Three agent benchmarks report gains over ten memory baselines.
2026-08-17	QUMem: Personalized Memory for Query-Conditioned User-State Inference in LLM Agents
• Infers query-specific user states from evolving personal memory. • Typed episodic records and sequential retrieval agents resolve temporal and contextual relevance. • PersonaMem and KnowU-Bench tests report strong personalization performance.
2026-08-17	FTA-Mem: Fact-Time-Affect Anchored Memory for Low-Density Long-Term Dialogue
• Structures low-density dialogue memory around facts, time, and affect. • Boundary-preserving segmentation creates coherent situation-level memory units. • ES-MemEval and LoCoMo tests improve memory QA with balanced construction granularity.
2026-08-17	MELD: A Protocol for Merging Knowledge Across Distributed Agentic Memories
• Reconciles knowledge across distributed agent memories without silently resolving contradictions. • Auditable patches and replicated claim states support merging and partition recovery. • HotpotQA and deployment tests report useful recall, storage, and convergence tradeoffs.
2026-08-17	What to Remember, What to Reveal: Privacy-Aware Memory for Conversational Agents
• Separates sanitized conversational memory from exact private values. • Isolated stores and consent-aware retrieval govern sensitive information across its lifecycle. • A privacy-aware benchmark reports personalization gains with less unnecessary exposure.
2026-08-17	Cross-Model Memory Transfer via Target-Side Reader Adaptation
• Transfers frozen addressable memory between model backbones. • Adapts lightweight target-side readers while preserving learned memory tables. • Controlled QA tests show reader alignment is crucial for effective reuse.
2026-08-13	ERSkill: Evolving for Skill-Guided Adaptive Memory Retrieval
• Evolves executable retrieval skills and their query router jointly. • An experience trie and separate exploration and deployment frontiers support stable updates. • Multiple memory benchmarks report gains over static and self-evolving baselines.
2026-08-13	Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence
• Converts verified spatial experience into transferable procedural memory. • Reflection and outcome-calibrated reliability scores guide frozen VLM inference. • Five spatial benchmarks report improvements across four base models.
2026-08-13	When Your Agent Opens the Chat App: Agent-Controlled Search over Raw Chat Logs Rivals Structured Memory
• Tests agent-controlled search over unmodified conversational archives. • Lexical indexing, session controls, temporal filters, and iterative search gather evidence. • MemoryAgentBench and LongMemEval results challenge the need for elaborate semantic indexes.
2026-08-13	LycheeMemory V2: Efficient Long-Term Memory for LLM Agents via Semantic Segment-Level Consolidation
• Consolidates dialogue at semantic segment boundaries rather than every turn. • Typed records and lightweight indexes support planned evidence retrieval. • LoCoMo and LongMemEval-S tests reduce construction tokens without increasing query cost.
2026-08-13	RippleMem: From Isolated Retrieval to Associative Recollection for Long-Term Agent Memory
• Replaces isolated retrieval with adaptive associative recollection. • Cue-rich episodic anchors expand along event-graph associations to complete evidence. • LoCoMo and LongMemEval-S tests improve accuracy with lower graph-construction cost.
2026-08-12	Towards a Formal Definition of Agent Memory: Basis, Span, Optimality, and the Sequential Memory Problem
• Formalizes agent memory through knowledge coverage under capacity constraints. • A sequential decision model treats writes as actions with delayed query utility. • An Odyssey case study illustrates utility–capacity frontiers and coverage–precision differences.
2026-08-12	ε-MemEvo: Adaptive Cross-Task Memory Transfer for LLM Program Evolution
• Transfers task-agnostic tactic memories between program-evolution tasks. • An adaptive gate controls whether and how strongly retrieved strategies are injected. • Eight held-out-task tests report faster improvement and less harmful transfer.
2026-08-11	From Faulty Memories to Corrected Actions: Dependency-Guided Rollback Repair for Memory-Augmented Agents
• Repairs answers and persistent state after diagnosed memory failures. • Dependency graphs invalidate unsupported descendants and selectively replay affected computation. • Controlled and trajectory-derived tests improve recovery while preserving benign memory.
2026-08-11	Self-Correcting Long-Horizon Search Agents via Tree-Structured Memory
• Repairs downstream reasoning when stored search evidence changes. • A provenance-linked evidence tree supports localized revision and affected-branch pruning. • Four QA and search benchmarks report higher accuracy with bounded per-step context.
2026-08-10	Muscle Memory for Agents: Compile not Merely Retrieve
• Compiles recurring user intent into specialized executable agents. • A quality-gated pipeline mines history and matches specialists through staged triggers. • A 90-scenario study reports personalization gains with a small accuracy tradeoff.
2026-08-10	Skills in Weights, Memory in Code: Hybrid Learning for Memory-Dependent Robot Manipulation
• Separates learned motor skills from executable memory-management code. • Coding-agent heuristics and multimodal completion checks steer a Markovian VLA. • RoboMemArena reports higher cumulative and task success than compared policies.
2026-08-10	MESA:Task-Adaptive Multi-Structure Evidence Selection for Long-Horizon Agent Memory
• Selects task-dependent combinations of complementary memory structures. • Prior-guided search learns which of five evidence views a frozen reader should use. • AMA-Bench reports improved answers with fewer tokens than using every structure.
2026-08-08	SodaMem: Evidence-Grounded Temporal Graph Memory for LLM Agents
• Tracks evidence provenance and temporal validity in graph memory. • Typed fact events and planner–reader retrieval distinguish updates and contradictions. • LongMemEval-S reports strong accuracy, with self-judging and estimated cross-system costs as limitations.
2026-08-08	LatticeMind: A Conflict-Aware Memory Primitive for Multi-Agent Systems
• Handles contradictory multi-agent claims at memory write time. • Explicit statuses and symbolic checks invoke model reconciliation only when needed. • ConflictBank improves strongly; planning results remain mixed against deliberation methods.
2026-08-08	Mitigating Over-Personalization in LLMs via Structured Memory
• Studies over-personalization caused by persistent user memory. • Partitions memory by domain at inference time without changing stored content. • PersistBench tests across seven models reduce cross-domain leakage while preserving utility.
2026-08-07	MemPrism: Task-Conditioned Relational Memory Views for Long-Horizon Agents
• Constructs task-conditioned working views over persistent experience. • A lightweight policy selects relations, evidence scope, outcomes, and granularity. • Embodied and web-agent tests improve performance while reducing memory tokens.
2026-08-07	Coupling Planning with Episodic Memory in LLM Agents for Software Issue Resolution
• Couples hierarchical planning with episodic memory for software repair. • Plan phases guide retrieval; memory statistics trigger replanning and execution-grounded verification. • SWE-bench Verified tests report more resolved issues and fewer repeated failures.
2026-08-07	Controlled Memory Interference in Continual LLM Agents
• Diagnoses interference between accumulated agent memories. • Controlled relationships and update-authority cues generate targeted learning examples. • Experiments show relationship-specific interference impairs updates beyond simple memory growth.
2026-08-07	MemOPD: On-Policy Distillation through Memory State Alignment for Long-Horizon Agents
• Aligns teacher supervision with the actual memory state of student rollouts. • Reconstructs each invocation's positions and visibility before packed on-policy distillation. • Matched controls report a 7.0% F1 improvement over persistent-history teacher scoring.
2026-08-07	MemWM: Memory-Augmented Text-Based World Model
• Augments text world models with explicit transition and fact memory. • World memory conditions predictions while retrieved skills guide a frozen planning policy. • Three environments report better state fidelity and downstream task success.
2026-08-07	Agent Memory Distillation: Empowering Small LLM Agents with Hierarchical Teacher Memory
• Transfers teacher experience into hierarchical memory for small agents. • Workflow, subtask, and function memories provide proactive and error-triggered guidance. • Three tool-use benchmarks report gains for 4B–8B students without parameter training.
2026-08-07	TEPA: Revoking Stale Memories for Conflict-Robust Language Agents
• Makes evolving memory validity explicit through revocable evidence records. • Keyed contradictions deactivate stale precedents while preserving audit history. • Drift tests improve current evidence use; multi-hop settings expose additional retrieval bottlenecks.
2026-08-06	Causal Episodic Memory for Feedback-Driven Agent Repair
• Reuses verified corrections and failed directions across SQL-repair episodes. • Failure-typed retrieval selects only earlier finalized experiences under oracle feedback. • Spider improves clearly; BIRD evidence is weaker and some retrieval controls remain comparable.
2026-08-05	FocusMem: Factorizing Content, Readout, and Trust in Latent GUI Memory
• Separates content retention, readout, and trust in latent GUI memory. • Role-aware representations and state-conditioned gates support a frozen action policy. • Five GUI benchmarks improve performance and reduce harm from irrelevant memories.
2026-08-05	MemoryCPT: An End-to-End Agent Memory Framework for Cost-Performance Trade-off
• Trains memory construction and query-time summarization for cost efficiency. • Distillation and cost-aware reinforcement learning produce compact retrieved contexts. • LoCoMo and LongMemEval tests improve measured quality per inference cost.
2026-08-05	Mimir: A Neuro-Symbolic Memory System with Dynamic Grounding for Embodied Agents in Interactive Environments
• Separates embodied world memory from task-progress memory. • Dynamic grounding binds active goals to recalled objects, locations, and evidence. • EB-ALFRED and EB-Habitat tests improve success across evaluated backbones.
2026-08-05	Hierarchical Graph Memory for LLM Agents with Path-level Localization and Rewrite
• Updates hierarchical graph memory through localized evidence paths. • Query- and update-conditioned subgraphs coordinate memory-unit and dependency rewrites. • Conversational and conflict-aware tests report better answers and token efficiency.
2026-08-04	DP-MemView: A Memory Interface for Attribute-Level Transcript Privacy in Long-Term LLM Agents
• Protects attributes across repeated memory-conditioned responses using differential privacy. • Private view selection and per-attribute budgets restrict what the response model receives. • Controlled and transfer tests preserve useful personalization under the stated interface contract.
2026-08-04	Verifiable Memory: Learning Unified Memory Management with Local and Global Verifiers for Large Language Model Agents
• Unifies long-term memory, active context, and episodic history management. • Trains seven memory operations using local and global verifiers with hierarchical rewards. • Five-benchmark experiments report strong accuracy and token-budget efficiency.
2026-08-04	Towards Improving Sequential Decision-Making in LLM Agents via Experience Memory
• Studies experience memory for sequential game decisions. • Post-game reflection extracts reusable rules without modifying model parameters. • Tic-tac-toe tests show measurable gains despite broader decision-making weaknesses.
2026-08-04	LeanMem: Simple and Efficient Long-Term Memory for LLM Agents
• Separates profile, event, and source-record memories by information properties. • Selective event updates and query-dependent budgets avoid uniform processing overhead. • LoCoMo and LongMemEval-S tests report accuracy gains with low construction costs.
2026-08-04	TARL: Transaction-Aware Reliable Ledgers for Executable Memory Management in Long-Term Agents
• Distinguishes five executable memory-update actions beyond write-or-hold. • Temporal scope and source reliability route claims into accepted, pending, or rejected ledgers. • TARL-Mem evaluations report better state recovery and reduced cumulative corruption.
2026-08-03	When Memory Updates but Behavior Does Not: Repairing Implicit Stale Dependencies in Personalized Agent Responses
• Audits stale behavioral dependencies from stored state toward response drafts. • Verified quotations and chronology gate repairs for candidate state transitions. • STALE improves, while external tests bound claims of general applicability.
2026-08-03	Salami Attack: Stealthy Collusive Memory Poisoning against OpenClaw
• Studies collusive poisoning from individually benign-looking memory fragments. • A controlled social-platform setting tracks storage and separate-session behavioral effects. • OpenClaw evaluations expose compositional risks that persist under dilution and memory defenses.
2026-08-03	PGMem: Tightly Coupled Persona-Memory Graph for Lifelong Personalized Agents
• Links evolving personas directly to supporting event memories. • Typed evidence edges and validity-aware ranking guide personalized retrieval. • Three benchmarks with small models outperform several memory representation baselines.
2026-08-03	CoEvo-Mem: Co-Evolving Retrieval Policy and Memory Bank for LLM Agents
• Co-evolves retrieval policy and the memory bank through task feedback. • Alternating router and memory updates control nonstationarity in their feedback loop. • Seven benchmarks report benefits from joint retrieval–memory adaptation.
2026-08-03	MemSIF: From Structured Interactions to Dual-Track Fact Memory for LLM Agents
• Combines structured interactions with stable and query-emergent fact memories. • Topical segments and event trajectories preserve continuity across time. • LoCoMo and LongMemEval-S tests report gains across five backbones.
2026-08-03	Benign Alone, Harmful Together: Exploiting Experience Composition in Self-Evolving LLM Agents
• Studies safety risks from composing individually benign persistent experiences. • Sequential experience acquisition tests how complementary memories jointly alter later behavior. • Cross-framework evaluations expose composition as a distinct persistent attack surface.
2026-08-03	MemArbiter: Decision-Time Memory Arbitration for Long-Horizon LLM Agents
• Arbitrates memory salience at decision time for long-horizon agents. • Functional banks and temporal presentation gates turn stored items into actionable context. • Budget-matched ALFWorld tests improve success and post-failure recovery.
2026-08-03	EvoGraph-Mem: Failure-Aware Editable Graph Memory for Long-Term Language Agents
• Maintains reusable insights through failure-aware graph editing. • Positive and negative evidence guide retrieval, revision, archival, and new writes. • Multi-backbone experiments favor editable maintenance over append-only memory.
2026-08-03	RoMeRL: Balancing Feedback Coverage and the Memory-Reward Trap in Self-Evolving Agent Memory via Reduced-Order Utility States
• Bounds memory-utility state to concentrate sparse learning feedback. • Outcome and dynamics factors replace expanding trajectory-level reward coordinates. • ALFWorld and LifelongAgentBench report better performance with smaller memory and fewer calls.
2026-08-02	PMMC: Prospective Multimodal Memory Compilation for Long-Term LVLM Agents
• Compiles prospective multimodal memory programs during consolidation. • Question prediction, planning, and verification build a reusable evidence-routing bank. • Multimodal benchmarks improve answer quality and visual recall with lower query cost.
2026-08-02	TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents
• Stores memory as source-grounded evolution trajectories. • Immutable snapshots, claim edits, and interlinked wiki pages support hierarchical retrieval. • LoCoMo and MedMT tests report improved dialogue performance and diagnostic traceability.
2026-08-02	Stop When Memory Suffices: Evidence-Conditioned Progressive Execution for LLM Agents
• Stops memory execution early when retrieved evidence is sufficient. • A trained router escalates only unresolved queries to deeper analysis. • AMA-Bench and BEAM tests reduce average inference time versus full execution.
2026-08-02	V-Mem: Modality-Routed Retrieval for Long-Term Multimodal Agentic Memory
• Routes multimodal memory retrieval by query and evidence modality. • Round-level alignment and generated search anchors address modality and relevance gaps. • Mem-Gallery and LoCoMo report gains, especially on image-containing queries.
2026-07-31	Tokenizer-Agnostic Engram Module
• Makes Engram lookup independent of tokenizer-specific token boundaries. • Polynomial byte-sequence hashing shares one embedding space across n-gram lengths. • Tests retain comparable performance and equal hashes for byte-equivalent sequences.
2026-07-31	Reproducing LightMem: Naive RAG Is Just as Good for Memory Management
• Reproduces LightMem and compares constructed memory with raw-turn retrieval. • Controlled retriever and budget changes isolate sources of performance variation. • Raw retrieval often matches or wins; constructed memory helps mainly under tight budgets.
2026-07-31	Zero-Mem: Zero-Token Memory Operations for LLM Agents
• Avoids generative model calls during memory operations. • Entity–context graphs and temporal hierarchies organize raw traces for query-specific retrieval. • Matched-reader tests retain competitive QA and reduce memory-operation time.
2026-07-31	Know It, Act on It: Investigating Memory Utilization in LLM Personalization
• Separates knowing user preferences from acting on them. • Paired recall and behavior tests cover sixteen systems and five memory architectures. • Results reveal substantial utilization gaps even when relevant preferences are recalled.
2026-07-31	Beyond Retrieval: Analytic Memory for Multimodal Agents
• Adds analytical operations over recurring multimodal observations. • Provenance-linked attributes support filtering, aggregation, ranking, and temporal comparison. • MemEye and MemGallery tests report gains over retrieval-focused memory systems.
2026-07-31	CrystalMem: Elastic Memory for Self-Evolving LLM Agents via Knowledge Crystallization
• Addresses capability loss when memory budgets shrink and later recover. • Uses four fidelity levels and verified reconstruction under explicit resource caps. • Seven-environment experiments report stronger capability recovery than compared baselines.
2026-07-30	ChronoMem: Version Control and Semantic Rollback for Large Language Model Agent Memory
• Adds semantic version control and rollback to agent memory. • Whole-memory snapshots and hybrid retrieval map natural-language undo requests to versions. • Post-exposure tests improve rollback-consistent QA and history summaries.
2026-07-30	MIND: Lightweight and Effective Memory Injection Defense for LLM Agents via Intent-Aware Information Bottleneck
• Detects memory injection through compact intent–behavior representations. • An information bottleneck filters repeated context before lightweight threat classification. • Experiments reduce attack success while retaining task accuracy and inference efficiency.
2026-07-30	RRM: Experience-Driven Reflective Retrieval Memory for Long-Horizon Multimodal Reasoning
• Stores reusable retrieval strategies separately from current-video facts. • Reflection over earlier trajectories guides queries while lifecycle policies control experience noise. • Three long-video benchmarks report improved multimodal reasoning.
2026-07-30	MemHarness: Memory Is Reconstructed, Not Replayed
• Reconstructs retrieved experience for the current agent state. • A unified policy critiques memory and learns context-grounded guidance through GRPO. • ALFWorld and WebShop tests improve performance and out-of-distribution robustness.
2026-07-29	MemRetriever: Learning to Search, Reflect, and Retrieve from Long-Term Memory
• Learns multi-step search and reflection over long-term memory. • Supervised warm starts and GRPO reward coverage, denoising, sufficiency, and stopping. • Five QA benchmarks report gains over static and supervised-only retrieval.
2026-07-29	Filesystem-Based Memory for LLM Agents: Organization, Evolution, and Sustainability
• Systematically studies filesystem-based declarative and procedural memory. • Varying organization, tools, and agent strengths tests quality, cost, and store health. • Organization saves search cost, but measured agents do not reliably turn it into better answers.
2026-07-29	Bridging Inference-Time Scaling and Episodic Memory with Action-Centric Graphs
• Connects episodic experience to inference-time search through action graphs. • Dual temporal-difference values encode useful actions and risky alternatives. • Multiple benchmarks report improved success and progress over vanilla agents.
2026-07-28	MemLens: A Value-Aware Memory Management System with Interactive Analytics for LLM-based Agents
• Exposes memory value through an interactive management dashboard. • Shapley-style evaluation guides storage and visualizes hierarchical memory records. • A study-copilot application supports comparisons of quality, latency, and token usage.
2026-07-27	MemChain: Learning Interpretable Memory Traces for Memory-Augmented LLM Agents
• Learns interpretable post-retrieval memory traces for answer generation. • Evidence plans and explicit memory actions are optimized with grounded trace rewards. • LoCoMo and LongMemEval-S tests improve accuracy while shrinking answer-facing context.
2026-07-26	Isolated but Exposed: Persistence-Based Memory Extraction Attack on LLM Agents
• Examines memory extraction through malicious tool interfaces. • Cross-session evaluations test exposure despite per-user memory isolation. • Results show isolation alone does not protect data passed into tool calls.
2026-07-25	ConsistencyGate: Preventing Memory Contamination in LLM Agents via Self-Consistency Admission Control
• Screens candidate memory writes through self-consistency support scores. • Repeated judgments or a log-probability variant gate unsupported facts without fine-tuning. • Four-backbone tests reduce contamination, with costs for implicitly stated facts.
2026-07-23	AttriMem: Attribution-Guided Process Feedback for Agent Memory Construction
• Assigns fine-grained process rewards to memory construction. • Token-level answer attribution supplements global task rewards during reinforcement learning. • Long-dialogue QA tests report stronger generalization and more stable optimization.
2026-07-22	PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning
• Uses programmatic access to complete structured interaction logs. • Coding agents search retained histories instead of relying on lossy fixed summaries. • ARC-AGI-3 tests report higher success and fewer tokens than specialized harnesses.
2026-07-20	Retain or Consolidate? Budget-Dependent Operator Selection for Language Agent Memory
• Chooses retention or consolidation according to memory budget pressure. • A calibrated utility model selects merging, abstraction, rewriting, or raw retention. • LongMemEval and LoCoMo show consolidation helps tight budgets while retention favors loose ones.
2026-07-20	Mechanistic Attention Guidance for Agent Memory Refinement
• Uses internal attention signals to refine agent memory. • Segment utilization patterns guide edits that are verified through re-execution. • Interactive benchmarks improve task outcomes and memory efficiency over text-only refinement.
2026-07-20	Exploratory and Assimilating Reflection: Reflective Recall Cycle for Long-term Memory
• Combines exploratory retrieval with experience-replayed reranker refinement. • Iterative search gathers feedback that updates a shared global retrieval policy. • Two dialogue benchmarks report improved recall and robustness to noisy feedback.
2026-07-17	Do Agents Dream of False Memories? Black-box Visual Attacks on Long-term Memory in Multimodal AI Agents
• Studies image-bounded attacks on persistent multimodal memory. • Black-box tests separate corrupted historical recall from unsupported visual injection. • Five memory architectures exhibit vulnerabilities without attacker access to text or model internals.
2026-07-17	LazyMem: Retrieve Broadly, Construct Selectively for Efficient Long-Term Agent Memory
• Defers memory construction until the query is known. • A trained small model selectively compresses broadly retrieved evidence in parallel windows. • LongMemEval reports strong accuracy with compact context and transfer to LoCoMo.
2026-07-16	MemPoison: Uncovering Persistent Memory Threats and Structural Blind Spots in LLM Agents
• Benchmarks direct, compositional, and dormant memory corruption. • Hand-validated cases span injection channels, memory substrates, and model families. • Write-time checks suppress direct attacks more reliably than composed or triggered ones.
2026-07-15	Memory as a Controlled Process: Learned Adaptive Memory Management for LLM Agents
• Learns adaptive control over retrieval, plan reuse, consolidation, and forgetting. • A lightweight online contextual bandit wraps existing memory backends. • Six benchmarks report higher success with reduced token consumption.
2026-07-15	CMI-Mem: Toward Generalizable Long-Term Memory Management via CMI-Augmented Reinforcement Learning
• Trains a lightweight memory manager with complementary intrinsic and task rewards. • Conditional mutual information rewards nonredundant updates independently of sampled queries. • Experiments report stronger transfer and more efficient training and inference.
2026-07-15	Experience Memory Graph: One-Shot Error Correction for Agents
• Represents agent failure recovery as experience-graph matching. • Successful subgraphs and corrective edit paths provide reusable action guidance. • ALFWorld and ScienceWorld tests improve recovery without test-time trial-and-error loops.
2026-07-15	MEMORA: Embodied Action Memory from Egocentric Videos for Reasoning and Planning
• Forms embodied action memory from accumulated egocentric video. • Multiple stores and online/offline consolidation derive reusable routines and preferences. • A 45-hour benchmark and robot demonstration support memory-grounded planning beyond observed episodes.
2026-07-14	Track, Rank, Crack: Epistemic Working Memory Scales Multi-Hop Reasoning in Language Agents
• Makes investigative working state explicit for multi-hop reasoning. • Source-grounded facts, ranked hypotheses, and open questions guide search and commitment. • Five benchmarks show larger gains as reasoning chains become harder.
2026-07-14	AutoMem: A Text-Gradient Recursive Self-Improvement Framework for Automated Memory Architectures Search
• Searches task-adaptive agent memory architectures rather than fixing one design. • Experience-guided search and failure diagnosis optimize encoders, stores, retrievers, and managers. • Authors report a 2.8-point average accuracy gain across six settings.
2026-07-13	ToolAtlas: Learning Once, Reusing Everywhere with Tool-Side Memory
• Places reusable tool knowledge in provider-side graph memory. • Execution-verified probes capture capabilities, failure boundaries, and tool compositions. • MCP benchmarks report gains and transfer across agents and environment instances.
2026-07-13	LightMem-Ego: Your AI Memory for Everyday Life
• Maintains streaming multimodal memory for everyday-life assistants. • Aligned egocentric audio and video feed current, short-term, and long-term stores. • A phone-and-glasses demonstration supports recall, object finding, and personalized assistance.
2026-07-12	The Compliance Trap: Diagnosing How AI Agents Consume Conflicting Memory
• Diagnoses how conflicting memory changes agent action trajectories. • Entry, propagation, and recovery analysis isolates stages of memory-induced failure. • WebArena and controlled tests reveal early compliance and weak recovery after divergence.
2026-07-09	What to Keep, What to Forget: A Rate--Distortion View of Memory Compaction in LLMs and Agents
• Unifies memory compaction through a rate–distortion perspective. • A shared taxonomy connects KV caches, prompts, recurrent states, and agent memory. • Identifies irreversible early deletion and missing repeated-compaction evaluation as open problems.
2026-07-09	Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
• Uses a separate memory agent for selective proactive reminders. • Structured state updates decide when intervention helps an unmodified action agent. • Terminal-Bench and tau-squared-bench tests outperform passive or always-on memory exposure.
2026-07-08	MILES: Modular Instruction Memory with Learnable Selection for Self-Improving LLM Reasoning
• Builds modular step-level instruction memory with learned selection. • Coarse-to-fine retrieval expands memory and trains lightweight correctness-oriented selection heads. • Experiments report favorable reasoning accuracy, efficiency, and transfer.
2026-07-08	Parametric Multimodal User Memory: Storing What Captions Cannot Carry
• Preserves perceptual user identity beyond textual captions. • Context grounding and dedicated identity encoders combine with a separate factual text store. • PerceptMem experiments show complementary roles for perceptual and factual memory.
2026-07-07	From Passive Retrieval to Active Memory Navigation: Learning to Use Memory as a Structured Action Space
• Trains active navigation over a multi-granularity user-memory pyramid. • Memory tools expose linked conversations, records, topics, and profiles for reinforcement learning. • Three memory benchmarks report competitive performance while largely retaining general abilities.
2026-07-07	MemDefrag: Latent Memory Defragmentation for Large Language Models
• Defragments latent memory without additional training. • Middle-layer attention guides fragment reordering, filtering, and information-aware forgetting. • Authors report substantially better retention after repeated memory updates.
2026-07-06	Your Agent's Memories Are Not Its Own: Forged Reasoning Attacks on LLM Agent Memory and Defenses
• Studies forged reasoning histories as persistent memory attacks. • A layered structural detector screens candidate reasoning entries for manipulation. • Small-scale evaluations report reduced attacks, with no observed false positives on tested benign traces.
2026-07-06	When Claws Remember but Do Not Tell: Stealthy Memory Injection in Persistent Personal Agents
• Tests stealthy memory injection through ordinary external email processing. • WhisperBench tracks silent adoption and later effects across persistent-agent systems. • Held-out evaluations expose cross-system persistence despite several defense layers.
2026-07-06	When Agents Remember Too Much: Memory Poisoning Attacks on Large Language Model Agents
• Studies poisoning of tool-using personal-agent memory. • Two-stage evaluations trace initial storage and later activation of poisoned records. • A memory-saving policy and retrieval screen reduce attack success while preserving utility.
2026-07-02	DRIFTLENS: Measuring Memory-Induced Reasoning Drift in Personalized Language Models
• DRIFTLENS is a ground-truth-free framework that maps each expressed reasoning step to a value-ontology symbol and measures divergence between a question's no-memory trajectory and its trajectory under injected user-attribute memory, revealing that personalization memory silently reshapes how a model reasons ("symbolic drift"), not just its answer. • Measures per-instance reasoning stability of personalized LLMs under memory perturbations using a value ontology and two drift metrics (DTW and SRI) on a benchmark of unverifiable, persona-indifferent questions, and evaluates GRPO- and DPO-based post-training as mitigation. • Across four LLMs and 10 user-attribute categories, irrelevant persona memory induces medium-to-large reasoning drift (Cohen's d ≈ 0.35–0.98); GRPO and DPO both reduce drift but neither dominates (e.g., GRPO lowers DTW to 0.186 vs. 0.309 on Gemma2-2B; DPO reaches 0.204 on Qwen3-4B).
2026-07-02	InduceKV: Fixed-Footprint Continual Adaptation of Multimodal LLMs via Inducing KV Memories
• Reframes continual multimodal-LLM adaptation as budgeted online inducing-set selection: task increments are stored as attention-compatible external KV memories (a frozen retrieval key plus compact layerwise KV payloads) injected into self-attention, keeping the backbone frozen under a strict fixed memory budget. • Fixed-footprint continual adaptation of MLLMs (task-incremental tuning, continual VQA, domain-incremental, lifelong tuning); extracts attention-ready memory entries and builds a compact inducing set via bilevel optimization (inner retrieval calibration; outer weight selection). • Consistently beats PEFT, MoE, replay, and prompt-retrieval baselines under matched budgets; improves over HiDe-LLaVA by 0.88 Avg/1.12 Last on UCIT and 1.35/1.43 on COIN, and raises continual-VQA AP 51.34→52.64 over CL-MoE while outperforming QUAD on VQACL.
2026-07-02	A-TMA: Decoupling State-Aware Memory Failures in Long-Term Agent Memory
• Identifies "ghost memory" — a state-coordination failure where old, current, and transition facts coexist and mislead answers — and proposes A-TMA (Adaptive Truth Maintenance Auditing), a state-aware overlay that decouples memory into three diagnosable levels (bank maintenance, retrieval, answer-time resolution). • Long-term agent memory under changing user facts; A-TMA keeps superseded/transition records with typed links (a lightweight Sentry gate plus a Qwen2.5-3B Judge), builds state-aligned evidence packets, and conditions QA on explicit labels, alongside a new conflict-heavy benchmark LTP (LoCoMo Temporal Plus). • On LTP, Graphiti/Zep +A-TMA improves conflict accuracy by 0.240 absolute (0.480→0.720) and InsideOut+A-TMA lifts Acc from 0.117 to 0.662; on LoCoMo, Graphiti/Zep +A-TMA raises temporal F1 from 0.0295 to 0.1705.
2026-07-02	Learning User-Aware Recall: Personalized Retrieval in Long-Term Conversational Memory
• Profile-guided Personalized Retrieval Optimization (PPRO) makes long-term conversational memory retrieval both user-aware and optimizable by injecting a derived user-profile embedding as an explicit personalized prior into the retrieval ranking score. • Personalized long-term conversational QA; PPRO builds episodic and semantic memory banks plus a user profile offline, performs profile-guided dual-path retrieval, and trains a query rewriter with GRPO using evidence-retrieval and answer quality as rewards while keeping memory banks and answer model frozen. • On LoCoMo, PPRO gives the best overall F1 across three backbones, beating prior-best SimpleMem by 7–19 points (e.g., GPT-4o overall F1 48.16 vs. 40.87); on LongMemEval-S it reaches 81.5 overall accuracy vs. 75.9 for the best baseline.
2026-07-02	ISM: Self-Improving Strategy Memory for Continual Mathematical Reasoning
• Intelligent Schema Memory (ISM) is a self-evolving external memory that lets a frozen LLM improve at math reasoning under hard episodic resets by maintaining a compact, bounded bank of strategy schemas with dual representation (stable content + online-adapting feature hook), where every update is gated by symbolic verification. • Continual mathematical reasoning under a streaming episode protocol with frozen parameters; ISM uses two-stage retrieval (operator filter + soft scoring), symmetric success/failure learning, and seven self-improvement mechanisms plus conditional schema synthesis. • 80.67% on MATH-Hard and 61.67% on OlympiadBench over a 300-episode stream, beating the strongest baseline by +2.00 points on each while storing 64% and 86% fewer schemas (up to 23× fewer entries), with positive backward transfer (+0.03) on OlympiadBench.
2026-07-01	Multi-Head Recurrent Memory Agents
• Decomposes recurrent-memory performance into capture vs. retention, diagnoses retention as the dominant bottleneck (caused by monolithic memory blocks), and proposes Multi-Head Recurrent Memory (MHM) — a training-free framework partitioning memory into independent heads with a stage-wise select-then-update strategy that structurally shields unselected heads from overwriting. • Reliable long-context reasoning over 100K–1M tokens; MHM-LRU is a lightweight instantiation that selects the least-recently-updated head each step, guaranteeing uniform head utilization with zero extra token overhead and no retraining. • On RULER-HQA at 896K tokens, MHM-LRU lifts retention from less than 30% to 73.96% and accuracy to 49.74% (vs. 21.62% MemAgent, 0.00% native LLM); on BABILong at 1M tokens it reaches 41.41% vs. 25.26% for MemAgent, staying stable where baselines collapse.
2026-07-01	AUTOMEM: Automated Learning of Memory as a Cognitive Skill
• Reframes memory management as an independently trainable "metamemory" skill by promoting file-system operations (read/write/search/append/create) to first-class memory actions alongside task actions, then automates its improvement along scaffold structure and model proficiency via meta-LLM-driven outer loops. • Long-horizon procedurally generated games (Crafter, MiniHack, NetHack); Loop 1 (a meta-LLM revises the agent scaffold/file schema) and Loop 2 (a meta-LLM curates good memory decisions to LoRA-finetune a dedicated "memory specialist" while the gameplay model stays frozen). • Optimizing memory alone yields ~2×–4× gains on a Qwen2.5-32B base — Crafter 25.0→51.36%, MiniHack 7.5→30.0%, NetHack 0.42→1.85% — bringing the 32B model to the level of frontier systems like Claude Opus 4.5 and Gemini 3.1 Pro Thinking.
2026-07-01	Imprint: Online Memory Compression for Long-Horizon Egocentric QA
• Imprint reframes long-horizon egocentric memory as an online memory compression problem (rather than hierarchical text summarization), representing observations as structured Interaction Records and consolidating them using cognitively-inspired signals of recurrence, recency, and distinctiveness. • Long-horizon egocentric QA; parses captions into (person, action, object, timestamps) records via Qwen2.5-7B-Instruct, groups them into event prototypes, scores importance, and consolidates online into a compact retrieval-oriented memory. • On the EgoLifeQA 7-day benchmark, improves QA accuracy 31.0%→35.8% and grounded accuracy 10.8%→64.8% (6× more evidence-grounded answers than EgoRAG), while reducing memory footprint 2.3× (109 MB vs. 254 MB) and retrieval latency 11.8× (1.7s vs. 20.1s/query).
2026-06-30	From Signals to Structure: How Memory Architecture Drives Language Emergence in LLM Agents
• Demonstrates that in Lewis signaling games with frozen LLM agents, memory architecture matters more than channel capacity for language emergence — a persistent private notebook lets agents externalize learned conventions and avoid the high-capacity collapse seen in stateless agents. • A two-agent referential signaling game (sender/receiver coordinating a code from scratch) run with gpt-5.4-mini; compares five memory architectures (memory only, env board, scratchpad, codebook, codebook meta) across channel capacities from 4 to 125. • The scratchpad notebook achieves the most reliable coordination (0.867 ± 0.023 at capacity=25) while stateless "memory only" peaks at cap=25 then collapses (collision 1.0 at cap=64); the information-bottleneck point (cap=8) is a bimodal fragility point, not a compositional optimum.
2026-06-30	The Past Is Prologue: A Plug-in Controller for Selective Updates in Sequentially Evolving LLM Memory
• Janus is a method-agnostic plug-in memory controller that treats each candidate memory update as an accept/reject deployment decision, combining a Memory Momentum Trigger (when to compare old vs. new memory) with a compact hybrid evaluation set of coverage, boundary, and fresh tasks (what to compare on). • Sequentially evolving LLM memory for task-solving agents; Janus wraps existing memory updaters (e.g., DC-RS, ExpeL) without changing their update rules, using directional deviation of the memory-update trajectory to trigger bounded-cost old-vs-new validation. • Across six datasets, two LLMs (Qwen3-8B, DeepSeek-V4-Flash), and two updaters, Janus improves average accuracy by +2.7 to +4.6 points (e.g., DC-RS 79.5→83.2 and ExpeL 78.3→81.5 on Qwen3-8B).
2026-06-29	Forensic Trajectory Signatures for Agent Memory Poisoning Detection
• Discovers a mechanistically-forced behavioral invariant ("recall fact before send email") that persistent memory-poisoning attacks imprint on an LLM agent's tool-call trajectory, enabling detection from operation-only tool logs without access to memory contents, model weights, or activations. • Detects memory-channel (delayed-trigger) poisoning by extracting 19 trajectory features from trigger-session tool logs and training LR/RF/GBM classifiers, evaluated via 5-fold CV, BCa bootstrap, and leave-one-model-out hold-out on 2,520 runs across 9 models (7B–120B). • The single invariant rule alone reaches AUC=0.9563; the full Random Forest reaches AUC=0.9904 (Recall 0.984), with AUC=1.000 on 6/9 cross-model hold-outs and zero-retraining transfer to GPT-4.1/GPT-4o; a prefix-only variant hits AUC=0.934 for inline blocking.
2026-06-29	Neural Procedural Memory: Empowering LLM Agents with Implicit Activation Steering
• NPM is a training-free framework that represents agent procedural memory as implicit activation-steering vectors in the residual stream rather than explicit textual instructions, distilled from dual-granularity contrastive experiences to overcome the text-action disconnect of RAG-injected guidelines. • Procedural memory for LLM agents; pre-computes steering vectors from contrastive success/failure trajectories, then retrieves and dynamically synthesizes a task-specific vector injected at inference time to modulate reasoning and action selection without parameter updates or context expansion. • On four benchmarks (ALFWorld, WebShop, ScienceWorld, BabyAI), NPM matches/exceeds explicit textual baselines (e.g., MiniCPM3-4B avg 22.60→28.87; Qwen3-8B 30.63→36.32) and the hybrid NPM+Workflows setting is best overall (Qwen3-8B avg 41.89, ALFWorld 66.42%).
2026-06-29	Mandol: An Agglomerative Agent Memory System for Long-Term Conversations
• Mandol consolidates fragmented vector/graph memory into a unified memory-native architecture combining a hierarchical memory model, an agglomerative SemanticMap/SemanticGraph structure that natively fuses key-value/vector/graph storage, and a quantitative retrieval mechanism that runs without invoking LLMs. • Long-term cross-session conversational memory; replaces RAG-style recall-then-rank with query-adaptive routing, MAD-based denoising/conflict resolution, and MMR token-constrained context generation over an in-memory unified store (with DuckDB persistence). • Best overall accuracy on LoCoMo (92.21%) and LongMemEval (88.40%), with ~5.4× mean retrieval and ~4.8× mean insertion speedup under 10 QPS load, cutting tokens 17.4–20.0% vs. EverMemOS.
2026-06-28	Manufactured Confidence: How Memory Consolidation Turns Hearsay into Confident Facts
• Diagnoses "manufactured confidence" — memory-consolidation products (mem0, LangMem) rewrite hedged, casual remarks into confident, dated standalone "facts" that agents then obey, showing agents key on the confidence of phrasing rather than the source, needing no attacker. • Uses judge-free access-control and budget-approval agents across five models/four providers to isolate the failure, running the same poisoning protocol against mem0, LangMem, and a verbatim-storage control, and testing framings, source attribution, uncertainty tags, and a hedge-preserving extraction prompt. • mem0 and LangMem launder hedged injections into confident facts at 100% (verbatim control 0%); confident framings grant unauthorized access ~0.81 while hedges collapse to ~0.00; a redundant directory restores 0.00 wrong-grant, and hedge-preserving extraction cuts wrong-grant from 0.45 to 0.10.
2026-06-28	Selective Memory Retention for Long-Horizon LLM Agents
• TraceRetain is a lightweight capacity-bounded memory-retention framework for frozen LLM agents that scores memory entries by interpretable features (success, age, access frequency, redundancy, specificity, similarity, downstream utility) and evicts the lowest-scoring ones. • Formulates external-memory management as a capacity-constrained retention problem on ALFWorld (gpt-5-mini, ReAct-style), comparing TraceRetain-Linear/CEM against cache heuristics (FIFO/LRU/LFU/Random/Ebbinghaus) and unbounded memory under a 75%-distractor noisy-write stress. • Methods saturate on clean ALFWorld (47–49/50 vs. 39/50 no-memory); under noisy writes, unbounded and FIFO Precision@5 collapse while TraceRetain-CEM stays stable (16.9%→16.6%) and preserves 97/100 task success, with bounded K=50 matching unbounded K=100.
2026-06-27	Memory as an Attack Surface in LLM Agents: A Study on Multiple-Choice Question Answering
• Frames the external memory of LLM agents as an attack surface, showing that misleading or corrupted memories inserted through ordinary natural-language interactions can silently flip an agent's answer even when the current query is clean. • Builds a planner-guided LLM QA agent with external memory for four-option MCQ, then applies two attacks — false-information memory injection and interaction-based answer-choice steering — across ML, cybersecurity, and networking on GPT-5.4/GPT-4o mini, Gemma2-9B, and Phi3-14B. • Clean baselines average 91.85% (closed) vs. 77.10% (open); false-memory injection causes 82/1064 answer changes (7.80% ASR) with Phi3-14B most vulnerable (34.48% shift in cybersecurity); feedback reinforcement biases answers more than example exposure.
2026-06-25	Supersede: Diagnosing and Training the Memory-Update Gap in LLM Agents
• Introduces "supersession" (keeping the current value of a changed fact) as a distinct, trainable failure mode, and releases Supersede — the first RL environment whose reward directly targets temporal fact-currency rather than a proxy, reframing the FAMA metric as a dense training signal. • Handling superseded facts in long multi-session dialogue under a bounded, self-maintained memory; diagnoses the gap on the LongMemEval knowledge-update subset and trains it down via GRPO fine-tuning of Qwen2.5-3B with a programmatic supersession-aware reward. • Bounded memory drops knowledge-update accuracy 92%→77% on frontier gpt-5.4 (p=0.0033); accuracy falls further as conversations grow 24× (68%→28%); GRPO training nearly doubles held-out supersession accuracy on real unseen conversations (9.0%→16.7%).
2026-06-25	Temporal Validity in Retrieval Memory: Eliminating Stale-Fact Errors for AI Agents over Evolving Knowledge
• MemStrata maintains temporal validity via a deterministic (subject, relation, object) supersession rule in a bi-temporal ledger — retiring stale facts with no similarity threshold and no LLM call on the read path — backed by a proof that cosine similarity cannot separate contradictions from duplicates (AUROC 0.59). • Keeps agent memory current under evolving knowledge (code renames, config/dependency/API changes); stores facts like RAG for full static recall but supersedes contradicted values, evaluated fully deterministically on a local 7B model across 6 benchmarks. • Ties RAG on static knowledge yet reaches 0.95–1.00 accuracy on evolving knowledge vs. RAG's 0.20–0.47 (2–5× gain), drives stale-fact-error rate from 15–40% to ~0%, and runs ~8× faster than LLM-reranking baselines (~2.1s vs. ~16–18s).
2026-06-24	Memory Makes the Difference: Evaluating How Different Memory Roles Shape Conversational Agents
• Introduces the first fine-grained taxonomy of conversational memory by functional role (answer, clarifying, enriching, distracting, irrelevant) plus a user-centric, context-aware LLM-as-judge evaluation framework covering accuracy, relevance, and informativeness. • Controlled comparative experiments on conversational RAG over two long-term multi-session datasets (LongMemEval-m, Long-MT-Bench+) with three frontier LLMs and three retrievers, varying context size and memory-type composition. • Clarifying memory reliably improves factual accuracy; distracting memory substantially harms accuracy/relevance; irrelevant memory reduces topic relevance; performance rises then declines as context grows (information overload), and answer memory remains essential.
2026-06-23	Reasoning as Attractor Dynamics: Latent Memory Retrieval via Gibbs-Weighted Energy Minimization
• Reframes LLM reasoning as retrieval from a Dense Associative Memory — correct chains as flat-minima attractor basins, hallucinations as sharp minima — and introduces a Gibbs-Weighted Basin Selection operator that reweights sampled paths by inverse-square spectral entropy (W ∝ E⁻²). • Math reasoning (GSM8K); sample K high-temperature trajectories, compute each path's trajectory energy as length-normalized NLL, then reweight via a post-hoc Gibbs measure to relax into the dominant attractor basin. • On GSM8K with Phi-3.5-mini (3.8B), Gibbs-Weighted Retrieval (K=12) reaches 90.07% vs. Standard Sampling/majority-vote 84.69% and Greedy Decoding 78.4% — a +5.38% gain over self-consistency.
2026-06-23	ReM-MoA: Reasoning Memory Sustains Mixture-of-Agents Scaling
• A memory-augmented Mixture-of-Agents framework built on a Ranked Reasoning Memory that persistently stores/ranks cross-layer reasoning traces via a comparative Reviewer Agent, paired with Curated Diversified Memory Routing to preserve both reasoning quality and exploration diversity. • Scalable multi-agent LLM reasoning; at each layer a Reviewer Agent comparatively scores traces with rationales and later agents receive distinct high/low/contrastive trace subsets, with an optional frontier-model (GPT-5.5) LoRA distillation pipeline to upgrade the Reviewer. • Across five reasoning benchmarks (MATH, MMLU-redux, Formal Logic, CRUX, HellaSwag) it consistently beats prior MoA variants and the gap widens with depth — e.g., MATH at L=9: ReM-MoA* 84.0% vs. AttentionMoA 76.9% vs. Standard MoA 61.0%.
2026-06-19	When Does Overlap Help? OSU-Mem and a Cell-Conditional Analysis of Trajectory Memory for LLM Agents
• OSU-Mem organizes agent trajectory memory into overlapping semantic units with budgeted coarse-to-fine retrieval, and shows via cell-conditional analysis that overlap helps only when evidence steps share tool calls or entities (T+E+) and hurts when they share neither (T−E−). • Budgeted retrieval from long-horizon LLM-agent trajectories under a strict token budget; builds OSUs from entity/tool/subgoal/similarity views, then query-adaptive centroid-scored expansion, evaluated on a synthetic benchmark, τ-bench, and ToolBench with a 2×2 tool/entity cross-tabulation. • +39.9% Recall and +61.5% Hit@2 over the strongest baseline at B=256 on the synthetic benchmark; wins on T+E+ but loses on T−E− in τ-bench; on ToolBench overlap beats disjoint construction with a monotonic dose-response.
2026-05-30	Memory Shot for Long-Term Dialogue
• MemShot renders raw dialogue spans directly into structured visual "memory shots" (images preserving speaker turns, timestamps, and turn boundaries) and leverages an MLLM's internal visual reasoning, avoiding fragile, heavyweight text-based memory construction. • Long-term dialogue memory-augmented QA; segments dialogue into contiguous spans, renders each into a hierarchical header+chat visual unit, retrieves top-k units (Qwen3-VL-Embedding-8B), and answers with Qwen3-VL-Instruct MLLMs (2B/8B/32B). • Competitive-to-superior on LoCoMo (79.61 overall Acc @32B) and LongMemEval (74.80 overall Acc @32B) while delivering ~70× faster memory construction (≈9.56s), beating visual-memory baseline MemOCR by over 10%.
2026-06-25	MIRROR: Novelty-Constrained Memory-Guided MCTS Red-Teaming for Agentic RAG
• This paper studies red-teaming for multimodal agentic RAG systems whose attack surface spans retrieved text, images, direct user queries, and orchestrator-level tool manipulation. • MIRROR combines an episodic memory bank of successful attack traces with Monte Carlo tree search; retrieved memories provide search priors, while a deterministic novelty gate blocks copying from known or retrieved attacks. • Across four attack surfaces, the framework validates candidates through deterministic replay or structured tool-call parsing, aiming to make memory-guided adversarial search both more effective and less dependent on recycled templates.
2026-06-23	Escaping the Self-Confirmation Trap: An Execute-Distill-Verify Paradigm for Agentic Experience Learning
• This paper studies experience-driven self-evolution for LLM agents and identifies a Self-Confirmation Trap: single-agent loops may treat wrong-but-self-consistent trajectories as successful experience, causing erroneous memories to be retrieved and reused later. • It proposes EDV, an Execute-Distill-Verify framework in which multiple heterogeneous agents first explore the same task space, a third-party distillation agent compares candidate trajectories to produce experience candidates, and an execution group verifies them through consensus before memory insertion. • By decoupling execution, experience distillation, and verification, EDV turns isolated self-reflection into collaborative experience construction and filters noisy or erroneous content before it enters shared or private memory. Experiments on tau2-bench, Mind2Web, and MMTB show consistent improvements over strong baselines, highlighting the importance of reliable memory construction for agent self-evolution.
2026-06-18	Grouped Query Experts: Mixture-of-Experts on GQA Self-Attention
• This paper proposes Grouped Query Experts (GQE), a mixture-of-experts layer on top of grouped-query attention that targets the high cost of dense self-attention at long context lengths. • Within each GQA group, a router selects k query-head experts per token while leaving all key-value heads dense and unchanged. This preserves the KV-cache advantages of GQA while reducing active query-head computation according to token difficulty or information content. • On a fixed 30B token budget at the 250M parameter scale, GQE matches the downstream accuracy of an all-active GQA baseline while activating only half of the query heads per token, suggesting a path to more efficient long-context processing.
2026-06-18	Multi-Agent Transactive Memory
• This paper introduces Multi-Agent Transactive Memory (MATM), which organizes action–observation trajectories generated by heterogeneous agents into group-level shared memory. Producer agents contribute execution experience, while consumer agents retrieve prior trajectories, enabling the reuse of procedural knowledge that would otherwise be discarded after a single task. • MATM employs a state-conditioned key–value index, using the current task and recent interaction history as retrieval keys and subsequent trajectory segments as values. It further applies a learning-to-rank model that integrates producer reliability, consumer characteristics, retrieval scores, and trajectory attributes to perform personalized reranking. • Experiments in ALFWorld and WebArena demonstrate that retrieving shared trajectories improves task performance and reduces interaction steps without requiring direct inter-agent coordination or joint training, with benefits extending across agents of varying capability levels. The work thereby extends individual experiential memory into a collective knowledge infrastructure for open agent ecosystems.
2026-06-17	What Must Generalist Agents Remember?
• This paper formally investigates the information that generalist agents must retain to sustain near-optimal behavior across multiple environments and objectives. It defines states with identical observations but conflicting optimal actions as observational bottlenecks that reveal the necessity of memory. • The separation theorem establishes that when different domains require mutually incompatible optimal actions at such a bottleneck, any unified near-optimal policy must induce distinct memory distributions. Consequently, a memoryless policy that relies solely on the current observation cannot simultaneously maintain a high success rate and cross-domain robustness. • The paper further proves that if memory is sufficient to estimate the value functions of a set of relevant objectives, local transition dynamics can be approximately reconstructed from it. Memory therefore serves simultaneously as a mechanism for domain disambiguation, environment-model reconstruction, and a substrate for planning. These findings establish theoretical necessity but do not directly prescribe a specific engineering architecture for memory.
2026-06-17	User as Engram: Internalizing Per-User Memory as Local Parametric Edits
• This paper decomposes personalized memory into user-specific content and reasoning skills shared across users. It proposes storing user facts through localized row edits in the hash-keyed memory table of the Engram model, while a single shared adapter provides fact interpretation and indirect reasoning capabilities. • Unlike per-user LoRA modules, which exert dense effects on global weights, localized Engram edits operate only at precisely triggered positions and leave all other positions unchanged. Facts belonging to different users are written into non-overlapping hash slots, enabling additive and lossless multi-user composition within a shared table. • The paper reports that the proposed design matches per-user LoRA in direct recall, improves indirect-reasoning accuracy by an average factor of 5.6, and reduces memory consumption by approximately 33,000 times. After roughly 100 facts, it also outperforms retrieval pipelines using larger models. Its applicability, however, depends on the underlying model possessing an editable Engram memory architecture.
2026-06-16	Closing the Feedback Loop: From Experience Extraction to Insight Governance in Verbal Reinforcement Learning
• This paper identifies a retention–forgetting dilemma faced by parameter-free verbal reinforcement learning in non-stationary environments: retaining obsolete rules indefinitely induces negative transfer, whereas permanently deleting prior knowledge causes catastrophic forgetting when similar conditions recur. • The authors propose a three-layer memory architecture comprising rules, evidence, and skills, connected through a feedback-driven curation loop. Rules distill experience, evidence records the cross-episode reliability of rules, and skills govern rule selection, conflict resolution, and abstention when necessary. • A financial forecasting case study shows that the same accumulated experience can produce performance below a zero-shot baseline when governance is absent, whereas incorporating the curation loop improves both predictive accuracy and risk-adjusted returns. This finding suggests that the central bottleneck in continual agent learning is not merely experience extraction, but also the governance of knowledge lifecycles and application permissions.
2026-06-16	Memory as a Wasting Asset: Pricing Flash Endurance for Embodied Agents, and the Limits of Doing So
• This paper models the finite program/erase endurance of robotic flash storage as non-renewable depreciating capital. It introduces an endurance shadow price η and uses it to construct a wear-augmented per-byte index for optimizing memory placement across RAM, onboard non-volatile storage, and cloud infrastructure. • Theoretical analysis shows that threshold-based placement policies can achieve cost optimality under different value–write correlation levels χ. Only when χ is positive can the optimal policy become non-monotonic, moving frequently written, high-value memories away from local flash storage. Real-world logs show that χ is positive for cyclic long-horizon operations, near zero for short-horizon tasks, and negative for non-cyclic teleoperation. • Endurance constraints are generally inactive for high-end TLC devices rated for approximately 3,000 P/E cycles, but may become binding for QLC or eMMC devices rated for approximately 1,000 P/E cycles. The paper also explicitly notes that wear-aware management primarily improves device longevity and cost efficiency; it has not yet been shown to increase task value or task success rates.
2026-06-15	Posterior Twins: Distributional Behavioral Simulation for Enterprise Decisions
• This paper introduces Posterior Twins, extending the output of enterprise digital twins from a single most likely behavior to a posterior distribution over behaviors conditioned on a specific decision context. This formulation captures states such as adoption, churn, hesitation, and risk migration across different population segments. • The system uses governed historical memory as its evidence base and integrates behavioral-model routing, scenario orchestration, distribution aggregation, and auditing mechanisms. Evaluation employs both modal accuracy and Wasserstein-1 distance to distinguish point-prediction correctness from overall distributional fidelity. • On 226 held-out samples, TL-Twin Alpha achieves the lowest Wasserstein-1 distance in the reported results, at 1.16, while Gamma and Delta exhibit more balanced operating points. The study demonstrates the need for distributional evaluation in enterprise simulation, although its conclusions remain constrained by the scale of a single benchmark and the particular system configurations examined.
2026-06-15	Trust-Aware Multi-Agent Traceability: Confidence-Calibrated Knowledge Graphs for Consistent Software Artifact Management
• This paper addresses error propagation in multi-agent software-engineering pipelines by employing a shared knowledge graph as both centralized semantic memory and a coordination interface. This enables downstream agents to evaluate and inherit upstream artifacts according to calibrated confidence estimates. • The method comprises two-stage traceability-link prediction combining embedding-based retrieval with LLM-based multi-criteria analysis, a trace-seed mechanism for comparing confidence at generation and verification time, and protocols for threshold gating, confidence-disagreement detection, and conflict resolution. • An automotive software-engineering case study and ablation experiments indicate that confidence calibration is essential to the effective operation of the coordination protocol. However, because the evidence is derived primarily from a domain-specific case, the work is better understood as an architectural validation of trustworthy multi-agent artifact governance than as a universally applicable conclusion across all collaborative settings.
2026-06-15	HiMPO: Hindsight-Informed Memory Policy Optimization for Less-Entangled Credit in Long-Horizon Agents
• This paper focuses on entangled causal credit assignment in memory-writing operations for long-horizon agents. A final failure may originate from tool invocation, noisy observations, or subsequent reasoning, yet trajectory-level rewards can mistakenly penalize an earlier memory update that was itself correct. • HiMPO first compares the task-relevant information recoverable from the new and previous memories under the same pre-write state to estimate the local utility of a memory update. It then calibrates this utility using a bounded hindsight-relevance filter and applies the resulting memory-specific advantage exclusively to memory tokens. • Across open-domain evaluations and compressed-memory question-answering tasks, the method outperforms multiple memory and reinforcement-learning baselines. Controlled interventions further demonstrate that HiMPO reduces responsibility leakage caused by tool errors and improves the accuracy and interpretability of credit attribution for memory writes.
2026-06-15	User as Code: Executable Memory for Personalized Agents
• This paper proposes the “User as Code” paradigm, implementing the user model of a personalized agent as a continuously evolving software project. Typed Python objects store user state, while ordinary functions encode constraints, aggregation logic, and response rules, allowing memory representation and memory reasoning to share the same executable medium. • The system employs a two-stage pipeline comprising an append-only fact log and periodic code checkpoints. This preserves the original history while organizing it into verifiable structured state, thereby supporting contradiction resolution, cross-record aggregation, and safety rules that are proactively triggered by state changes. • UaC achieves a conventional factual recall rate of 78.8% on LOCOMO and approximately 99% on historical aggregation questions, substantially exceeding the 6%–43% performance of retrieval-based memory. Its central contribution extends beyond improved recall: it transforms user memory from a passive query database into a proactively serving system capable of deterministic execution.
2026-06-15	TokenPilot: Cache-Efficient Context Management for LLM Agents
• This paper observes that although conventional context pruning and dynamic memory eviction reduce input-token counts, they also alter sequence boundaries and prompt layouts, resulting in prefix mismatches and KV-cache invalidation. Context sparsity and cache continuity therefore exhibit a systematic tension. • TokenPilot adopts a dual-granularity management strategy. At the global level, Ingestion-Aware Compaction compresses environmental noise as information enters the context and stabilizes the prompt prefix. At the local level, Lifecycle-Aware Eviction performs conservative batched offloading according to the remaining task utility of individual context segments. • In both independent and continuous operation modes on PinchBench and Claw-Eval, the paper reports cost reductions of 56%–61% and 61%–87%, respectively, while maintaining task performance comparable to existing systems. These findings establish cache-compatible stable layout as an independent design objective for cost optimization in long-horizon agents.
2026-06-14	FragFuse: Bypassing Access Control of Large Language Model Agents via Memory-Based Query Fragmentation and Fusion
• This paper reveals that long-term memory introduces a cross-turn temporal channel into agent access control. A prohibited request that would ordinarily trigger refusal can be decomposed into superficially benign fragments, written separately into memory, and subsequently recombined during retrieval. • FragFuse comprises three stages: refusal-sensitive fragment identification, memory injection using marker-based carriers, and fusion-oriented retrieval attacks. It further uses surrogate models to optimize fusion instructions and marker design, enabling automated attack generation under a black-box threat model. • The paper reports an average access-control bypass rate of 86.3% and an end-to-end harmful-task success rate of 41.1%. Prompt-injection detectors and perplexity-based detectors provide limited defensive effectiveness. These results demonstrate that security review must encompass the complete lifecycle of memory writing, storage, retrieval, and fusion rather than inspecting only the current user query.
2026-06-14	DYNA: Dynamic Episodic Memory Networks for Augmenting Large Language Models with Temporal Knowledge Graphs in Continuous Learning
• This paper introduces DYNA, which uses a frozen large language model as its reasoning core and constructs a continuously updatable external episodic memory by representing events as nodes and temporal relations such as “before,” “after,” and “co-occurs with” as timestamped directed edges. • During querying, the system uses random walks, node centrality, and graph-structural information to identify relevant events, then combines the retrieved results with the model’s internal knowledge to generate answers. This avoids the training costs and parameter-level knowledge interference associated with continual fine-tuning. • Across three categories of temporal-memory tasks, the paper reports that DYNA reduces catastrophic forgetting by approximately 7% relative to fine-tuning and improves temporal-ordering accuracy by approximately 5% relative to standard RAG. The positive correlation between graph clustering coefficients and retrieval performance further indicates that memory topology itself is an important determinant of performance.
2026-06-12	AgentSpec: Understanding Embodied Agent Scaffolds Through Controlled Composition
• This paper introduces AgentSpec, formalizing embodied agents as typed compositions of perception, memory, reasoning, reflection, action, and optional learning modules. Standardized interfaces support the independent replacement, ablation, and controlled recombination of these components. • The study systematically compares different reasoning, memory, reflection, and reinforcement-learning components across DeliveryBench, ALFRED, MiniGrid, and RoboTHOR. The results show that agent performance is not determined by the strength of any single module, but is significantly influenced by module compatibility, task environment, and component interactions. • The experiments specifically demonstrate that structured, multi-granularity memory provides more stable support for long-horizon state tracking, although its benefits depend on alignment between the memory representation and the downstream reasoning strategy. The work thus transforms agent scaffolding from an empirically configured engineering artifact into an analyzable compositional design space.
2026-06-11	EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments
• Introduces EvoArena, a benchmark suite for simulating dynamic environment changes to evaluate memory robustness in LLM agents. • Proposes EvoMem, a patch-based memory paradigm that explicitly tracks structured memory evolution over time. • Results show that current agents perform poorly under evolving environments, while EvoMem significantly improves performance.
2026-06-11	MemRefine: LLM-Guided Compression for Long-Term Agent Memory
• Proposes MemRefine, an LLM-driven memory compression framework under fixed storage budgets. • Uses semantic and factual consistency evaluation to iteratively delete, merge, or retain memory entries. • Maintains or exceeds baseline performance under strict memory constraints.
2026-06-11	Getting Better at Working With You: Compiling User Corrections into Runtime Enforcement for Coding Agents
• Introduces TRACE, which compiles user corrections into runtime constraints for LLM agents. • Converts feedback into executable rules for continuous behavior correction. • Significantly reduces preference violations in interactive tasks.
2026-06-11	G-Long: Graph-Enhanced Memory Management for Efficient Long-Term Dialogue Agents
• Proposes G-Long, a graph-based memory system using structured triples for long-term dialogue management. • Employs lightweight LMs for structured extraction and attention-aware importance scoring. • Achieves efficient and high-quality memory retrieval with reduced computational cost.
2026-06-11	Multi-Turn Reasoning When Context Arrives in Pieces: Scalable Sharding and Memory-Augmented RL
• Addresses multi-turn reasoning failures caused by fragmented context input in dialogue systems. • Introduces a compact rolling memory mechanism replacing full history tracking. • Improves zero-shot reasoning and generalization across long-context tasks.
2026-06-10	Arbor: Tree Search as a Cognition Layer for Autonomous Agents
• Introduces Arbor, a multi-agent framework using tree search as a shared cognitive layer. • Maintains an explicit evolving search tree as shared working memory for coordination. • Improves scalability, reproducibility, and reasoning efficiency in large state spaces.
2026-06-10	Substrate Asymmetry in User-Side Memory: A Diagnostic Framework
• Decomposes user-side memory into orthogonal axes rather than treating personalization as a single capability. • Identifies three key dimensions: behavioral consistency, factual presence, and factual absence. • Reveals asymmetries across models and highlights alignment and routing trade-offs.
2026-06-10	Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents
• Proposes HORMA, a hierarchical memory organization framework for efficient navigation and retrieval. • Structures experience into layered representations to reduce loss and latency in unstructured retrieval. • Improves performance and efficiency in long-horizon dialogue tasks.
2026-06-05	Position: Hippocampal Explicit Memory Is the Cornerstone for AGI
• Argues that explicit hippocampal-like memory is fundamental for achieving AGI-level capabilities. • Claims implicit statistical learning alone is insufficient for planning, reasoning, and meta-cognition. • Outlines computational requirements for implementing explicit memory systems in AI.
2026-06-09	Trace Only What You Need: Structure-Aware On-Demand Hypergraph Memory for Long-Document Question Answering
• Introduces DocTrace, a multi-agent RAG framework with structure-aware hypergraph memory. • Builds document structure trees and on-demand shared memory graphs for reasoning reuse. • Achieves strong improvements over structured RAG baselines on long-document QA tasks.
2026-06-09	REAL: A Reasoning-Enhanced Graph Framework for Long-Term Memory Management of LLMs
• Proposes REAL, a time- and confidence-aware directed graph for long-term memory management. • Supports non-destructive updates with multi-version factual memory retention. • Achieves +22.72% average improvement in long-term memory tasks.
2026-06-09	Infini Memory: Maintainable Topic Documents for Long-Term LLM Agent Memory
• Proposes Infini Memory, which organizes agent memory into maintainable topic-centric documents. • Uses buffer-and-consolidation updates to continuously refine semantic memory over time. • Enables iterative retrieval via tool-based memory reading rather than single-pass lookup.
2026-06-09	ActiveMem: Distributed Active Memory for Long-Horizon LLM Reasoning
• Introduces ActiveMem, decoupling planning and memory systems for long-horizon reasoning tasks. • Uses lightweight planners with distributed memory modules running in parallel. • Improves efficiency while achieving state-of-the-art performance on complex benchmarks.
2026-06-08	Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents
• Proposes DCPM, a dual-process memory system separating explicit and implicit cognitive memory layers. • Uses day-time writing and night-time consolidation mechanisms for long-term evolution. • Demonstrates strong performance in cross-session reasoning and personalization tasks.
2026-06-10	MemToolAgent: Leveraging Memory for Tool Using Agents Based on Environment and User Feedback
• Proposes MemToolAgent, which enhances tool-using agents via structured memory extraction and retrieval. • Converts past interactions into reusable memory entries for continuous improvement. • Improves personalization and task accuracy through memory-augmented reasoning.
2026-06-05	AdMem: Advanced Memory for Task-solving Agents
• Proposes AdMem, a unified memory framework integrating semantic, episodic, and procedural memory. • Uses multi-agent coordination for memory generation, reward annotation, and adaptive retrieval. • Improves robustness and success rate in long-horizon task-solving scenarios.
2026-06-04	AdaMEM: Test-Time Adaptive Memory for Language Agents
• Addresses the difficulty of using past experience when language agents face dynamic test-time conditions. • Combines offline long-term trajectory memory with online short-term policy memory to guide decisions. • Shows consistent gains over static memory baselines, highlighting test-time adaptation as a practical memory direction.
2026-06-04	Ask Only When Needed: Proactive Retrieval from Memory and Skills for Experience-Driven Lifelong Agents
• Studies when a lifelong agent should consult prior experience rather than always retrieving memory. • Organizes experience into factual, episodic, and skill memory, and learns retrieval as an explicit policy action. • Improves task success while reducing unnecessary interactions, showing the value of proactive memory control.
2026-06-04	Beyond Semantic Organization: Memory as Execution State Management for Long-Horizon Agents
• Argues that semantic similarity alone fragments decision trajectories in long-horizon agent tasks. • Proposes MAGE, a hierarchical execution-state tree with grow, compress, maintain, and revise operations. • Improves task success and reduces token use by preserving valid state paths and isolating erroneous branches.
2026-06-04	Beyond Similarity: Trustworthy Memory Search for Personal AI Agents
• Identifies semantic-similarity memory search as a trust boundary for personal AI agents. • Introduces MemGate, a lightweight query-conditioned gate that decides whether retrieved memories should enter context. • Reduces cross-domain leakage, sycophancy, tool drift, and memory-induced jailbreak risks while preserving utility.
2026-06-04	EMBER: Efficient Memory via Budgeted Evidence Retention for Long-Horizon Agents
• Focuses on which evidence should survive when long-horizon agents have fixed retained-memory budgets. • Learns to store source-backed evidence capsules with retrieval keys and update metadata during ingestion. • Improves retained-evidence recall and answer quality, showing that memory quality depends on budgeted evidence survival.
2026-06-04	Membrane: A Self-Evolving Contrastive Safety Memory for LLM Agent Defense
• Targets evolving jailbreak attacks where static safety classifiers and naive memory guardrails are insufficient. • Builds contrastive safety memory cells pairing harmful queries with superficially similar benign requests. • Improves agent-level defense while reducing over-refusal and maintaining robustness under memory poisoning.
2026-06-04	Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents
• Challenges the static retrieve-then-reason pipeline used by many memory-augmented agents. • Represents memory as a Cue-Tag-Content graph and lets the agent actively reconstruct relevant paths during reasoning. • Outperforms strong baselines on long-memory benchmarks while reducing token and runtime cost.
2026-06-04	TOKI: A Bitemporal Operator Algebra for Contradiction Resolution in LLM-Agent Persistent Memory
• Treats contradiction resolution in persistent agent memory as a write-time consistency problem. • Defines bitemporal operators with explicit isolation assumptions, provenance annotations, and audit rows. • Clarifies the correctness contract needed by production memory systems when beliefs evolve or conflict.
2026-06-03	ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction
• Addresses the cost of keeping full interaction histories in long sequential decision tasks. • Learns recursively updated natural-language belief states and directly supervises their information content. • Reduces summary errors and memory footprint, narrowing the gap with full-context agents.
2026-06-03	PersonaTree: Structured Lifecycle Memory for Person Understanding in LLM Agents
• Focuses on building durable person understanding from long-term interactions. • Organizes evidence, claims, confidence, and query-conditioned paths into a structured persona tree. • Improves personalization by returning only the evidence depth needed for each query.
2026-06-03	RAMPART: Registry-based Agentic Memory with Priority-Aware Runtime Transformation
• Addresses runtime context assembly for LLM agents with explicit memory policies. • Defines a registry-based memory model with primitives such as promote, gate, write, evict, and rollback. • Shows that grouping and priority management of memory blocks can improve task success across models.
2026-06-03	Scaling Self-Evolving Agents via Parametric Memory
• Addresses the limitation that retrieval memories influence prompts but rarely change future agent behavior. • Compresses historical experience into explicit memory and uses lightweight online updates as parametric memory. • Establishes a scaling direction for agents that continuously improve after deployment.
2026-06-03	Temporal Order Matters for Agentic Memory: Segment Trees for Long-Horizon Agents
• Argues that event order is central to long-horizon memory but often lost in similarity-based organization. • Introduces SegTreeMem, an online segment-tree structure that preserves temporal order while forming hierarchical memory. • Improves long-memory answering and shows performance depends on keeping order during memory construction.
2026-06-03	Topology Matters: Measuring Memory Leakage in Multi-Agent LLMs
• Studies how graph topology affects leakage of private information across multi-agent LLM systems. • Introduces MAMA with Engram seeding and Resonance extraction phases over controlled synthetic PII documents. • Finds denser connectivity, shorter distance, and higher target centrality increase leakage risk.
2026-06-02	DMF: A Deterministic Memory Framework for Conversational AI Agents
• Targets nondeterminism and token cost in LLM-summary-based conversational memory pipelines. • Uses classical NLP, vector geometry, mathematical scoring, survival scores, and decay rules. • Achieves competitive accuracy while reducing token usage and improving reproducibility.
2026-06-02	InfoMem: Training Long-Context Memory Agents with Answer-Conditioned Information Gain
• Addresses sparse rewards in training chunk-wise memory agents for long-context tasks. • Uses answer-conditioned information gain to evaluate whether final memory supports the ground-truth answer. • Provides a more targeted training signal for deciding what information should be retained.
2026-06-02	MemTrain: Self-Supervised Context Memory Training
• Targets the lack of high-quality labeled data for training memory agents. • Uses self-supervised masked reconstruction and intermediate memory recall objectives on unlabeled corpora. • Improves memory-intensive reasoning on long-text and search-based QA tasks.
2026-06-02	RGMem: Renormalization Group-inspired Memory Evolution for Language Agents
• Targets long-term user-state modeling under evolving and potentially conflicting conversational evidence. • Uses a renormalization-group-inspired multi-scale process with hierarchical coarse-graining and thresholded updates. • Improves cross-session continuity and adaptation to changing user preferences over flat retrieval or static summaries.
2026-06-02	SaliMory: Orchestrating Cognitive Memory for Conversational Agents
• Addresses memory-related failures in conversational agents by supervising memory operations more explicitly. • Uses hierarchical staged rewards and contrastive refinement to train filtering, consolidation, and recall behavior. • Improves end-to-end accuracy and personalization while reducing memory-operation errors.
2026-06-02	Training-Free Lexical-Dense Fusion for Conversational-Memory Retrieval
• Studies retrieval over multi-session conversational history where lexical and semantic cues have complementary strengths. • Combines BM25 with dense late-interaction scoring without additional training. • Improves controllable, reproducible retrieval for multi-hop, temporal, and adversarial memory questions.
2026-06-01	DELTAMEM: Incremental Experience Memory for LLM Agents via Residual Trees
• Targets redundancy and retrieval conflict when agents accumulate many task experiences. • Organizes goal-conditioned experience and scene-level knowledge into residual trees with self-organization. • Enables compact experience reconstruction and improves performance across interactive environments.
2026-06-01	Memory Retrieval for Changing Preferences
• Addresses personalization when user preferences change and older memories may conflict with newer evidence. • Formulates retrieval as selecting historical turns that provide evidence about the latent preference state. • Outperforms embedding-only retrieval in preference-dense long-context dialogue tasks.
2026-05-31	Don't Ask the LLM to Track Freshness: A Deterministic Recipe for Memory Conflict Resolution
• Shows that LLM-based judgment is unreliable for resolving freshness conflicts in evolving memory. • Replaces direct LLM tracking with candidate extraction and version-aware deterministic aggregation. • Improves conflict resolution and clarifies that aggregation, not storage, is the main bottleneck.
2026-05-31	Honest Lying: Understanding Memory Confabulation in Reflexive Agents
• Studies how reflexive agents may store false self-explanations as persistent memory. • Introduces diagnostic signals for reflection dependence and replaces open-ended self-diagnosis with trajectory-level failure extraction. • Reduces confabulated memory use across environments and improves the reliability of reflective agents.
2026-05-31	Joint Agent Memory and Exploration Learning via Novelty Signals
• Explores the mutual dependence between memory and exploration in open-ended agent environments. • Uses novelty-driven interaction to jointly train memory and exploration policies. • Improves generalization in unseen environments while reducing token consumption.
2026-05-30	MemPro: Agentic Memory Systems as Evolvable Programs
• Challenges fixed memory construction-retrieval pipelines that stop improving after deployment. • Treats the entire memory system as an evolvable program with version trees and failure-driven edits. • Shows memory pipelines can continuously improve across benchmarks through iterative diagnosis and debugging.
2026-05-07	Belief Memory: Agent Memory Under Partial Observability
• Proposes BeliefMem, a memory framework that shifts the memory paradigm from storing deterministic conclusions to maintaining an attribute-level belief representation to combat self-reinforcing errors in partially observable environments. • Maintains multiple candidate conclusions with probabilities updated via Noisy-OR evidence merge, and features a belief-aware retrieval mechanism to preserve uncertainty for agent decision-making. • Demonstrates superior average performance on the LoCoMo and ALFWorld benchmarks over existing deterministic memory methods, showing strong memory correction capabilities and data efficiency.
2026-05-07	MemReranker: Reasoning-Aware Reranking for Agent Memory Retrieval
• Proposes MemReranker (0.6B/4B), a family of reasoning-aware reranking models for agent memory systems that moves beyond simple semantic matching by leveraging a multi-stage LLM knowledge distillation pipeline. • Combines Elo/Bradley-Terry calibrated scoring, BCE pointwise distillation, and InfoNCE contrastive fine-tuning to achieve well-calibrated relevance scores and hard-sample discrimination. • Demonstrates state-of-the-art performance on memory retrieval benchmarks (LOCOMO, LongMemEval), matching the ranking quality of larger closed-source models (e.g., GPT-4o-mini) with significantly lower inference latency.
2026-05-07	Event-Causal RAG: A Retrieval-Augmented Generation Framework for Long Video Reasoning in Complex Scenarios
• Proposes Event-Causal RAG (EC-RAG), a lightweight framework for infinite long-video reasoning that asynchronously segments video streams into semantically complete events and abstracts them into State-Event-State (SES) graph memory, replacing traditional fixed-length clip memory. • Designs a Dual-Store Memory system integrating a vector database for semantic matching and a graph database for causal-topological retrieval, achieving low-overhead storage and efficient spatiotemporal memory merging in streaming environments. • Introduces a bidirectional graph retrieval strategy to efficiently identify relevant event causal chains, significantly improving causal reasoning accuracy and preventing out-of-memory issues without requiring expensive long-sequence fine-tuning of backbone video foundation models.
2026-05-06	Tree-based Credit Assignment for Multi-Agent Memory System
• Proposes TreeMem, a tree-based reinforcement learning framework for multi-agent memory systems that derives agent-specific credit directly from final downstream rewards without task-specific annotations. • Expands each memory agent's (builder, summarizer, retrieval) outputs into multiple subsequent branches, estimating the contributions of intermediate actions via Monte Carlo averaging. • Converts coarse final rewards into granular optimization signals, enabling heterogeneous memory agents to specialize effectively and outperforming strong baselines on long-horizon benchmarks.
2026-05-05	MemFlow: Intent-Driven Memory Orchestration for Small Language Model Agents
• Proposes MemFlow, a training-free memory orchestration framework for Small Language Models (SLMs) that replaces open-ended reasoning loops with intent-driven memory routing to handle long-horizon memory tasks. • Features a specialized multi-agent pipeline (Router, Memory, Answer, and Validator) with dynamic context packing to ensure deterministic evidence preparation and grounded responses under strict token budgets.
2026-05-05	Governed Collaborative Memory as Artificial Selection in LLM-Based Multi-Agent Systems
• Proposes governed collaborative memory as an artificial selection regime for LLM-based multi-agent systems, determining which candidate memories persist as durable, shared institutional state. • Introduces a layered architecture separating agent-local, shared institutional, archive, and project-continuity memory, emphasizing memory evaluation for provenance fidelity, selection traceability, and role preservation.
2026-05-05	Learning to Forget -- Hierarchical Episodic Memory for Lifelong Robot Deployment
• Proposes H²-EMV, a hierarchical episodic memory framework for lifelong robot deployment that uses LLM-based relevance judgment and selective forgetting to manage memory scale. • Integrates user feedback to update natural-language forgetting rules, achieving personalized memory management while maintaining QA accuracy and reducing query overhead.
2026-05-05	MEMSAD: Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents
• Formalizes multiple attack scenarios for persistent external memory poisoning in retrieval-augmented agents and proposes MEMSAD, a semantic anomaly detection defense. • Proves via gradient-coupling theorem that anomaly score gradients align with retrieval objective gradients, enabling certifiable detection radii and optimal calibration sample complexity.
2026-05-05	ScrapMem: A Bio-inspired Framework for On-device Personalized Agent Memory via Optical Forgetting
• Introduces ScrapMem, a bio-inspired on-device agent memory framework that uses optical forgetting to progressively reduce old memory resolution for storage efficiency. • Constructs an Episodic Memory Graph with causal-temporal event links to maintain semantic consistency, achieving superior retrieval performance on ATM-Bench with significantly lower storage cost.
2026-05-04	A Semantic Autonomy Framework for VLM-Integrated Indoor Mobile Robots: Hybrid Deterministic Reasoning and Cross-Robot Adaptive Memory
• Proposes the Semantic Autonomy Stack (SAS), a six-layer framework for VLM-integrated indoor robots, featuring a hybrid reasoning mechanism that resolves routine instructions deterministically to bypass VLM inference latency. • Introduces a five-category semantic memory framework that compiles learned preferences into a shared digest, enabling cross-session learning and cross-robot knowledge transfer without retraining.
2026-05-04	The Dynamic Gist-Based Memory Model (DGMM): A Memory-Centric Architecture for Artificial Intelligence
• Proposes the Dynamic Gist-Based Memory Model (DGMM), a memory-centric architectural framework that represents experience as an explicit, persistent, and graph-structured episodic-semantic memory. • Formally characterizes memory operations into four distinct regimes (ingestion, consolidation, recall, and analysis), decoupling memory storage from downstream interpretation. • Defines architectural invariants such as episodic persistence and locality of cue-conditioned surprise, enabling stable memory structures to support evolving interpretation over time without retraining.
2026-05-04	MAGE: Safeguarding LLM Agents against Long-Horizon Threats via Shadow Memory
• Proposes MAGE, inspired by shadow stack in system security, maintaining an independent safety memory that distills and preserves critical security context across long task trajectories. • Evaluates pending actions against the safety memory before execution, enabling earlier detection of long-horizon attacks with minimal impact on agent utility.
2026-05-04	Symmetry-Protected Lyapunov Neutral Modes in Equivariant Recurrent Networks
• Proves that group-orbit tangent directions in equivariant recurrent networks produce symmetry-protected zero Lyapunov exponents, forming long-lived neutral memory modes. • Demonstrates through S¹, T^q, SO(n), U(m) systems and equivariant RNN experiments that strict equivariance enhances long-range memory retention, generalization, and stability.
2026-05-03	Planner Matters! An Efficient and Unbalanced Multi-agent Collaboration Framework for Long-horizon Planning
• Proposes an unbalanced multi-agent collaboration framework with planner, actor, and memory manager roles, showing that planning contributes most to task performance. • Introduces planner-only RL optimization with trajectory-level rewards, validated on web navigation, system control, and tool-use benchmarks for efficient long-horizon automation.
2026-05-02	MemORAI: Memory Organization and Retrieval via Adaptive Graph Intelligence for LLM Conversational Agents
• Proposes MemORAI, a memory organization and retrieval framework that utilizes selective memory filtering with dual-layer compression to retain user-persona-relevant content and preserve global context. • Constructs a provenance-enriched multi-relational knowledge graph to track factual origins at the turn level, enabling fine-grained and transparent memory auditing. • Introduces Dynamic Weighted PageRank for query-adaptive subgraph retrieval, applying query-conditioned edge weighting to significantly improve context-sensitive retrieval precision and personalized response generation.
2026-05-01	From Unstructured Recall to Schema-Grounded Memory: Reliable AI Memory via Iterative, Schema-Aware Extraction
• Argues AI memory should be schema-constrained rather than retrieval-based recall, proposing an iterative schema-aware write pipeline with object detection, field detection, and value extraction. • Adds verification, retry, and state control to the ingestion process, outperforming baselines on structured extraction and end-to-end memory tasks requiring stable facts and state updates.
2026-05-01	Learning How and What to Memorize: Cognition-Inspired Two-Stage Optimization for Evolving Memory
• Proposes MemCoE, a cognition-inspired two-stage optimization framework: stage one induces global memory criteria via contrastive feedback, stage two uses multi-round RL to learn criteria-compliant memory evolution strategies. • Validates on three personalized memory benchmarks, demonstrating improvements in preference memory, robustness, transferability, and efficiency.
2026-04-30	Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses
• Proposes AHE, an observability-driven closed-loop mechanism that automatically evolves coding-agent harnesses across component, experience, and decision layers. • Compresses large trajectory volumes into usable evidence via self-prediction and outcome verification, significantly improving coding-agent performance with cross-model transferability.
2026-04-30	EviMem: Evidence-Gap-Driven Iterative Retrieval for Long-Term Conversational Memory
• Introduces EviMem with IRIS closed-loop iterative retrieval that detects evidence gaps via sufficiency assessment and rewrites queries directionally for scattered multi-turn evidence. • Proposes LaceMem hierarchical memory architecture for coarse-to-fine evidence diagnosis, significantly improving temporal and multi-hop accuracy on LoCoMo while reducing retrieval latency.
2026-04-30	MemRouter: Memory-as-Embedding Routing for Long-Term Conversational Agents
• Proposes MemRouter, replacing per-turn LLM-based memory management with embedding routing to decide which turns should be written to external memory, training only ~12M parameters. • Outperforms LLM-style memory managers on LoCoMo while significantly reducing memory management latency, with retrieval pipeline, prompts, and QA backbone held constant.
2026-04-29	Detecting Clinical Discrepancies in Health Coaching Agents: A Dual-Stream Memory and Reconciliation Architecture
• Proposes a dual-stream memory architecture for health coaching LLM agents that separates patient self-reports from structured EHR/FHIR records with a reconciliation engine. • Quantifies error cascades from memory extraction and identifies discrepancy types and severity, emphasizing the necessity of memory verification in clinical longitudinal agents.
2026-04-19	Seeing Isn't Believing: Mitigating Belief Inertia via Active Intervention in Embodied Agents
• Studies belief inertia in embodied agents, where agents ignore new observations and remain anchored to prior beliefs. • Proposes Estimate-Verify-Update, which predicts, verifies, and evidence-updates textual belief states to actively manage beliefs and improve success rates across embodied benchmarks.
2026-04-22	Memanto: Typed Semantic Memory with Information-Theoretic Retrieval for Long-Horizon Agents
• Proposes Memanto with 13-class typed semantic memory schema, achieving 89.8% on LongMemEval and 87.1% on LoCoMo without knowledge graphs. • Introduces Moorcheh Information-Theoretic Search for index-free, deterministic retrieval under 90ms with zero write latency. • Addresses constraint drift through built-in conflict resolution and temporal versioning with supersede mechanism.
2026-04-19	Memory Intelligence Agent
• Proposes MIA, a Memory Intelligence Agent framework featuring a Manager-Planner-Executor architecture that addresses storage and retrieval bottlenecks by compressing historical search trajectories into structured non-parametric memory. • Introduces a two-stage alternating reinforcement learning paradigm to align high-level planning with low-level execution, alongside a continual test-time learning mechanism for on-the-fly parametric memory updates. • Incorporates reflection and unsupervised evaluation mechanisms, achieving state-of-the-art performance in deep research tasks and demonstrating strong self-evolution capabilities across diverse benchmarks.
2026-04-26	ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems
• Introduces ZenBrain, a seven-layer long-term memory architecture for autonomous agents that combines neuroscience-inspired mechanisms — Hebbian learning, FSRS spaced repetition, sleep-cycle consolidation, and Bayesian confidence propagation — released as an open-source, zero-dependency TypeScript library (Apache-2.0). • On LongMemEval-500 it attains 91.3% of oracle accuracy at 1/106 of the token budget, with 9/9 Bonferroni-corrected head-to-head wins over Letta, Mem0, and A-Mem.
2026-04-18	HeLa-Mem: Hebbian Learning and Associative Memory for LLM Agents
• Introduces HeLa-Mem, modeling long-term LLM-agent memory as a dynamic graph inspired by association, consolidation, and spreading activation. • Combines an episodic memory graph with semantic knowledge distilled through Hebbian Distillation, improving LoCoMo performance while saving context tokens.
2026-04-18	Freshness-Aware Prioritized Experience Replay for LLM/VLM Reinforcement Learning
• Proposes Freshness-Aware Prioritized Experience Replay, adding exponential age decay to priority sampling. • Mitigates stale priorities as LLM/VLM policies change quickly, outperforming standard on-policy methods and ordinary PER on multi-step reasoning, agent, and math tasks.
2026-04-18	MEMRES: A Memory-Augmented Resolver with Confidence Cascade for Agentic Python Dependency Resolution
• Introduces MEMRES, a memory-augmented agentic Python dependency resolver that uses a multi-level confidence cascade and treats the LLM as a last resort. • Combines self-evolving memory, an error-pattern knowledge base, semantic import analysis, and Python 2 heuristics to improve dependency resolution success.
2026-04-18	Visual Inception: Compromising Long-term Planning in Agentic Recommenders via Multimodal Memory Poisoning
• Introduces Visual Inception, an attack that embeds triggers in user-uploaded images to poison long-term memory in agentic recommender systems and hijack future planning chains. • Proposes CognitiveGuard, a dual-process defense combining perceptual sanitization and counterfactual consistency checks to detect abnormal memory-driven planning and reduce risk.
2026-04-18	On Safety Risks in Experience-Driven Self-Evolving Agents
• Studies safety risks in experience-driven self-evolving LLM agents across web and embodied environments, showing that even benign-task experience can reinforce action-taking over refusal. • In mixed benign/malicious tasks, refusal-related experience mitigates safety degradation but causes over-refusal, exposing a safety-utility trade-off in evolving experience memory.
2026-04-18	OASIS: On-Demand Hierarchical Event Memory for Streaming Video Reasoning
• Introduces OASIS, an on-demand hierarchical event memory framework for streaming video reasoning that organizes long histories into event hierarchies. • Performs short-context reasoning first and triggers semantic retrieval under uncertainty, using high-level intent to guide memory access while reducing noise and token cost.
2026-04-17	AdaExplore: Failure-Driven Adaptation and Diversity-Preserving Search for Efficient Kernel Generation
• Introduces AdaExplore, which distills repeated execution failures into reusable validity-rule memories for future Triton kernel generation. • Uses tree-structured candidate organization, local repair, and structural regeneration for diversity-preserving search, improving both correctness and performance optimization.
2026-04-17	StageMem: Lifecycle-Managed Memory for Language Models
• Introduces StageMem, treating language-model memory as a stateful lifecycle process rather than static storage. • Separates memory into transient, working, and persistent layers, using confidence and strength to support low-cost writing, promotion, updating, and eviction while reducing deep-memory pollution.
2026-04-17	Federation over Text: Insight Sharing for Multi-Agent Reasoning
• Proposes Federation over Text, where agents upload reasoning traces after independent work and a central server semantically aggregates them. • Builds a shared insight repository across tasks and domains, improving downstream accuracy and reducing reasoning cost without gradient updates or supervision.
2026-04-13	The Past Is Not Past: Memory-Enhanced Dynamic Reward Shaping
• Proposes MEDS, a memory-enhanced dynamic reward shaping framework that incorporates historical behavioral signals into reward design to mitigate error collapse during LLM reinforcement learning. • Constructs an error memory by reusing layer-wise logits as lightweight representations of reasoning trajectories, applying HDBSCAN clustering to dynamically penalize frequently recurring failure patterns. • Demonstrates consistent improvements in both reasoning performance and exploration diversity across multiple math reasoning benchmarks and base models.
2026-04-11	MemCoT: Test-Time Scaling through Memory-Driven Chain-of-Thought
• Proposes MemCoT, a test-time memory scaling framework that recasts long-context reasoning as iterative, stateful information search. • Uses multi-view long-term memory perception to surface evidence and a task-conditioned short-term memory module to track search history and refine subsequent queries. • Achieves state-of-the-art results across several long-memory reasoning benchmarks.
2026-04-11	SinkTrack: Attention Sink based Context Anchoring for Large Language Models
• Introduces SinkTrack, a training-free context anchoring method that exploits the model's attention sink at <BOS> to preserve critical context. • Injects key information into the <BOS> representation so the model can retain initial context throughout generation instead of gradually forgetting it. • Delivers consistent gains on text and multimodal tasks while reducing hallucination and context forgetting.
2026-04-11	Self-Distilled Reinforcement Learning for Co-Evolving Agentic Recommender Systems
• Proposes CoARS, a self-distilled reinforcement learning framework for co-evolving agentic recommender systems. • Replaces reliance on external text memory with interaction rewards and self-distilled credit assignment that convert multi-turn supervision into parameter updates. • Improves recommendation quality and user alignment over memory-only baselines.
2026-04-11	CodeComp: Structural KV Cache Compression for Agentic Coding
• Introduces CodeComp, a structural KV-cache compression method for agentic coding tasks. • Combines static program analysis with code property graphs to preserve structurally critical tokens that attention-only pruning often removes. • Outperforms attention-based compression under the same memory budget and integrates into the SGLang pipeline without changing the model.
2026-04-11	ClawVM: Harness-Managed Virtual Memory for Stateful Tool-Using LLM Agents
• Proposes ClawVM, a harness-managed virtual memory layer for stateful tool-using LLM agents. • Organizes agent state as typed pages with multi-resolution representations and validates writeback at lifecycle boundaries to avoid stale or destructive state updates. • Significantly reduces controllable failures while adding very little policy-engine overhead.
2026-04-11	CodaRAG: Connecting the Dots with Associativity Inspired by Complementary Learning
• Presents CodaRAG, an associative retrieval framework inspired by complementary learning for RAG systems. • Consolidates fragmented knowledge, navigates semantic and functional links to recover evidence chains, and suppresses distractors during retrieval. • Improves both retrieval recall and generation accuracy over prior methods.
2026-04-11	Astrolabe: A Content-Addressable Hypergraph for Semantic Knowledge Management
• Introduces Astrolabe, a content-addressable hypergraph for semantic knowledge management. • Uses content hashes and ordered reference lists to store structured knowledge, while a plugin system interprets records and decomposes structures by width or depth. • Demonstrates the extensibility of the design through plugins that bridge informal and formal mathematics.
2026-04-10	MemReader: From Passive to Active Extraction for Long-Term Agent Memory
• Proposes the MemReader family for long-term agent memory extraction, covering both passive structured extraction and active write decisions. • MemReader-4B uses GRPO and a ReAct-style process to decide whether information should be written, deferred, retrieved, or discarded. • Outperforms prior methods on knowledge updating, temporal reasoning, and hallucination reduction, and has been integrated into MemOS.
2026-04-10	HyperMem: Hypergraph Memory for Long-Term Conversations
• Introduces HyperMem, a hypergraph-based memory architecture for long-term conversation. • Models higher-order relations through hyperedges, organizes memory into topic, event, and fact layers, and couples lexical-semantic indexing with coarse-to-fine retrieval. • Improves recall and consistency on long-conversation tasks and achieves the best results on LoCoMo.
2026-04-10	StreamMeCo: Long-Term Agent Memory Compression for Efficient Streaming Video Understanding
• Proposes StreamMeCo for long-term memory compression in streaming video understanding. • Compresses memory graphs through edge-aware pruning and isolated-node sampling, and adds time-decayed retrieval to offset compression loss. • Delivers large memory savings and faster retrieval while maintaining or improving accuracy on multiple benchmarks.
2026-04-10	Towards Lifelong Aerial Autonomy: Geometric Memory Management for Continual Visual Place Recognition in Dynamic Environments
• Introduces a geometric memory management framework for continual visual place recognition in dynamic environments. • Decomposes knowledge into static satellite anchors and a dynamic replay buffer, then allocates limited storage with spatially constrained sampling. • Improves long-term retention and spatial generalization, and establishes a 21-task benchmark for evaluation.
2026-04-10	SPASM: Stable Persona-driven Agent Simulation for Multi-turn Dialogue Generation
• Proposes SPASM, a stable persona-driven framework for multi-turn dialogue generation. • Uses persona creation, client-responder simulation, and an ECP representation that stores dialogue history in a viewpoint-invariant form before projecting it back to each agent. • Reduces persona drift, role confusion, and echoing across long conversations and releases a large-scale dialogue dataset.
2026-04-10	SkillMOO: Multi-Objective Optimization of Agent Skills for Software Engineering
• Introduces SkillMOO, a multi-objective optimization framework for software engineering agent skills. • Combines LLM-generated edits with NSGA-II selection to evolve skill bundles based on task failures and cost-performance trade-offs. • Improves pass rates while lowering cost, and shows that replacing or pruning instructions can outperform simply accumulating them.
2026-04-10	RecaLLM: Addressing the Lost-in-Thought Phenomenon with Explicit In-Context Retrieval
• Proposes RecaLLM, which interleaves explicit in-context retrieval with reasoning to address the lost-in-thought problem. • Uses constrained decoding to copy evidence spans and trains on both lexical and semantic retrieval tasks. • Outperforms baselines on long-context benchmarks such as RULER and HELMET and scales to 128K contexts with relatively modest training data.
2026-04-10	Constraint-Aware Corrective Memory for Language-Based Drug Discovery Agents
• Introduces a constraint-aware corrective memory framework for language-based drug discovery agents. • Audits protocols and diagnoses constraint violations, then writes compact corrective feedback into static, dynamic, and correction memory channels. • Significantly improves task success rates by keeping planning context compact and actionable.
2026-04-10	ADAM: A Systematic Data Extraction Attack on Agent Memory via Adaptive Querying
• Presents ADAM, a systematic data extraction attack on agent memory via adaptive querying. • Estimates the memory distribution of the victim and uses entropy-guided queries to maximize leakage. • Achieves much higher attack success than prior methods, reaching up to 100% in some settings and highlighting major privacy risks in agent memory.
2026-04-10	EE-MCP: Self-Evolving MCP-GUI Agents via Automated Environment Generation and Experience Learning
• Proposes EE-MCP, a self-evolving framework for MCP-GUI agents. • Builds a closed loop of environment generation, trajectory collection, task synthesis, and quality filtering, then stores distilled rules in an experience bank for inference-time improvement. • Shows that experience-enhanced agents outperform pure distillation in GUI-intensive settings without requiring model fine-tuning.
2026-04-10	MEMENTO: Teaching LLMs to Manage Their Own Context
• Introduces MEMENTO, which teaches LLMs to manage their own context by summarizing long reasoning traces into dense intermediate mementos. • Splits reasoning into blocks, generates compact summaries for each block, and continues reasoning from the summaries instead of the full trace. • Reduces context length, KV cache usage, and compute cost while preserving reasoning quality, and releases the OpenMementos dataset.
2026-04-10	Formal Architecture Descriptors as Navigation Primitives for AI Coding Agents
• Studies formal architecture descriptors as navigation primitives for AI coding agents. • Shows that structured architecture context can reduce aimless repository exploration and improve behavioral consistency across agent runs. • Compares JSON, YAML, and S-expression formats, and introduces intent.lisp and the Forge toolchain.
2026-04-09	Task-Adaptive Retrieval over Agentic Multi-Modal Web Histories via Learned Graph Memory
• Introduces ACGM, a learned graph memory retriever for task-adaptive retrieval over long multimodal web histories. • Optimizes sparse relevance graphs with policy gradients and models different temporal decay patterns across visual and textual observations. • Outperforms multiple baselines on WebShop, VisualWebArena, and Mind2Web.
2026-04-09	TSUBASA: Improving Long-Horizon Personalization via Evolving Memory and Self-Learning with Context Distillation
• Proposes TSUBASA for long-horizon personalization with evolving memory and self-learning. • Improves memory writing through dynamic memory evolution and strengthens memory reading through context-distilled self-learning that internalizes user experience. • Outperforms systems such as Mem0 on long-horizon personalization benchmarks while achieving a better quality-efficiency trade-off.
2026-04-09	Dynamic Attentional Context Scoping: Agent-Triggered Focus Sessions for Isolated Per-Agent Steering in Multi-Agent LLM Orchestration
• Introduces DACS, a dynamic attentional context scoping mechanism for multi-agent LLM orchestration. • Keeps only lightweight state summaries in registry mode and switches into a focus mode that injects full context for the selected agent while compressing others. • Improves steering accuracy and reduces interference from irrelevant agents.
2026-04-09	LogAct: Enabling Agentic Reliability via Shared Logs
• Proposes LogAct, a shared-log execution framework that turns agents into state machines over an explicit log. • Makes planned actions visible before execution, enabling intervention, recovery, semantic replay, and LLM-based introspection from historical traces. • Improves reliability, fault recovery, and debugging for agentic systems.
2026-04-09	PASK: Toward Intent-Aware Proactive Agents with Long-Term Memory
• Introduces the DD-MM-PAS paradigm and the Pask system for intent-aware proactive agents with long-term memory. • Combines IntentFlow for latent-need detection, a hybrid workspace-user-global memory design, and a full perception-memory-action infrastructure. • Achieves deeper user-intent recognition under low-latency constraints and releases LatentNeeds-Bench.
2026-04-09	Beyond Stochastic Exploration: What Makes Training Data Valuable for Agentic Search
• Proposes HiExp, a framework that extracts reusable experience knowledge from raw agentic search trajectories. • Uses comparative analysis and hierarchical clustering to organize trajectories into layered experience representations and trains agents with experience alignment. • Turns stochastic exploration into more strategic search and improves stability, task performance, and cross-task generalization.
2026-04-09	"Theater of Mind" for LLMs: A Cognitive Architecture Based on Global Workspace Theory
• Introduces Global Workspace Agents, a cognitive architecture for LLM agents inspired by global workspace theory. • Combines a central broadcast hub with heterogeneous agents, plus entropy-driven intrinsic motivation and a two-tier memory branching strategy. • Improves semantic diversity and long-term cognitive continuity in multi-agent execution.
2026-04-09	ACF: A Collaborative Framework for Agent Covert Communication under Cognitive Asymmetry
• Proposes ACF, a collaborative framework for covert communication among memory-augmented agents under cognitive asymmetry. • Decouples covert signaling from semantic reasoning and uses prefix-independent decoding with shared steganographic configurations to avoid symmetry assumptions. • Maintains semantic consistency and covert communication quality even under severe asymmetry, with theoretical error and capacity guarantees.
2026-04-09	Distributed Multi-Layer Editing for Rule-Level Knowledge in Large Language Models
• Studies rule-level knowledge editing in LLMs and shows that rule knowledge is distributed across layers rather than concentrated locally. • Extends the RuleEdit benchmark and uses causal tracing to distinguish how formulas, descriptions, and instances are represented in different layers. • Proposes DMLE, which improves rule transfer and understanding while remaining competitive on standard editing metrics.
2026-04-09	SkillClaw: Let Skills Evolve Collectively with Agentic Evolver
• Proposes SkillClaw for collective skill evolution in multi-user LLM agent environments. • Continuously aggregates user trajectories, detects repeated behaviors and failure modes, and converts them into skill revisions or new skills for a shared repository. • Enables cross-user experience transfer and improves performance in real agent settings.
2026-04-09	Verify Before You Commit: Towards Faithful Reasoning in LLM Agents via Self-Auditing
• Introduces SAVeR, a self-auditing framework for faithful reasoning in LLM agents. • Generates diverse candidate beliefs, audits them adversarially to locate violations, and applies minimal constrained edits before the agent commits to action. • Reduces behavior drift caused by unfaithful reasoning while preserving task performance.
2026-04-09	SkillForge: Forging Domain-Specific, Self-Evolving Agent Skills in Cloud Technical Support
• Proposes SkillForge, a framework for building and self-evolving domain-specific agent skills in cloud technical support. • Combines domain-context skill creation with a closed loop of failure analysis, skill diagnosis, and skill rewriting driven by deployment feedback. • Improves initial skill quality and continues to raise performance over repeated deployment cycles.
2026-04-09	Efficient RL Training for LLMs with Experience Replay
• Revisits experience replay for LLM post-training and questions the assumption that only fresh on-policy data is useful. • Analyzes the trade-off between staleness, diversity, and generation cost in replay buffer design. • Shows that well-designed replay buffers can cut inference cost substantially without hurting and sometimes even improving performance.
2026-04-09	Artifacts as Memory Beyond the Agent Boundary
• Formalizes the idea that the environment itself can serve as external memory for an agent. • Introduces artifacts as observations that reduce the information needed to represent history and connects the idea to reinforcement learning theory. • Shows that observing spatial paths can lower the internal memory needed to learn strong policies and motivates future environment-backed memory designs.
2026-04-09	MT-OSC: Path for LLMs that Get Lost in Multi-Turn Conversation
• Proposes MT-OSC, a framework that condenses multi-turn chat history in the background to preserve salient information. • Uses a few-shot condenser and a lightweight decision module to retain key facts while compressing irrelevant content. • Cuts token usage by up to 72% while maintaining or improving multi-turn dialogue accuracy and latency.
2026-04-09	M^\star: Every Task Deserves Its Own Memory Harness
• Proposes M^\star, which represents agent memory as executable Python memory programs rather than fixed schemas. • Evolves task-specific memory data structures, storage logic, and interaction workflows through reflective code evolution guided by population search and failure analysis. • Outperforms fixed memory baselines across dialogue, embodied planning, and expert reasoning benchmarks while discovering markedly different memory designs per task.
2026-04-08	From Business Events to Auditable Decisions: Ontology-Governed Graph Simulation for Enterprise AI
• Proposes LOM-action, an ontology-governed graph simulation framework for auditable enterprise AI decisions. • Triggers scenario conditions from business events, performs deterministic graph transformations in isolated sandboxes, and bases decisions only on the resulting scenario-valid graph. • Improves accuracy and tool-chain F1 over baselines while generating fully traceable audit logs.
2026-04-08	LAST: Leveraging Tools as Hints to Enhance Spatial Reasoning for Multimodal Large Language Models
• Proposes LAST, a tool-as-hints framework for improving multimodal spatial reasoning. • Packages heterogeneous tool invocations into reusable spatial skills through LAST-Box and returns multimodal prompts that can be consumed directly by the model. • Uses a three-stage curriculum to teach adaptive tool use and significantly improves performance on four spatial reasoning datasets.
2026-04-04	LightThinker++: From Reasoning Compression to Memory Management
• The paper introduces LightThinker, a method that achieves efficiency through representation-level thought compression by condensing lengthy reasoning chains into compact hidden states via gist tokens. • It proposes LightThinker++, an Explicit Adaptive Memory Management framework that utilizes behavioral primitives such as commit, expand, and fold to dynamically regulate context resolution and mitigate information loss in complex scenarios. • The work establishes a specialized trajectory synthesis pipeline to train purposeful memory scheduling, enabling significant reductions in peak token usage and performance gains in both standard reasoning and long-horizon agentic tasks.
2026-04-18	Experience Compression Spectrum: Unifying Memory, Skills, and Rules in LLM Agents
• Unifies agent memory, skill discovery, and rule learning as different levels of experience compression, defining a compression function C_L mapping traces to knowledge at levels L0-L3 (raw trace → episodic memory → procedural skill → declarative rule). • Reveals that the memory and skill communities are deeply fragmented with only <1% cross-citation rate among 22 core papers (1136 citations), and proposes a full-spectrum agent learning system that adaptively selects compression granularity across levels. • Optimization targets include reducing context consumption, lowering retrieval latency, and improving cross-task/model/scenario migration ability; the workflow follows a cyclic pattern where new problems go to memory, repeated patterns compress to skills, and cross-scenario principles become rules.
2026-04-03	Poison Once, Exploit Forever: Environment-Injected Memory Poisoning Attacks on Web Agents
• The paper introduces Environment-injected Trajectory-based Agent Memory Poisoning (eTAMP), the first attack to achieve cross-session and cross-site compromise of web agents through environmental observations without requiring direct memory access. • It discovers Frustration Exploitation, a phenomenon where environmental stress and task failures significantly amplify an agent's susceptibility to malicious instructions by up to eight times. • The work introduces Chaos Monkey, a methodology inspired by chaos engineering, to systematically evaluate the robustness of LLM-based web agents under realistic deployment conditions such as network latency and input errors.
2026-04-03	Aligning Progress and Feasibility: A Neuro-Symbolic Dual Memory Framework for Long-Horizon LLM Agents
• Identifying long-horizon agent failures as arising from two coupled but distinct objectives: global progress alignment and local feasibility alignment. • Proposing the Neuro-Symbolic Dual Memory Framework that instantiates a neural Progress Memory for stage-aware semantic guidance and a symbolic Feasibility Memory for executable action verification. • Demonstrating superior performance across diverse benchmarks including ALFWorld, WebShop, and TextCraft while significantly reducing invalid action rates and trajectory lengths.
2026-04-02	ByteRover: Agent-Native Memory Through LLM-Curated Hierarchical Context
• Proposes an agent-native memory architecture where the LLM itself curates, structures, and retrieves knowledge, eliminating the architectural separation and subsequent semantic drift between reasoning agents and external storage pipelines. • Introduces the Context Tree, a hierarchical file-based knowledge graph managed by an Adaptive Knowledge Lifecycle (AKL) that employs importance scoring, maturity tiers, and recency decay to evolve knowledge over time. • Designs a 5-tier progressive retrieval strategy that minimizes latency by resolving queries through multi-level caching and indexing before escalating to agentic reasoning for novel questions.
2026-04-02	MemFactory: Unified Inference & Training Framework for Agent Memory
• MemFactory presents the first comprehensive framework that unifies the training, evaluation, and inference pipelines for memory-augmented AI agents. • The framework features a highly modular architecture that decouples the memory lifecycle into atomic, plug-and-play components such as Extractors, Updaters, and Retrievers. • It natively incorporates Group Relative Policy Optimization (GRPO) to facilitate efficient fine-tuning of internal memory management policies driven by multi-dimensional environmental rewards.
2026-04-02	MEMRERANK: Preference Memory for Personalized Product Reranking
• Introduction of a new benchmark for personalized product reranking that incorporates user purchase histories, candidate sets, and human-annotated relevance labels. • Development of the MEMRERANK framework that distills long purchase histories into structured, query-independent within-category and cross-category preference memory. • Implementation of a reinforcement learning post-training objective to optimize the preference memory extractor for downstream reranking utility.
2026-04-02	OMNI-SIMPLEMEM: Autoresearch-Guided Discovery of Lifelong Multimodal Agent Memory
• OMNI-SIMPLEMEM is proposed as a unified multimodal memory framework for lifelong AI agents that utilizes selective ingestion , progressive pyramid retrieval , and structured knowledge graph augmentation. • The system architecture and configurations were autonomously discovered and optimized through AUTORESEARCHCLAW , an autonomous research pipeline capable of code modification, bug diagnosis, and architectural redesign fundamentally beyond the reach of traditional AutoML. • The framework achieves state-of-the-art results on LoCoMo and Mem-Gallery benchmarks , demonstrating that autonomously identified bug fixes and architectural changes contribute more significantly to performance gains than cumulative hyperparameter tuning.
2026-04-02	SelRoute: Query-Type-Aware Routing for Long-Term Conversational Memory Retrieval
• Introduction of SelRoute, a selective routing framework that assigns queries to specialized retrieval pipelines based on their specific query types to optimize long-term conversational memory retrieval. • Identification of the enrichment-embedding asymmetry, demonstrating that storage-time vocabulary expansion improves lexical search performance while simultaneously degrading embedding search quality. • Achievement of state-of-the-art results on LongMemEval_M using a CPU-only architecture that requires no LLM inference at query time and generalizes across multiple benchmarks.
2026-03-31	Multi-Layered Memory Architectures for LLM Agents: An Experimental Evaluation of Long-Term Context Retention
• Proposes a Multi-Layer Memory Framework (MLMF) that decomposes dialogue history into working, episodic, and semantic layers to separate short-term interaction from long-term abstraction. • Introduces an adaptive retrieval gating mechanism and a retention stability objective to regulate semantic drift and preserve persona consistency across extended sessions. • Validates the framework through experiments on long-horizon benchmarks, achieving superior retention stability and multi-hop reasoning while reducing context usage and false memory rates.
2026-03-31	OBLIVION: Self-Adaptive Agentic Memory Control through Decay-Driven Activation
• OBLIVION introduces a read/write decoupled memory control paradigm that separates the decision of when to retrieve information from the selection of what to reinforce. • The framework implements a hierarchical memory structure consisting of L1 procedural, L2 semantic, and L3 episodic layers integrated with Ebbinghaus-inspired decay-driven activation to manage accessibility without explicit deletion. • Empirical evaluations on static and dynamic benchmarks demonstrate that self-adaptive memory control effectively balances learning and forgetting while significantly reducing interference and computational costs in long-horizon interactions.
2026-03-30	GEMS: Agent-Native Multimodal Generation with Memory and Skills
• GEMS is proposed as an agent-native multimodal generation framework that utilizes a structured multi-agent loop for iterative refinement and closed-loop optimization. • A persistent Agent Memory mechanism is introduced to manage historical context through hierarchical compression, preserving factual artifacts while distilling high-level strategic experiences. • An extensible Agent Skill module is developed to provide domain-specific expertise via an on-demand loading mechanism, effectively addressing specialized downstream applications.
2026-03-30	Understand and Accelerate Memory Processing Pipeline for Disaggregated LLM Inference
• Diverse long-context large language model (LLM) inference optimizations are unified into a common four-step memory processing pipeline, which is identified as a primary performance bottleneck through systematic profiling. • This research characterizes the quantitative and qualitative computational heterogeneity across pipeline stages, distinguishing between compute-bound regular operations and memory-bound irregular tasks. • A GPU-FPGA heterogeneous system is developed to accelerate inference by offloading irregular and memory-bounded operations to FPGAs while retaining compute-intensive tasks on GPUs, achieving significant speedups and energy savings.
2026-03-28	Codebase-Memory: Tree-Sitter-Based Knowledge Graphs for LLM Code Exploration via MCP
• CODEBASE-MEMORY provides a persistent, Tree-Sitter-based knowledge graph architecture supporting 66 languages through a multi-phase parallel build pipeline and incremental synchronization. • The system introduces an MCP-based tool interface exposing 14 structural query tools, such as call-path tracing and impact analysis, with sub-millisecond query latency. • Empirical evaluation across 31 repositories demonstrates that the approach achieves competitive answer quality while reducing token consumption by ten times and tool calls by 2.1 times compared to traditional file-exploration agents.
2026-03-27	Scaling Teams or Scaling Time? Memory Enabled Lifelong Learning in LLM Multi-Agent Systems
• A joint scaling perspective is introduced for LLM multi-agent systems that connects team size and lifelong learning as an interacting scaling space rather than isolated dimensions. • LLMA-Mem is proposed as a lifelong memory framework that integrates episodic, procedural, and transactive memory modules to enable cross-task transfer and coordination modeling under flexible memory topologies. • Systematic empirical evaluations reveal a non-monotonic scaling landscape where effective memory design enables smaller teams to achieve superior long-horizon performance and token efficiency compared to larger collectives.
2026-03-27	MemBoost: A Memory-Boosted Framework for Cost-Aware LLM Inference
• Proposes MemBoost, a memory-boosted LLM serving framework that integrates an Associative Memory Engine, a Meta Controller, and a Large-LLM Oracle to optimize the cost-quality trade-off. • Introduces a retrieve-or-escalate decision loop combined with a continuous write-back mechanism to enable efficient semantic answer reuse and memory growth. • Demonstrates through experiments on the MMLU-Pro dataset that the framework significantly reduces inference costs and latency while achieving accuracy comparable to or exceeding the oracle model.
2026-03-20	PersonaVLM — Long-Term Personalized Multimodal LLMs
• Proposes PersonaVLM, an innovative personalized multimodal agent framework that transforms a general-purpose MLLM (e.g., Qwen2.5-VL) into a personalized assistant via proactive memory management and self-evolving personality alignment. • Implements a multi-type memory architecture (core, semantic, episodic, procedural) for complex multi-turn reasoning, with proactive remembering that automatically extracts and summarizes multimodal interaction fragments into a persistent personalized database. • Introduces Momentum-based Personality Evolving (PEM) mechanism for consistent response generation, and releases Persona-MME benchmark (2,000+ cases across 14 fine-grained tasks, CVPR 2026) along with full model weights, 80k+ training samples, and evaluation code.
2026-03-19	MemMA: Coordinating the Memory Cycle through Multi-Agent Reasoning and In-Situ Self-Evolution
• Proposes MEMMA, a plug-and-play multi-agent framework that coordinates the memory cycle along both the forward and backward paths to address strategic blindness and delayed feedback. • Introduces a Meta-Thinker for strategic reasoning during construction and iterative retrieval, and designs an in-situ self-evolving memory mechanism to convert downstream probe failures into direct memory repairs.
2026-03-11	Governing Evolving Memory in LLM Agents: Risks, Mechanisms, and the Stability and Safety Governed Memory (SSGM) Framework
• The memory systems of large-model agents are shifting from static retrieval to dynamic autonomous updating, which enhances agent adaptability but also raises serious stability and security concerns. These include semantic drift, program drift that solidifies erroneous workflows, and memory poisoning caused by malicious external injections. • It proposes the Stable and Secure Governance Memory (SSGM) framework. The core design principle of this framework is to completely decouple an agent’s “generative cognitive strategies” from the underlying memory storage medium. Between the two, it introduces an actively intercepting governance middleware, so that memory updates are no longer blindly written directly, but must instead pass through multiple gateway checks. • Pre-merge validation performs logical consistency checks before writes, rejecting updates that contradict core facts to prevent hallucinations from being solidified. Temporal and permission filtering combines decay functions at read time to filter out outdated or invalid data and uses access control to prevent cross-user privacy leakage. Reversible periodic alignment adopts a dual-track storage structure of “mutable activity graph + immutable situational log,” whereby the system regularly aligns current memory with the immutable log and rolls back errors, thereby imposing a strict mathematical upper bound on long-term semantic drift.
2026-03-11	Taming OpenClaw: Security Analysis and Mitigation of Autonomous LLM Agent Threats
• The paper systematically analyzes the security threats faced by autonomous large language model agents such as OpenClaw across five lifecycle stages: initialization, input, reasoning, decision-making, and execution. • A detailed case study on OpenClaw demonstrates the destructiveness of these threats. For example, an attacker can turn transient malicious inputs into long-term behavioral control through “memory poisoning”; and during the decision-making and execution stages, ambiguous instructions may trigger “intent drift,” causing the agent to escalate a simple safety-check task into destructive firewall modifications and high-risk command execution. • Mitigation strategies include: plugin verification and signing in the initialization phase; semantic firewall isolation in the input phase; dynamic memory integrity checks and state rollback in the reasoning phase; intent consistency verification in the decision-making phase; and kernel-level sandboxing and least-privilege control in the execution phase.
2026-03-12	ReMem-VLA: Empowering Vision-Language-Action Model with Memory via Dual-Level Recurrent Queries
• ReMem-VLA introduces two sets of learnable recurrent memory queries: frame-level queries, which are updated frame by frame to capture short-term memory, and chunk-level queries, which are updated over longer temporal spans to stably maintain long-term memory. • A visual prediction head is added, introducing past observation prediction as an auxiliary training objective, which forces the model to recall visual details by reconstructing historical RGB frames. To address the batching challenge of recurrent training on variable-length sequences, a slot-based streaming training paradigm is proposed, which preserves temporal continuity while preventing state leakage across episodes. • To overcome the bottleneck of traditional truncated backpropagation through time on long-sequence optimization, the model creatively adopts a gradient-free recurrent update path that combines a frozen VLM with a fixed exponential moving average, allowing the queries to focus solely on learning “what task-relevant information to extract” rather than “how to propagate it.”
2026-03-11	Think While Watching: Online Streaming Segment-Level Memory for Multi-Turn Video Reasoning in Multimodal Large Language Models
• Existing streaming multimodal large models typically adopt a serial “perception–generation alternation” paradigm, where text decoding blocks the continuous intake of video, and as long videos progress, the model is prone to forgetting key information from earlier segments. To address this, this paper proposes a novel streaming inference framework that “thinks while watching.” • The framework divides a video into multiple segments and, during system operation, dynamically generates and maintains persistent segment-level memory notes online, which support multi-turn question answering via implicit retrieval. • Constructs a dedicated three-stage streaming chain-of-thought (CoT) dataset—covering single-turn adaptation, multi-turn interaction, and long-range capability training—and pairs it with segment-level streaming causal masks to ensure strict temporal causality.
2026-03-09	MEMO: Memory-Augmented Model Context Optimization for Robust Multi-Turn Multi-Agent LLM Games
• In long-horizon multi-agent games, early mistakes are easily amplified, and fixed prompts can lead to rigid strategies and highly variable evaluation results. To address this, this paper proposes MEMO, a self-play framework that requires no updates to model weights. • MEMO cleverly decouples and combines the mechanisms of “retention” and “exploration.” It builds a persistent memory bank that uses CRUD (create, read, update, delete) operations to extract structured strategic insights from self-play trajectories and injects them as prior knowledge for subsequent reasoning; meanwhile, it employs tournament-style prompt evolution based on TrueSkill ratings and a prioritized experience replay mechanism to efficiently explore strategies and revisit critical decision states. • In five text-based game benchmarks, MEMO demonstrates remarkable learning efficiency: with just 2,000 self-play episodes, it boosts GPT-4o-mini’s average win rate from 25.1% to 49.5%, while simultaneously causing a substantial reduction in the variance of its performance.
2026-03-12	Collaborative Multi-Agent Optimization for Personalized Memory System
• Existing personalized large-model memory systems are typically composed of multiple agents, but most prior methods optimize them independently and locally, neglecting cross-agent collaboration, which means that locally optimal performance cannot guarantee the best global question-answering performance of the overall system. • To address the optimization challenges caused by asynchronous execution of heterogeneous agents, the paper proposes the CoMAM framework, which models the execution pipelines of agents—such as fine-grained extraction, coarse-grained profiling, and memory retrieval—as a sequential Markov decision process. • To align local task improvements with global system performance, CoMAM quantifies each agent’s contribution by computing the group-level ranking consistency between its local rewards and the global system reward.
2026-03-13	Structured Distillation for Personalized Agent Memory: 11× Token Reduction with Retrieval Preservation
• It proposes extracting each conversational interaction into a structured composite object that includes core content, specific context, topic classification, and related files. This approach follows a “surviving vocabulary” principle, avoiding arbitrary rewriting of technical terms, and successfully compresses the average number of tokens per interaction from 371 to 38, achieving an 11× compression efficiency. • In experiments covering 107 retrieval configurations, it was found that the best pure distilled-text setup can retain 96% of the retrieval quality (MRR) of the original verbatim text. Furthermore, retrieval performance is highly mechanism-dependent: vector search shows virtually no noticeable degradation even under 11× compression, whereas keyword search (BM25) degrades significantly. • The agent carries only the compressed distilled text within the context as a “routing index” for efficient retrieval, while the original full conversation text is stored locally and is only brought up for display when the user needs to inspect it in detail.
2026-03-09	TA-Mem: Tool-Augmented Autonomous Memory Retrieval for LLM in Long-Term Conversational QA
• Structured episodic memory extraction: TA-Mem introduces an episodic memory extraction agent that uses one-shot prompting to adaptively chunk long contexts based on semantic topic shifts, extracting them into structured memory notes containing summaries, keywords, entities, and events. • Tool-driven autonomous memory exploration: It constructs a multi-indexed memory database supporting diverse query methods (e.g., string matching and vector similarity). A tool-augmented retrieval agent can autonomously select appropriate tools and run iterative agentic loops to explore the memory space. • Enhanced reasoning and token efficiency: By leveraging precise tool calls and a per-session memory caching mechanism to filter out redundant context, the framework significantly outperforms existing baselines on complex long-range inference tasks (such as the LoCoMo benchmark) while maintaining high token efficiency.
2026-03-09	EVOKING USER MEMORY: PERSONALIZING LLM VIA RECOLLECTION-FAMILIARITY ADAPTIVE RETRIEVAL
• It points out that current memory retrieval in personalized large models either feeds in the entire history, causing context overload, or relies only on a single similarity search, leading to shallow understanding. To address this, the research team draws inspiration from the dual-process theory of human memory (“recollection–familiarity”) and proposes an adaptive memory retrieval framework called RF-Mem. • RF-Mem measures “familiarity” by probing the average similarity scores and entropy of the retrieval. When familiarity is high and uncertainty is low, the system takes a fast “familiarity route,” directly returning the single top-K result; when familiarity is low and uncertainty is high, it instead activates a deeper “recollection route.” • When the “recollection route” is triggered, the system clusters the candidate memories and uses an α-mix strategy to update the original query by blending it with the cluster centroids.
2026-03-10	A Control-Theoretic Foundation for Agentic Systems
• It proposes a control-theoretic framework that embeds AI agent systems into a feedback control loop for analysis. Rather than treating AI memory, learning, tool invocation, interaction signals, and goal specification as isolated, fragmented functions, this framework unifies and integrates them into a single closed-loop dynamical architecture. • It creatively defines a five-level hierarchy of agents based on the extent of “decision authority” that AI holds within the control architecture. • Applying this framework to both nonlinear and linear systems, the paper points out that as the agent hierarchy increases, the system will inevitably introduce more complex dynamical mechanisms, such as time-varying adaptation, endogenous switching, decision-induced delays, and structural reconfiguration of the control pipeline.
2026-03-11	When OpenClaw Meets Hospital: Toward an Agentic Operating System for Dynamic Clinical Workflows
• Constrain the agents to an isolated environment where they are only allowed to read and write specific files and invoke a pre-vetted “medical skill library,” cutting off arbitrary code execution and network access at the operating-system level to ensure data security and compliance. • Instead of traditional vector retrieval, clinical documents are organized into a tree structure augmented with manifests. Relying on natural language understanding to read these manifests, the agents perform “progressive disclosure” navigation, thereby acquiring long-term medical record context in a precise and interpretable manner. • Instead of communicating directly, multiple agents collaborate implicitly by append-only writes to shared clinical documents and event subscriptions. This allows the agents to perform on-the-fly task orchestration and flexibly handle complex, long-tail clinical needs that traditional systems cannot address.
2026-03-10	Trajectory-Informed Memory Generation for Self-Improving Agent Systems
• To address the agents’ “amnesia” problem, it systematically and automatically extracts reusable practical experience from their historical execution traces, including clean successes, inefficient successes, failures, and error recoveries. • It proposes a complete learning feedback loop that includes trajectory intelligence extraction, decision attribution analysis, contextual learning generation, and adaptive memory retrieval. • It achieves comprehensive improvements in agent task performance on the AppWorld benchmark, with particularly pronounced gains on long-horizon tasks that require complex planning.
2026-03-09	AutoAgent: Evolving Cognition and Elastic Memory Orchestration for Adaptive Agents
• It structures the agent’s cognition into “internal cognition” and “external cognition.” This cognition, treated as an updatable state of the agent, is continuously self-corrected and evolved based on the actual outcomes of task interactions, thereby providing a more accurate and less biased knowledge foundation for decision-making. • It adopts a dynamic “select–execute–update” loop for problem solving. The action space is uniformly divided into “emic actions” (self-driven) and “etic actions” (collaborative), enabling the agent to adaptively plan its next step entirely based on the current context and real-time cognition. • To tackle context redundancy and token consumption in long-sequence reasoning, the system designs an Elastic Memory Orchestrator (EMO) that dynamically compresses historical trajectories, filters redundant information, and extracts reusable episodic memories.
2026-03-09	Multi-Agent Memory from a Computer Architecture Perspective: Visions and Challenges Ahead
• It compares the multi-agent memory system to a classic computer system, distinguishing between two basic architectural prototypes: shared memory and distributed memory. • It proposes an architecture-inspired three-layer memory hierarchy (I/O layer, cache layer, and memory layer), and identifies two critical missing protocols: an agent cache sharing protocol, and an agent memory access protocol that regulates read/write permissions and granularity. • In the future, the most pressing challenge in building multi-agent systems is ensuring memory consistency. This requires the system to properly handle read-time conflicts, the visibility and ordering of update operations when multiple agents concurrently read and write shared memory, and to establish clear versioning and conflict resolution rules in order to maintain coherence of the global context.
2026-03-04	Adaptive Memory Admission Control for LLM Agents
• The absence of memory admission control allows long-term memory stores to be rapidly polluted by low-quality, redundant, or ineffective noise features, ultimately undermining multi-turn reasoning. • Introduces Adaptive Memory Admission Control (A-MAC), a structured admission mechanism that uses lightweight estimation of five interpretable factors—such as future utility, confidence, and novelty—to rigorously intercept write operations before they occur. • Establishes admission control as a core engineering principle in memory system design, cleansing the memory store at the source and delivering simultaneous latency and performance gains on the LoCoMo benchmark.
2026-03-02	MemSifter: Offloading LLM Memory Retrieval via Outcome-Driven Proxy Reasoning
• When executing long-horizon tasks, repeatedly invoking the primary model for complex memory retrieval is prohibitively expensive, making it difficult to balance computational efficiency and accuracy. • Proposes the MemSifter framework, which leverages an outcome-driven reinforcement learning paradigm to offload retrieval-intensive computation directly to a lightweight proxy model. • Achieves low-cost decoupling of the retrieval architecture, attaining state-of-the-art retrieval hit rates with minimal inference overhead, and offering a highly scalable industrial solution for long-term memory.
2026-03-02	GAM-RAG: Gain-Adaptive Memory for Evolving Retrieval in Retrieval-Augmented Generation
• Traditional RAG systems rely on static indexing structures, which introduce substantial traversal redundancy and computational waste when handling frequently recurring queries. • Inspired by cognitive science, the paper proposes the training-free GAM-RAG framework, which uses Kalman gain rules to dynamically strengthen memory states for repeated retrieval based on query feedback. • Successfully balances index stability and adaptability, effectively avoiding unproductive retrieval while substantially reducing inference compute overhead without sacrificing accuracy.
2026-03-02	Diagnosing Retrieval vs. Utilization Bottlenecks in LLM Agent Memory
• Clearly identifies a prevailing optimization misconception in current memory systems: the true impact of the writing, retrieval, and utilization stages on overall performance must be quantitatively diagnosed. • Develops a diagnostic framework that performs rigorous cross-ablation studies of mainstream writing and retrieval strategies, isolating and disentangling the contribution of each stage. • Demonstrates that retrieval is the dominant performance bottleneck by a wide margin (with an impact weight of up to 20%), and that raw chunk-based storage remains highly effective, guiding the industry to prioritize retrieval enhancement over complex compression schemes.
2026-02-28	MemPO: Self-Memory Policy Optimization for Long-Horizon Agents
• Long-horizon agents passively depend on external RAG retrieval and lack autonomous mechanisms for assessing the value of historical information and managing what should be retained. • Proposes the Self-Memory Policy Optimization algorithm (MemPO), which introduces an improved credit assignment mechanism to grant the model the ability to proactively summarize and filter high-value memories. • By precisely removing redundant information, the method significantly reduces token consumption while comprehensively outperforming prior approaches in both F1 score and overall task performance.
2026-02-26	ParamMem: Augmenting Language Agents with Parametric Reflective Memory
• Introduces ParamMem, a parameterized memory module that addresses output repetition caused by self-reflection through diversified reflection generation. • Employs temperature-controlled sampling and cross-sample memory techniques to build the high-performance ParamAgent framework. • Significantly outperforms existing baselines on tasks such as code generation and mathematical reasoning, validating the positive impact of reflection diversity on task success rates.
2026-02-26	Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language
• Innovatively proposes a general neural memory system whose updates can be controlled via natural language instructions. • Addresses the limitation of traditional models that prevent users from intervening in memory content, enabling agents to selectively learn from heterogeneous sources. • This approach shows great application potential in scenarios such as healthcare and customer service, where memory accuracy and controllability are critically important.
2026-02-26	Exploratory Memory-Augmented LLM Agent via Hybrid On- and Off-Policy Optimization
• Proposes the EMPO² hybrid policy optimization framework, aimed at overcoming exploration bottlenecks in reinforcement learning training for LLM agents. • Combines memory-guided exploration with policy updates, ensuring strong robustness in both memory-enabled and memory-free scenarios. • Achieves a 128.6% performance leap on the ScienceWorld task and demonstrates excellent adaptability on out-of-distribution tasks.
2026-02-26	AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications
• Introduces AMA-Bench, a long-term memory evaluation framework specifically designed for autonomous agent–environment interactions rather than dialogue-only settings. • Simultaneously proposes the AMA-Agent system, which leverages causal graphs and tool-augmented retrieval to optimize memory retrieval quality. • The study reveals shortcomings of existing memory systems in continuous environment interaction tasks and provides effective pathways for improvement.
2026-02-25	Towards Autonomous Memory Agents
• Proposes the autonomous memory agent U-Mem, which shifts memory behavior from passive storage to active knowledge acquisition, verification, and curation. • Introduces a cost-aware knowledge extraction cascade, incorporating multi-level validation from self-signals to expert feedback, combined with semantically aware sampling strategies. • Successfully mitigates cold-start bias and outperforms previous memory baselines on challenging knowledge-intensive tasks such as HotpotQA.
2026-02-25	Structurally Aligned Subtask-Level Memory for Software Engineering Agents
• To address the challenges faced by software engineering agents when handling tasks with similar descriptions but differing logic, this work proposes structure-aligned, subtask-level memory. • Through fine-grained task decomposition and memory alignment, the mechanism significantly enhances agents’ long-horizon reasoning capabilities. • Its superiority in solving complex, logically entangled tasks is demonstrated across multiple software engineering benchmark evaluations.
2026-02-24	ARCHITECTING AGENTOS: FROM TOKEN-LEVEL CONTEXT TO EMERGENT SYSTEM-LEVEL INTELLIGENCE
• This paper explores a framework—AgentOS—that enables large language models (LLMs) to evolve into dynamic autonomous cognitive systems. • By redefining LLMs as a “reasoning kernel,” AgentOS introduces the concept of deep context management to address information dilution in long-context tasks and temporal drift in multi-agent collaboration found in current applications. • The paper provides a detailed exposition of AgentOS’s core components, such as the Cognitive Synchronization Pulse (CSP), semantic slicing theory, and perception alignment mechanisms, and explains how they facilitate the emergence of collective intelligence in multi-agent systems.
2026-02-24	Pancake: Hierarchical Memory System for Multi-Agent LLM Serving
• Proposes Pancake, a multi-level agent memory system designed to address multi-agent memory fragmentation in large-scale LLM services. • Unifies multi-level index caching, coordinated index management, and collaborative GPU–CPU acceleration techniques. • Compatible with mainstream frameworks such as LangChain, Pancake achieves more than a 4.29× throughput improvement under real-world workloads.
2026-02-23	Agents of Chaos
• The paper reports on red-teaming tests of autonomous language model agents, highlighting the security and privacy vulnerabilities that agents can introduce in real-world environments. • The study examines 11 cases, revealing various issues and failure modes of agents in handling sensitive information, following instructions, and managing resources. • The text discusses the complexity of responsibility attribution and the shortcomings of existing technical and legal frameworks in ensuring the safety of autonomous systems.
2026-02-23	Structured Prompt Language: Declarative Context Management for LLMs
• Introduces SPL, an SQL-inspired declarative language that treats the LLM context window as a constrained resource for efficient management. • Natively integrates RAG and persistent memory, providing an automatic query optimizer and transparent EXPLAIN-style debugging capabilities. • Experiments confirm that SPL significantly reduces prompt redundancy and lowers computational costs in tasks such as multilingual translation and logical chunking.
2026-02-23	Agentic AI as a Cybersecurity Attack Surface: Threats, Exploits, and Defenses in Runtime Supply Chains
• Systematically analyzes the security risks in the reasoning dependencies of LLM agents, categorizing them as data and tool supply chain attacks. • Introduces the concept of a “viral agent loop,” warning that agents may become carriers of self-propagating generative worms. • Advocates for the adoption of a zero-trust runtime architecture, constraining tool execution through cryptographically verified provenance and treating context as an untrusted data stream.
2026-02-22	Anatomy of Agentic Memory: Taxonomy and Empirical Analysis of Evaluation and System Limitations
• The paper analyzes agentic memory systems—particularly the state persistence capabilities of large language models (LLMs) during long-term interactions—revealing that the empirical foundations of their design and evaluation remain fragile. • The article categorizes memory structures, including lightweight semantic memory, entity-centric and personalized memory, episodic memory, and reflective memory. Through empirical analysis, it shows that current systems fall short of expectations and identifies a mismatch between evaluation metrics and semantic utility. • The paper also explores Memory-Augmented Generation (MAG) in the context of long-context processing, examining different memory architectures, operations, and management strategies, and emphasizes the need for optimized memory management and improved evaluation methodologies.
2026-02-20	From Lossy to Verified: A Provenance-Aware Tiered Memory for Agents
• TierMem is a hierarchical memory system designed to improve the efficiency and accuracy of information retrieval, addressing information compression challenges arising from long-term interaction histories. • By constructing a fast summary layer and an immutable raw log layer, the system effectively manages information storage and retrieval, accelerating query response times while reducing costs. • TierMem’s design incorporates intelligent routing optimization, error analysis, and hierarchical recall strategies to balance accuracy and efficiency.
2026-02-20	Memory-Based Advantage Shaping for LLM-Guided Reinforcement Learning
• This study proposes an innovative “memory-based advantage shaping” approach aimed at improving the sample efficiency of reinforcement learning (RL) guided by large language models (LLMs) in sparse or delayed reward environments. • By constructing a memory graph and introducing a utility function, the method effectively integrates external guidance to optimize the learning process of RL models. • Experimental results demonstrate high sample efficiency and strong performance across multiple benchmark environments.
2026-02-20	MIRA: MEMORY-INTEGRATED REINFORCEMENT LEARNING AGENT WITH LIMITED LLM GUIDANCE
• MIRA (Memory-Integrated Reinforcement Learning Agent) is a novel reinforcement learning agent that integrates guidance from large language models (LLMs) by constructing a structured memory graph, improving learning efficiency and exploration capability in sparse-reward environments. • The core of MIRA lies in combining offline and online LLM guidance, filtering high-confidence outputs, computing utility signals, and optimizing the advantage function to ensure autonomous learning while reducing reliance on real-time LLM queries. • Experimental results show that MIRA outperforms traditional methods in both sample efficiency and convergence across multiple environments, demonstrating its potential in complex tasks.
2026-02-20	REMem: Reasoning with Episodic Memory in Language Agent
• To address the lack of effective recall of interaction history in existing language agents, the episodic memory reasoning framework REMem is proposed. • Its core consists of two stages—offline indexing and online reasoning—enabling agents to mimic human-like reasoning within spatiotemporal contexts. • Experimental results show that this framework demonstrates significant advantages on specialized episodic memory benchmark tests.
2026-02-17	Mnemis: Dual-Route Retrieval on Hierarchical Graphs for Long-Term LLM Memory
• Mnemis is a novel memory framework designed to enhance the organization and retrieval capabilities of large language models (LLMs). • The framework combines two retrieval mechanisms: System 1 similarity search and System 2 global selection. By constructing a base graph and a hierarchical graph, it can handle complex queries more efficiently. • Mnemis performs exceptionally well on long-term memory benchmarks, achieving high scores such as 93.9 and 91.6, and demonstrates significant advantages in entity recognition and information retrieval.
2026-02-17	ZOMBIE AGENTS: PERSISTENT CONTROL OF SELF-EVOLVING LLM AGENTS VIA SELF-REINFORCING IN-JECTIONS
• The paper introduces the “Zombie Agent” attack model, focusing on self-evolving large language model (LLM) agents. • The model demonstrates how malicious payloads can be injected into an agent’s long-term memory through a black-box two-stage attack framework (infection and trigger), leading to persistent compromise. • The study shows that existing prompt-filtering defenses are insufficient to protect such self-evolving agents and highlights the particular vulnerability of memory architectures.
2026-02-17	ER-MIA: Black-Box Adversarial Memory Injection Attacks on Long-Term Memory-Augmented Large Language Models
• This paper explores a framework called ER-MIA, which investigates black-box adversarial memory injection attacks (AMIAs) targeting long-term memory–augmented large language models (LLMs). • The ER-MIA framework reveals attack mechanisms that inject malicious text through normal interactions, potentially leading the model to produce incorrect reasoning. • The study designs multiple automated adversarial memory generation strategies and conducts empirical evaluations, demonstrating that current long-term memory systems exhibit significant vulnerabilities when facing such attacks.
2026-02-17	Improving MLLMs in Embodied Exploration and Question Answering with Human-Inspired Memory Modeling
• Proposes a non-parametric memory framework that explicitly distinguishes episodic memory from semantic memory, enhancing the exploration capabilities of embodied agents. • Employs a program-style rule extraction mechanism to transform environmental experiences into structured semantic memory that can generalize across scenarios. • Significantly improves observation reuse and question-answering efficiency of multimodal large models in non-stationary environments.
2026-02-16	HyperRAG: Reasoning N-ary Facts over Hypergraphs for Retrieval Augmented Generation
• HyperRAG is an innovative retrieval-augmented generation (RAG) framework that replaces traditional binary knowledge graphs with n-ary hypergraphs to improve the accuracy and efficiency of complex question answering and knowledge retrieval tasks. • Through the new modules HyperRetriever and HyperMemory, the framework enables more precise multi-hop reasoning and enhanced context awareness. • Experimental results show that HyperRAG achieves superior performance across multiple benchmark datasets.
2026-02-16	PANINI: Continual Learning in Token Space via Structured Memory
• This paper provides a detailed introduction to PANINI, a non-parametric continual learning framework designed to enhance continual learning and multi-hop question answering capabilities. • By incorporating structured memory and a Generative Semantic Workspace (GSW), PANINI can efficiently process new information and support reasoning through chains of question answering. • Compared with various baseline models, PANINI demonstrates superior performance across multiple evaluation benchmarks, showing particular robustness in scenarios with missing evidence.
2026-02-15	Choosing How to Remember: Adaptive Memory Structures for LLM Agents
• Proposes the FluxMem framework, which enables LLM agents to adaptively select different memory structures based on interaction characteristics. • Introduces a three-layer memory hierarchy and employs probabilistic gating based on a Beta mixture model to enhance the robustness of memory fusion. • FluxMem demonstrates outstanding performance in handling heterogeneous interaction patterns, achieving significant performance gains across multiple long-term benchmarks.
2026-02-14	Hippocampus: An Efficient and Scalable Memory Module for Agentic AI
• Proposes Hippocampus, an efficient and scalable memory module for agentic AI that replaces dense-vector or graph-heavy retrieval with compact binary signatures and lossless token-ID streams. • Introduces a Dynamic Wavelet Matrix (DWM) to jointly compress and index semantic signatures and reconstructable content, enabling ultra-fast search directly in the compressed domain. • Experiments on LoCoMo and LongMemEval show up to 31× lower end-to-end retrieval latency and up to 14× fewer per-query tokens while maintaining accuracy.
2026-02-14	HyMem: Hybrid Memory Architecture with Dynamic Retrieval Scheduling
• Proposes a hybrid memory architecture, HyMem, which addresses the efficiency trade-offs in long conversations through dual-granularity storage (summary-level / deep-level). • Introduces a dynamic, on-demand scheduling mechanism: simple queries invoke efficient summaries, while complex queries activate deep-level modules. • Experiments show that HyMem maintains high performance while reducing computational costs by 92.6%.
2026-02-14	Neuromem: A Granular Decomposition of the Streaming Lifecycle in External Memory for LLMs
• Introduces Neuromem, a scalable benchmarking platform for evaluating the full dynamic lifecycle of external memory modules for LLMs, from ingestion and maintenance to retrieval and integration. • The study reveals that memory accuracy and cost are jointly influenced by the entire lifecycle rather than any single stage in isolation. • The analysis indicates that performance generally degrades as memory scale increases, and that time-dependent queries remain the most significant challenge for current systems.
2026-02-13	Learning to Remember: End-to-End Training of Memory Agents for Long-Context Reasoning
• This paper explores the development of the Unified Memory Agent (UMA) and its application in dynamic long-term state tracking and information retrieval. • By integrating memory operations with question answering, UMAs overcome the challenges faced by traditional large language models (LLMs) when handling long inputs. • UMA leverages reinforcement learning and novel policy optimization algorithms, demonstrating significant performance improvements in dynamic scenarios, with accuracy increasing from 61.38% to 76.46%.
2026-02-12	Learning to Forget Attention: Memory Consolidation for Adaptive Compute Reduction
• The paper finds that 88% of attention operations in LLMs retrieve predictable information, a redundancy that persists throughout standard training without being eliminated. • It proposes CRAM, an architecture that adaptively reduces compute by gradually distilling frequently accessed episodic retrievals into efficient parametric semantic memory via a consolidation-aware router. • CRAM achieves a 37.8x reduction in attention compute while its learned consolidation dynamics quantitatively match the power-law transition curves observed in human cognitive psychology.
2026-02-12	Scene-Aware Memory Discrimination: Deciding Which Personal Knowledge Stays
• It introduces the "memory discrimination" task, which acts as a filter during memory construction to identify and store only valuable personal knowledge from vast daily interactions. • It proposes the Scene-Aware Memory Discrimination (SAMD) framework, combining a Gating Unit Module (GUM) for efficient noise filtering and a Cluster Prompting Module (CPM) for adaptive memory standards. • Evaluations demonstrate that SAMD significantly enhances the efficiency and quality of memory construction, successfully recalling critical data while reducing computational costs for personalized AI agents.
2026-02-12	Recurrent Preference Memory for Efficient Long-Sequence Generative Recommendation
• It proposes the Rec2PM framework with a tripartite memory mechanism that compresses long user histories into compact Preference Memory tokens, overcoming the computational bottlenecks of long-sequence modeling. • It introduces a self-referential teacher-forcing strategy that leverages a global view of history to generate reference targets, enabling fully parallelized training for recurrent memory updates. • It demonstrates superior storage and inference efficiency while acting as a denoising Information Bottleneck to filter interaction noise, achieving higher accuracy than full-sequence models.
2026-02-12	TS-Memory: Plug-and-Play Memory for Time Series Foundation Models
• It proposes TS-Memory, a plug-and-play lightweight memory adapter that uses Parametric Memory Distillation to adapt frozen Time Series Foundation Models (TSFMs) to downstream domain shifts. • It employs a two-stage training strategy: first constructing an offline kNN teacher to generate privileged supervision signals, then distilling retrieval-induced distributional corrections into a parametric module via confidence-gated supervision. • It enables retrieval-free inference with constant-time complexity, consistently improving forecasting accuracy across benchmarks while maintaining the efficiency of the frozen backbone and avoiding catastrophic forgetting.
2026-02-11	Understand Then Memory: A Cognitive Gist-Driven RAG Framework with Global Semantic Diffusion
• CogitoRAG is a retrieval-augmented generation (RAG) framework based on cognitive memory mechanisms, designed to enhance large language models’ (LLMs) capabilities in knowledge integration and reasoning. • The framework simulates human cognitive processes and optimizes information retrieval and answer generation through the collaboration of three main modules: query decomposition, entity diffusion, and CogniRank re-ranking. • CogitoRAG achieves strong performance across multiple question-answering benchmarks, particularly outperforming traditional RAG methods on complex reasoning tasks, emphasizing the principle that “understanding is superior to memorization.”
2026-02-11	When to Memorize and When to Stop: Gated Recurrent Memory for Long-Context Reasoning
• It proposes GRU-Mem, a gated recurrent memory framework that processes long context chunk-by-chunk to overcome performance degradation and context window limits in LLMs. • It introduces two text-controlled gates, an update gate and an exit gate, to selectively update memory and enable early termination, effectively preventing memory explosion and reducing redundant computation. • Optimized via end-to-end reinforcement learning, GRU-Mem significantly outperforms existing methods in reasoning tasks and achieves up to 400% inference speed acceleration compared to vanilla recurrent memory agents.
2026-02-11	Towards Compressive and Scalable Recurrent Memory
• It proposes Elastic Memory, a novel recurrent memory architecture grounded in the HiPPO framework that encodes long-range history into fixed-size states via optimal online function approximation. • It develops a parallelized block-level update and a flexible "polynomial sampling" mechanism for efficient retrieval, enabling the reconstruction of history summaries from compressed states without extra trainable parameters. • Experiments demonstrate that it outperforms SOTA baselines on 32k+ context tasks with superior efficiency, while its decoupled design allows for injecting inductive biases at test-time without retraining.
2026-02-11	UMEM: Uniffed Memory Extraction and Management Framework for Generalizable Memory
• The paper proposes UMEM, a framework that jointly optimizes memory extraction and management to resolve the policy misalignment and poor generalization found in traditional static extraction methods. • It introduces Semantic Neighborhood Modeling and a Marginal Utility Reward via GRPO, forcing the agent to distill generalizable principles across clusters of related tasks instead of memorizing instance-specific noise. • Experiments show that UMEM significantly outperforms state-of-the-art baselines across multiple benchmarks and maintains a stable, monotonic performance growth curve during continuous self-evolution.
2026-02-10	TraceMem: Weaving Narrative Memory Schemata from User Conversational Traces
• It proposes TraceMem, a cognitively-inspired framework that weaves disjointed conversational traces into structured narrative memory schemata through a three-stage pipeline of processing and consolidation. • It mimics human memory consolidation by using topic segmentation and hierarchical clustering to transform episodic snippets into coherent, time-evolving narrative threads and structured user memory cards. • It implements an agentic search mechanism to enable human-like source attribution, achieving state-of-the-art performance in multi-hop and temporal reasoning for long-term dialogues.
2026-02-09	AMEM4Rec: Leveraging Cross-User Similarity for Memory Evolution in Agentic LLM Recommenders
• It proposes AMEM4Rec, an agentic framework that models collaborative filtering signals end-to-end by introducing an evolving memory module without relying on pre-trained CF models. • It designs a cross-user memory evolution mechanism that aggregates abstract behavior patterns into a global pool, using dual validation to link and iteratively evolve shared memory entries across users. • Extensive experiments demonstrate that AMEM4Rec significantly outperforms existing baselines, particularly showing superior performance and generalization in sparse interaction scenarios.
2026-02-09	Position: Stateless Yet Not Forgetful: Implicit Memory as a Hidden Channel in LLMs
• Introduces the concept of "Implicit Memory," demonstrating that LLMs can bypass their stateless nature by encoding state information in their own outputs to create a persistent hidden channel across independent sessions. • Implements and validates "Time Bombs," a new class of temporal backdoors that use implicit memory to accumulate hidden states over multiple interactions, activating only after a specific sequence of conditions is met. • Systematically analyzes risks such as covert communication and benchmark contamination, while outlining future research directions for detecting, evaluating, and controlling unintended persistence in LLMs.
2026-02-09	MemAdapter: Fast Alignment across Agent Memory Paradigms via Generative Subgraph Retrieval
• It proposes MemAdapter, a framework that unifies heterogeneous agent memory paradigms (explicit, parametric, and latent) using a paradigm-agnostic generative subgraph retrieval approach. • It employs a two-stage training strategy: distilling a generative retriever from a unified memory space and then efficiently adapting it to new paradigms via lightweight alignment modules and contrastive learning. • MemAdapter consistently outperforms existing baselines, completing paradigm alignment in just 13 minutes and enabling effective zero-shot fusion across different memory types.
2026-02-07	MemPot: Defending Against Memory Extraction Attack with Optimized Honeypots
• This paper proposes MemPot, the first defense framework against memory extraction attacks that proactively injects optimized honeypots (trap documents) into the agent's memory systems. • It utilizes a two-stage optimization strategy to maximize the statistical separability between attacker and user retrieval patterns, generating safe and inconspicuous trap texts through safety-constrained embedding inversion. • Based on Wald’s Sequential Probability Ratio Test (SPRT), MemPot achieves near-perfect detection accuracy with zero online inference latency while preserving the agent's core utility and performance.
2026-02-06	Agentic Unlearning: When LLM Agent Meets Machine Unlearning
• A new framework called Synchronized Backflow Unlearning (SBU) is proposed to effectively remove sensitive information from large language models (LLMs) while ensuring the integrity of shared knowledge. • By coordinating the collaboration between parameter paths and memory paths, SBU addresses the limitations of traditional privacy-preserving methods, successfully eliminating residual effects of deleted information and reducing the risk of information recontamination. • Experimental results show that this approach significantly improves privacy protection on medical question-answering benchmarks while maintaining accuracy. It not only outperforms baseline methods in efficiency but also overcomes computational resource constraints.
2026-02-05	Learning to Share: Selective Memory for Efficient Parallel Agentic Systems
• Studied how to share memory selectively across parallel agents to reduce redundancy and coordination overhead. • Proposed a selective sharing mechanism/policy to decide what to broadcast vs keep private per agent. • Evaluated on multi-agent settings, showing improved efficiency while maintaining (or improving) task performance.
2026-02-02	Live-Evo: Online Evolution of Agentic Memory from Continuous Feedback
• Formulated memory as an evolving object updated online from dense/continuous feedback signals rather than sparse endpoints. • Proposed an update/evolution loop that revises stored memories based on feedback to improve future behavior. • Demonstrated online improvement over time in agent tasks under continuous supervision.
2026-02-02	Beyond RAG for Agent Memory: Retrieval by Decoupling and Aggregation
• Revisited agent memory retrieval as a two-stage process: decouple candidate fetching from evidence aggregation. • Proposed an aggregation mechanism to combine multi-source/multi-hop evidence for downstream reasoning. • Showed gains vs vanilla RAG-style retrieval pipelines in agent memory usage and answer quality.
2026-01-30	Field-Theoretic Memory for AI Agents: Continuous Dynamics for Context Preservation
• Proposes an AI agent memory system that treats stored information as a continuous field (governed by partial differential equations) rather than discrete entries. • Drawing inspiration from field theory, memory diffuses within a semantic space, decays thermodynamically based on importance, and enables multi-agent interaction through field coupling. • The approach achieves significant improvements in multi-session and temporal reasoning performance on the LongMemEval benchmark, demonstrating the advantages of collective intelligence.
2026-01-30	Traversal-as-Policy: Log-Distilled Gated Behavior Trees as Externalized, Verifiable Policies for Safe, Robust, and Efficient Agents
• Proposes the Traversal-as-Policy paradigm, which distills execution logs into structured Gated Behavior Trees (GBT) and replaces unconstrained generation with deterministic tree traversal at runtime, achieving externalized and verifiable Agent policies. • Introduces Spine Memory to record traversal paths, effectively solving the problems of implicit strategy and post-hoc safety in long-horizon tasks. • Experiments validate that this method can simultaneously improve success rates and significantly reduce violation risks and token consumption across software engineering, Web Agent, and safety domains.
2026-01-29	E-mem: Multi-agent based Episodic Context Reconstruction for LLM Agent Memory
• Proposed the E-mem framework, shifting from traditional memory preprocessing to Episodic Context Reconstruction to prevent information loss caused by de-contextualization. • Adopted a heterogeneous Master-Assistant architecture where assistant agents maintain uncompressed context as memory nodes while the master agent handles global planning. • Introduced a routing mechanism allowing assistants to reason within locally restored original contexts, achieving SOTA performance on LoCoMo and HotpotQA while reducing token costs by over 70%.
2026-01-29	ShardMemo: Masked MoE Routing for Sharded Agentic LLM Memory
• Proposed ShardMemo, a tiered memory architecture: Tier A (working state), Tier B (sharded evidence), and Tier C (versioned skill library). • Enforced a "scope-before-routing" strategy in Tier B and modeled shard selection as a Masked MoE routing problem under fixed budgets, using cost-aware gating. • Improved F1 by +6.87 on LoCoMo and HotpotQA compared to cosine similarity routing, while reducing retrieval work and latency by 20.5%.
2026-01-28	MemCtrl: Using MLLMs as Active Memory Controllers on Embodied Agents
• Proposed MemCtrl, a framework using MLLMs as active memory controllers to filter redundant observations online for embodied agents. • Introduced a trainable memory head ($\mu$) acting as a gate to dynamically determine whether to retain, update, or discard observations during exploration. • Trained via offline supervision and online RL, MemCtrl improved task completion rates by ~16% for small MLLMs on EmbodiedBench, with >20% gains on specific instruction subsets.
2026-01-28	AMA: Adaptive Memory via Multi-Agent Collaboration
• Proposed AMA (Adaptive Memory via Multi-Agent Collaboration), leveraging Constructor, Retriever, Judge, and Refresher agents to manage multi-granularity memory. • Utilized a hierarchical memory design (Raw Text, Fact Knowledge, Episode), where the Retriever dynamically routes queries and the Judge detects conflicts. • The Refresher maintains long-term consistency via logic-driven updates. AMA significantly outperformed baselines on LoCoMo and LongMemEval while reducing token consumption by 80%.
2026-01-27	GLOVE: Global Verifier for LLM Memory-Environment Realignment
• Proposed the Global Verifier (GLOVE) framework to address memory-environment misalignment caused by dynamic environmental drifts. • Established "relative truth" via active probing to detect cognitive dissonance by comparing retrieved memories with fresh observations, realigning memory without ground truth. • Significantly improved agent adaptability and success rates in web navigation, discrete planning, and continuous control tasks under explicit and implicit environment drifts.
2026-01-26	FadeMem: Biologically-Inspired Forgetting for Efficient Agent Memory
• Proposed FadeMem, a memory architecture inspired by the Ebbinghaus forgetting curve, introducing active forgetting mechanisms to prevent information overload. • Implemented a dual-layer hierarchy with adaptive exponential decay governed by semantic relevance, access frequency, and temporal patterns. • Combined with LLM-guided conflict resolution, FadeMem achieved superior multi-hop reasoning on Multi-Session Chat and LoCoMo while reducing storage by 45%.
2026-01-26	MemWeaver: Weaving Hybrid Memories for Traceable Long-Horizon Agentic Reasoning
• MemWeaver proposes a tri-layer memory framework (Graph, Experience, Passage) that consolidates long-term interactions into temporally grounded structures, ensuring evidence traceability. • It employs a dual-channel retrieval strategy to "weave" structured relational facts with original textual evidence, supporting complex multi-hop and temporal reasoning tasks. • Experiments on the LoCoMo benchmark show it improves reasoning accuracy while actively reducing input context length by over 95% compared to long-context baselines.
2026-01-24	Clustering-driven Memory Compression for On-device Large Language Models
• Proposed a clustering-based memory compression strategy designed for on-device personalization under limited context windows. • Groups similar memories and merges them within clusters (instead of simple concatenation), reducing redundancy while preserving semantic coherence. • Demonstrated significant reduction in token usage and improved personalized generation quality compared to naive concatenation baselines.
2026-01-13	Chain-of-Memory: Lightweight Memory Construction with Dynamic Evolution for LLM Agents
• Proposed CoM (Chain-of-Memory), advocating a shift from expensive structured construction to lightweight construction with dynamic utilization. • Introduced Dynamic Memory Chain Evolution to organize retrieved fragments into coherent inference paths with adaptive truncation to prune noise. • Achieved 7.5%–10.4% accuracy gains on LoCoMo and LongMemEval while reducing token consumption to ~2.7% compared to complex memory structures.
2026-01-15	TeleMem: Building Long-Term and Multimodal Memory for Agentic AI
• TeleMem introduces a unified long-term and multimodal memory framework that extracts narrative-grounded information to maintain coherent user profiles without schema-driven hallucinations. • It employs a structured writing pipeline for batching, retrieval, and consolidation, significantly improving storage and token efficiency, and incorporates a multimodal memory module with ReAct-style reasoning for video understanding. • Experimental results on the ZH-4O benchmark show TeleMem outperforming the state-of-the-art Mem0 baseline by 19% in accuracy while reducing token usage by 43% and speeding up operations by 2.1×.
2026-01-15	Grounding Agent Memory in Contextual Intent
• Proposes STITCH, an agentic memory system that indexes trajectory steps using "Contextual Intent"—comprising thematic scope, event type, and key entity types—to disambiguate recurring information in long-horizon tasks. • Introduces a retrieval mechanism that filters and prioritizes memory snippets based on structural intent compatibility rather than just semantic similarity, effectively suppressing context-incompatible history. • Presents CAME-Bench, a multi-domain benchmark designed to evaluate context-aware retrieval in realistic, goal-oriented trajectories, where STITCH achieves state-of-the-art performance.
2026-01-14	PersonalAlign: Hierarchical Implicit Intent Alignment for Personalized GUI Agent with Long-Term User-Centric Records
• Introduces PersonalAlign, a new task requiring GUI agents to align with implicit user intents—specifically resolving vague instructions and anticipating routines—by leveraging long-term user records. • Presents AndroidIntent, a benchmark constructed from 20k long-term records, featuring hierarchically annotated user preferences and routines to evaluate personalization capabilities. • Proposes HIM-Agent (Hierarchical Intent Memory Agent), which utilizes a streaming aggregation module and hierarchical filters (Execution-based and State-based) to continuously update and organize user memory for improved reactive and proactive performance.
2026-01-13	AtomMem: Learnable Dynamic Agentic Memory with Atomic Memory Operation
• Introduces AtomMem, a dynamic memory framework that reframes agentic memory management as a learnable sequential decision-making problem rather than a static, hand-crafted workflow. • Deconstructs memory processes into atomic CRUD (Create, Read, Update, Delete) operations and employs reinforcement learning (GRPO) to learn a task-aligned policy for autonomously orchestrating these operations. • Experimental results on long-context benchmarks (HotpotQA, 2WikiMultihopQA, Musique) demonstrate that AtomMem consistently outperforms static memory baselines by dynamically tailoring memory strategies to specific task demands.
2026-01-13	Fine-Mem: Fine-Grained Feedback Alignment for Long-Horizon Memory Management
• Fine-Mem is a unified reinforcement learning framework designed to optimize long-horizon memory management for LLM agents by aligning fine-grained feedback with memory operations. • It addresses reward sparsity through Chunk-level Step Reward (CSR), which provides immediate supervision via constructed QA tasks, and solves credit assignment issues with Evidence-Anchored Reward Attribution (EARA) by linking global rewards to specific memory operations. • Experimental results demonstrate that Fine-Mem consistently outperforms strong baselines on benchmarks like Memalpha and MemoryAgentBench, showing superior adaptability and generalization across different models.
2026-01-12	Active Context Compression: Autonomous Memory Management in LLM Agents
• Proposed an autonomous strategy to compress/retain context so agents can operate under tight context budgets. • Treated memory management as an active decision problem: what to keep, summarize, discard, or externalize. • Demonstrated improved long-horizon performance vs passive truncation or naive summarization baselines.
2026-01-12	MemoBrain: Executive Memory as an Agentic Brain for Reasoning
• MemoBrain introduces an "executive memory" paradigm for tool-augmented agents, functioning as a co-pilot to construct dependency-aware memory and actively manage context under bounded budgets. • The framework employs specific memory operations—Trajectory Folding and Selective Flush—to organize reasoning progress, retaining a high-salience structural backbone while discarding transient execution artifacts. • Experiments on benchmarks like GAIA, WebWalker, and BrowseComp-Plus demonstrate that MemoBrain consistently outperforms strong baselines by enabling coherent, goal-directed reasoning over long horizons.
2026-01-12	Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents
• TSM is a memory framework that models semantic time for point-wise memory and supports the construction and utilization of durative memory. • It builds a semantic timeline to organize episodic interactions and consolidates them into time-aware durative memories (topics and personas) to capture long-term user states. • During memory utilization, TSM incorporates the query’s temporal intent to retrieve temporally appropriate durative memories, significantly improving performance on benchmarks like LONGMEMEVAL and LOCOMO.
2026-01-10	Bi-Mem: Bidirectional Construction of Hierarchical Memory for Personalized LLMs via Inductive-Reflective Agents
• Bi-Mem is an agentic framework that constructs hierarchical memory (fact, scene, persona) bidirectionally using an inductive agent for bottom-up aggregation and a reflective agent for top-down calibration to mitigate noise and hallucination. • It employs an associative retrieval mechanism that leverages spreading activation to connect memory units across granularities, enabling coherent recall of both contextual scenes and specific facts. • Empirical evaluations on the LoCoMo benchmark demonstrate that Bi-Mem significantly outperforms leading memory baselines in long-term personalized conversational tasks.
2026-01-10	HiMem: Hierarchical Long-Term Memory for LLM Long-Horizon Agents
• HiMem is a hierarchical long-term memory framework designed for long-horizon dialogues, integrating fine-grained "Episode Memory" (via topic-aware segmentation) with abstract "Note Memory" (via knowledge extraction) to bridge concrete events and stable knowledge. • It employs a conflict-aware "Memory Reconsolidation" mechanism that uses retrieval feedback to revise and supplement stored knowledge, enabling continual self-evolution and correction of memory over time. • Evaluations on long-horizon benchmarks demonstrate that HiMem outperforms baselines in accuracy, consistency, and reasoning, validating the effectiveness of its hierarchical organization and dynamic updating strategies.
2026-01-10	Structured Episodic Event Memory
• SEEM introduces a dual-layer memory framework combining a Graph Memory Layer for static facts and an Episodic Memory Layer for narrative progression, both anchored by provenance pointers to raw interaction passages. • The system employs a "Reverse Provenance Expansion" (RPE) mechanism to reconstruct coherent narrative contexts from fragmented evidence during retrieval, addressing the "scattered retrieval" problem in long-term interactions. • Experiments on benchmarks like LoCoMo and LongMemEval show SEEM significantly outperforms competitive memory-augmented baselines (like HippoRAG 2) in narrative coherence and logical consistency.
2026-01-09	MemBuilder: Reinforcing LLMs for Long-Term Memory Construction via Attributed Dense Rewards
• MemBuilder is a reinforcement learning framework that trains LLMs to actively construct and manage a multi-dimensional memory system (Core, Episodic, Semantic, and Procedural) rather than relying on static prompting. • It introduces "Attributed Dense Rewards Policy Optimization" (ADRPO) to solve reward sparsity and credit assignment issues by using synthetic session-level QA for immediate feedback and gradient weighting based on memory component contribution. • Experimental results show that a lightweight 4B model trained with MemBuilder outperforms state-of-the-art closed-source models (including Claude 4.5 Sonnet) on long-term dialogue benchmarks like LoCoMo and LongMemEval.
2026-01-08	Beyond Static Summarization: Proactive Memory Extraction for LLM Agents
• ProMem Framework: Addresses the limitations of "one-off" static summarization by proposing a proactive memory extraction framework inspired by Recurrent Processing Theory (RPT). • Recurrent Feedback Loop: Introduces a self-questioning mechanism where the agent actively probes dialogue history to verify facts and recover missing details, ensuring memory completeness and accuracy. • Performance: Outperforms state-of-the-art baselines (e.g., Mem0) on HaluMem and LongMemEval benchmarks, demonstrating high robustness in token compression and cost-effectiveness with Small Language Models.
2026-01-08	Memory Matters More: Event-Centric Memory as a Logic Map for Agent Searching and Reasoning
• Proposed CompassMem, an event-centric memory framework inspired by Event Segmentation Theory, organizing memory as an Event Graph with explicit logical relations (causal, temporal). • Transforms memory from passive storage into a Logic Map, enabling agents to actively navigate structured dependencies via a Planner-Explorer-Responder mechanism. • Features active multi-path memory search that dynamically expands or skips nodes based on subgoal satisfaction, avoiding redundant retrieval. • Demonstrates superior performance on LoCoMo and NarrativeQA benchmarks, significantly improving multi-hop and temporal reasoning compared to baselines like HippoRAG and Mem0.
2026-01-08	Inside Out: Evolving User-Centric Core Memory Trees for Long-Term Personalized Dialogue Systems
• PersonaTree Framework: Introduces a globally maintained **PersonaTree** grounded in the Biopsychosocial model as a dynamic user profile. By constraining the trunk with a schema and iteratively updating branches, it enables controllable memory growth and compression. • MemListener Training: Trains a lightweight MemListener model via **Reinforcement Learning with process-based rewards** to generate structured, executable memory operations (ADD, UPDATE, DELETE), achieving performance comparable to large reasoning models. • Adaptive Response Generation: Implements a dual-mode strategy that utilizes PersonaTree directly for low-latency responses or triggers an agentic recall mode guided by the tree for complex queries, outperforming baselines in consistency and noise suppression.
2026-01-07	Membox: Weaving Topic Continuity into Long-Range Memory for LLM Agents
• Membox Architecture: Addresses the "fragmentation-compensation" flaw in existing systems by proposing a hierarchical architecture centered on **Topic Continuity** to preserve temporal and causal flow. • Topic Loom & Trace Weaver: Introduces a *Topic Loom* to group continuous dialogue into cohesive "memory boxes" and a *Trace Weaver* to link these boxes into long-range event timelines across discontinuities. • Performance: Achieves up to 68% F1 improvement on temporal reasoning tasks in the LoCoMo benchmark compared to baselines like Mem0, while using significantly fewer context tokens.
2026-01-06	HiMeS: Hippocampus-inspired Memory System for Personalized AI Assistants
• HiMeS is a memory framework for AI assistants that emulates the hippocampus–neocortex interaction by integrating short-term dialogue compression with long-term user profile storage. • It utilizes a short-term memory extractor trained via reinforcement learning to proactively pre-retrieve knowledge, and a partitioned long-term memory network to re-rank results based on historical user interactions. • Evaluations on real-world industrial datasets demonstrate that HiMeS significantly outperforms traditional RAG baselines in personalized question-answering tasks.
2026-01-06	SYNAPSE: Empowering LLM Agents with Episodic-Semantic Memory via Spreading Activation
• SYNAPSE is a brain-inspired memory architecture that replaces static vector retrieval with a unified episodic–semantic graph, addressing the “context isolation” issue in traditional RAG systems, where semantically distant yet causally related memories cannot be effectively associated. • It introduces cognitive dynamics such as spreading activation, lateral inhibition, and temporal decay to dynamically propagate relevance and filter noise within the graph, rather than relying solely on precomputed links or vector similarity. • SYNAPSE achieves state-of-the-art performance on the LoCoMo benchmark, significantly improving multi-hop reasoning capabilities and robustness to adversarial queries through an uncertainty-aware gating mechanism.
2026-01-06	CODEMEM: AST-Guided Adaptive Memory for Repository-Level Iterative Code Generation
• Proposed CODEMEM, a memory management system tailored for repository-level iterative code generation. • Introduces Code Context Memory: Uses AST-guided selection to dynamically update and merge repository context, keeping it relevant while discarding noise. • Introduces Code Session Memory: Uses AST-based change analysis to detect conflicts and forgetting, organizing history into code-centric units (diffs) rather than just text. • Achieves SOTA on CodeIF-Bench and CoderEval, improving instruction following by ~12% and reducing interaction rounds by 2–3.
2026-01-06	Implicit Graph, Explicit Retrieval: Towards Efficient and Interpretable Long-horizon Memory for Large Language Models
• LatentGraphMem Framework: Proposes a memory framework combining implicit graph memory for stability and efficiency with explicit subgraph retrieval for interpretability, storing graph structures in latent space. • Three-Stage Training Strategy: Involves training a graph builder (global representation), a subgraph retriever (budgeted edge selection), and joint fine-tuning (coordination optimization) for effective end-to-end QA. • Performance: Consistently outperforms explicit-graph and latent-memory baselines on long-horizon benchmarks like HotpotQA, NarrativeQA, and WikiHop across multiple model scales, achieving up to 63.34% average accuracy.
2026-01-06	MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents
• MAGMA Architecture: Proposes a multi-graph agentic memory architecture that explicitly models memory items across orthogonal semantic, temporal, causal, and entity graphs, overcoming the limitations of monolithic memory stores. • Adaptive Topological Retrieval: Introduces an intent-aware Adaptive Traversal Policy that dynamically routes retrieval through relevant relational views, decoupling memory representation from retrieval logic for transparent reasoning. • Performance: Consistently outperforms state-of-the-art agentic memory systems (e.g., Nemori, A-MEM) on long-horizon benchmarks like LoCoMo and LongMemEval, while reducing retrieval latency and token consumption.
2026-01-06	TiMem: Temporal-Hierarchical Memory Consolidation for Long-Horizon Conversational Agents
• TiMem Framework: Introduces a temporal-hierarchical memory framework using a Temporal Memory Tree (TMT) to progressively consolidate raw dialog into abstract persona representations, emphasizing temporal continuity. • Core Mechanisms: Features semantic-guided consolidation (fine-tuning free) and complexity-aware memory recall (Recall Planner + Gating) to balance precision and efficiency across query types. • Performance: Achieves SOTA accuracy on LoCoMo (75.30%) and LongMemEval-S (76.88%) benchmarks, while significantly reducing recalled context length (-52.20%) on LoCoMo.
2026-01-06	MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory
• MemRL Framework: Proposes a non-parametric reinforcement learning framework that enables frozen LLM agents to self-evolve by optimizing episodic memory, avoiding the costs and forgetting issues of fine-tuning. • Intent-Experience-Utility Triplet: Introduces a Two-Phase Retrieval mechanism (semantic recall + value-aware selection) and a runtime utility update rule, using Q-values to distinguish high-utility strategies from noise. • Performance: Significantly outperforms MemP and RAG on benchmarks like HLE, BigCodeBench, and ALFWorld, demonstrating that agents can continuously improve via runtime trial-and-error without weight updates.
2026-01-05	SimpleMem: Efficient Lifelong Memory for LLM Agents
• Introduces SimpleMem, an efficient memory framework tailored for lifelong LLM agents based on semantic lossless compression. • The system operates via a three-stage pipeline: Semantic Structured Compression to filter low-entropy noise, Recursive Memory Consolidation to synthesize abstract representations, and Adaptive Query-Aware Retrieval to minimize token usage. • Experiments on the LoCoMo benchmark demonstrate a 26.4% improvement in F1 score and up to 30× reduction in inference token consumption compared to full-context models, significantly outperforming baselines like Mem0.
2026-01-05	Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents
• AgeMem Framework: Proposes a unified framework that integrates both Long-Term (LTM) and Short-Term (STM) memory management directly into the agent's policy via tool-based actions (e.g., Add, Update, Filter). • Three-Stage Progressive RL: Introduces a step-wise GRPO algorithm and a three-stage training strategy (LTM construction, STM control, integrated reasoning) to address sparse rewards and enable end-to-end optimization. • Performance: Outperforms strong baselines like LangMem and Mem0 across five long-horizon benchmarks (e.g., ALFWorld, HotpotQA), achieving higher task success rates, better memory quality, and more efficient context usage.
2025-12-31	Nested Learning: The Illusion of Deep Learning Architecture
• Full arXiv version including all appendices — not the previously released trimmed version. • Presents a Nested Learning paradigm that unifies a large portion of optimizer and TTT-layer modules. • Architectural innovation: HOPE — composed of modified Titans attention and self-modified FFNs. By controlling the chunksize of self-modification of FFN parameter, FFN layers operating at different frequencies implicitly retain memories at different hierarchical levels during runtime. • Empirical results are modest.
2025-12-25	Beyond Heuristics: A Decision-Theoretic Framework for Agent Memory Management
• Provides TeleAI background and introduces a decision-theoretic memory framework (DAM) that formulates the timing and content of memory read/write as an optimal decision problem, with relevance to RL-style formulations. • Contains minimal or no experimental validation.
2025-12-21	MemEvolve: Meta-Evolution of Agent Memory Systems
• An OPPO-affiliated paper proposes a two-layer framework that, in RL settings, separates learning to extract memories (level-1) from learning the memory-extraction method itself (level-2). • Experiments using Flash-Searcher and GPT-5-Mini achieve SOTA on benchmarks including GAIA.
2025-12-20	MemR³: Memory Retrieval via Reflective Reasoning for LLM Agents
• MemR³ closed-loop retrieval controller: designed for long-term conversational memory, it can dynamically choose among three actions—retrieve, reflect, and respond. • Evidence–gap state tracker: the system maintains a global (evidence, gap) state that explicitly tracks "what is known" and "what is missing," making the process interpretable. • Experiments show that on the LoCoMo benchmark, MemR³ significantly improves answer quality across different underlying memory systems (e.g., RAG, Zep).
2025-12-18	Learning Hierarchical Procedural Memory for LLM Agents through Bayesian Selection and Contrastive Refinement
• A Bayesian procedural memory (experience) framework: MACLA. • Overall still a rule-based algorithm; operations include extraction, retrieval/storage, and refinement (Bayesian posterior calibration). • On unseen tasks in ALFWorld, performance (90.3%) exceeded that on seen tasks (87.2%), achieving +3.1% positive generalization.
2025-12-14	HINDSIGHT IS 20/20: BUILDING AGENT MEMORY THAT RETAINS, RECALLS, AND REFLECTS
• HINDSIGHT is a unified memory architecture that treats memory as a structured, first-class substrate for reasoning, organizing information into four logical networks: world facts, agent experiences, synthesized entity summaries, and evolving beliefs. • The system introduces TEMPR (Temporal Entity Memory Priming Retrieval) for building temporal entity graphs and CARA (Coherent Adaptive Reasoning Agents) for preference-conditioned reasoning, enabling agents to epistemically distinguish evidence from inference. • Experimental results on LongMemEval and LoCoMo benchmarks demonstrate that HINDSIGHT significantly outperforms existing memory systems and full-context frontier models in multi-session consistency and open-domain question answering.
2025-12-11	Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution
• ReMe (paper version): an Alibaba-affiliated framework for enhancing LLM procedural memory (experience), including the ReMe algorithm and the reme.library dataset. • Core idea: maintain an experience pool with operations—Acquisition, Reuse, and Refinement. • Experiments on BFCL-V3 and AppWorld show dynamic experience pools > static pools > baseline, with scaling studies for both model and judge models.
2025-12-10	LightSearcher: Efficient DeepSearch via Experiential Memory
• LightSearcher is an efficient reinforcement learning (RL)-based search architecture grounded in experiential memory. During large language model (LLM)-driven reasoning, it autonomously optimizes agent tool invocation without relying on external data by transforming implicit reasoning trajectories into explicit experiential guidance through contrastive experiential memory. • Evaluated on four multi-hop question answering benchmarks—Natural Questions (NQ), HotpotQA, MuSiQue, and 2WikiMultihopQA—LightSearcher achieves accuracy comparable to the state-of-the-art DeepSearch baseline while significantly reducing both tool invocation latency and model response time. • The method reduces tool invocations by 39.6%, shortens reasoning time by 48.6%, and decreases token consumption by 21.2%, substantially improving tool-use efficiency without compromising task performance.
2025-12-3	MemVerse: Multimodal Memory for Lifelong Learning Agents
• A lifelong learning memory framework for multimodal agents. • Retrieval-based long-term memory + parameterized fast memory + periodic distillation. • Multimodal handling: unified conversion into textual descriptions. • Experiments show improvements over baselines on ScienceQA (text) and MSR-VTT (video); LoCoMo (text) results remain unpublished (in appendix).
2025-11-12	ComoRAG: A Cognitive-Inspired Memory-Organized RAG for Stateful Long Narrative Reasoning
• Introduces ComoRAG, a retrieval-augmented generation framework inspired by the human Prefrontal Cortex, designed to achieve stateful reasoning in long narrative contexts. • The framework employs a dynamic memory workspace and a metacognitive regulation loop (including Self-Probe, Mem-Fuse, and Mem-Update) to iteratively fuse fragmented evidence into coherent context. • Experimental results demonstrate that ComoRAG consistently outperforms strong baselines on challenging benchmarks like NarrativeQA and ∞BENCH, particularly excelling in complex narrative queries requiring global understanding.
2025-11-04	MemSearcher Training LLMs to Reason, Search and Manage Memory via End-to-End Reinforcement Learning
• MemSearcher is a large language model (LLMs) agent trained through end-to-end Reinforcement Learning (RL), aiming to enhance the efficiency of knowledge acquisition tasks. • MemSearcher optimizes memory management by adopting a new framework called multi-context Group Relative Strategy Optimization (Multi-Context GRPO), which enables the model to self-evolve in multiple conversations. • Compared with traditional ReAct search agents, MemSearcher offers significant performance improvements while maintaining low token consumption, especially on smaller models.
2025-10-15	D-SMART: Enhancing LLM Dialogue Consistency via Dynamic Structured Memory And Reasoning Tree
• Proposes D-SMART, a model-agnostic framework designed to maintain logical and factual consistency in multi-turn dialogues by coupling a Dynamic Structured Memory (DSM) with a Reasoning Tree (RT). • DSM incrementally builds an OWL-compliant knowledge graph from conversation history to prevent context decay, while RT guides the LLM through explicit, traceable multi-step reasoning over this graph. • Comprehensive experiments on MT-Bench-101 demonstrate that D-SMART significantly outperforms state-of-the-art baselines, improving consistency scores by over 48% and exhibiting strong stability in extended dialogues.
2025-10-14	Memory as Action Autonomous Context Curation for Long-Horizon Agentic Tasks
• Memory-as-action (MemAct) addresses the issue of working Memory management for large language models (LLMS) in long-duration tasks. • MemAct transforms memory management into a learnable intrinsic capability, enabling agents to dynamically manage memories while performing tasks, and introduces the Dynamic Context Policy Optimization (DCPO) algorithm to handle the trajectory breakage problem caused by memory editing. • MemAct performs exceptionally well in multi-objective question answering tasks, demonstrating higher accuracy and robustness than traditional models.
2025-10-12	MemGen Weaving Generative Latent Memory for Self-Evolving Agents
• MemGen is a dynamic generative memory framework designed to enhance the reasoning and decision-making capabilities of agents based on large language models (LLMS). • MemGen simulates human cognitive patterns by interweaving memory with the reasoning process. • This framework consists of two parts: memory triggers and memory weavers, which can dynamically determine when to invoke potential memories and integrate them into the reasoning process.
2025-10-10	How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior
• The paper investigates memory management in large language model (LLM) agents and its impact on long-term performance. • It identifies issues such as error propagation and misaligned experience replay, highlighting the importance of high-quality memory. • By comparing multiple memory insertion and deletion strategies, the study finds that selective insertion performs better for long-term learning, while historical deletion is particularly effective at reducing low-quality memory records.
2025-10-09	Enabling Personalized Long-term Interactions in LLM-based Agents through Persistent Memory and User Profiles
• Introduces a framework for adaptive, user-centered AI agents that combines persistent memory, dynamic coordination, and evolving user profiles to enable personalized long-term interactions. • The approach integrates established agentic AI patterns—such as Multi-Agent Collaboration and Multi-Source Retrieval—with mechanisms like self-validation and implicit user profiling to tailor responses to individual needs. • Evaluations on three public datasets and a pilot user study demonstrate improvements in retrieval accuracy, response correctness, and perceived personalization compared to standard RAG baselines.
2025-10-08	ToolMem: Enhancing Multimodal Agents with Learnable Tool Capability Memory
• TOOLMEM, a memory-augmented agent that learns from past tool use. It stores summarized, retrievable “what this tool is good/bad at” knowledge and injects the relevant memories into context to better predict tool quality and choose the right tool for new tasks. • TOOLMEM maintains structured capability entries per tool. From each experience, it retrieves similar memories and updates them via a RAG-style merge/refinement, keeping a compact, evolving capability memory. At inference time, it retrieves the most relevant capability memories to guide scoring and tool selection. • They evaluate on text generation tools and text-to-image tools, comparing against no-memory and few-shot baselines. TOOLMEM improves quality prediction and makes better tool choices overall.
2025-10-07	CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension
• The paper introduces CAM, a Constructivist Agentic Memory system inspired by Jean Piaget’s theory, designed to enhance Large Language Models (LLMs) in long-form document comprehension. • CAM features structured schemata, flexible assimilation, and dynamic accommodation, utilizing an incremental overlapping clustering algorithm for efficient memory development and an adaptive Prune-and-Grow strategy for retrieval. • Experimental results across diverse benchmarks show that CAM achieves dual advantages in both performance and efficiency compared to existing structured and unstructured memory approaches.
2025-09-30	MEM-α: LEARNING MEMORY CONSTRUCTION VIA REINFORCEMENT LEARNING
• Proposes Mem-α, a reinforcement learning framework that trains agents to effectively manage complex memory systems (comprising core, episodic, and semantic components) through interaction and feedback. • Unlike approaches relying on pre-defined instructions, Mem-α treats memory construction as a sequential decision-making problem, optimizing directly for downstream question-answering accuracy. • Experimental results show that Mem-α significantly outperforms existing baselines and demonstrates remarkable generalization, effectively handling contexts exceeding 400k tokens despite being trained on 30k token sequences.
2025-09-29	ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
• ReasoningBank, a test-time learning framework that distills an agent’s own successful and failed trajectories into reusable reasoning memories. For new tasks, the agent retrieves relevant memories to guide decision-making and then writes new experience back into the bank, forming a self-improving loop without requiring ground-truth feedback. • Each memory is stored as a compact structured item and retrieved via embedding similarity (top-k) to augment the agent’s prompt. After task execution, an LLM-as-a-judge provides proxy success/failure signals: successful trajectories yield transferable strategies, while failed ones yield pitfalls and corrective rules. In addition, MaTTS expands test-time computation through parallel trajectory sampling and serial self-reflection, both of which generate stronger memory signals. • Experiments are conducted on WebArena and Mind2Web and SWE-Bench-Verified, comparing against No Memory and prior memory-based baselines. Performance is evaluated using success rate, efficiency (steps), and task-specific metrics. Results show consistent improvements across different backbone models
2025-09-29	Pretraining with hierarchical memories: separating long-tail and common knowledge
• Proposes a "pretraining-with-memories" architecture that decouples reasoning capabilities (anchor model) from long-tail world knowledge (hierarchical memory bank). • The system dynamically retrieves and attaches context-dependent parameter blocks from a massive memory bank to a small anchor model during inference, enabling efficient scaling. • Experiments demonstrate that a 160M model augmented with memories matches the performance of a standard model with over twice the parameters, specifically excelling at long-tail knowledge tasks.
2025-09-27	Look Back to Reason Forward: Revisitable Memory for Long-Context LLM Agents
• Addressed long-context reasoning where relevant evidence is dispersed across very long inputs. • Proposed a “revisitable” memory design that allows the agent to look back and selectively retrieve from the broader history. • Evaluated on long-context QA settings to show improved evidence recovery and reasoning accuracy.
2025-09-26	Conflict-Aware Soft Prompting for Retrieval-Augmented Generation
• The "Conflict-Aware Retrieval Enhancement Generation" (CARE) model aims to address the context-memory conflict problem that occurs in Retrieval Enhancement Generation (RAG). • CARE optimizes the performance of large language models (LLMs) by introducing context evaluators, especially in dealing with conflicts between external and internal knowledge. • This method significantly enhances the accuracy and reliability of the model in multiple tasks through techniques such as conflict-aware fine-tuning, soft prompts, and adversarial soft prompts.
2025-09-26	PRIME Planning and Retrieval-Integrated Memory for Enhanced Reasoning
• PRIME is a multi-agent inference framework. PRIME provides intuitive answers to simple questions through fast-response agents. • PRIME performs complex reasoning through multiple specific agents, such as memory, planning, search and reading agents. • PRIME still needs to improve its belief correction mechanism and optimize the interaction among agents.
2025-09-25	SGMEM: Sentence Graph Memory for Long-Term Conversational Agents
• SGMem is a hierarchical memory management framework designed to address memory fragmentation in long-term conversational agents by organizing dialogue into sentence-level graphs. • It explicitly models associations across turns, rounds, and sessions, and uses a multi-hop retrieval mechanism to integrate raw dialogue history with generated memory such as summaries, facts, and insights. • Extensive experiments on LongMemEval and LoCoMo benchmarks demonstrate that SGMem consistently improves retrieval coherence and outperforms strong baselines in question answering accuracy.
2025-09-22	PRINCIPLES: Synthetic Strategy Memory for Proactive Dialogue Agents
• PRINCIPLES builds a retrievable memory of dialogue strategy principles from offline self-play. At inference time, the model retrieves and applies these principles to guide strategy selection and response generation, without any additional training. • In the offline stage, the agent conducts multi-turn self-play with a user simulator and uses rewards to identify success or failure. Successful cases directly yield principles, while failed cases trigger strategy revision and rollback until success; principles are then extracted by contrasting failure-to-success trajectories in a structured form. In the online stage, relevant principles are retrieved using contextual embeddings, reinterpreted to fit the current dialogue, and then used to guide planning and response generation. • Experiments on emotional support and persuasion tasks show that PRINCIPLES improves success rates and strategy prediction performance while increasing strategy diversity. Ablation studies confirm the importance of retrieval and reinterpretation, and human evaluations indicate overall preference for the proposed method.
2025-09-16	WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research
• Introduces WebWeaver, a dual-agent framework comprising a Planner and a Writer designed to tackle open-ended deep research (OEDR) by emulating human research processes. • The Planner uses a dynamic cycle to interleave evidence acquisition with outline optimization, building a memory bank of evidence; the Writer performs hierarchical, citation-grounded retrieval to compose the report section by section. • WebWeaver achieves state-of-the-art performance on benchmarks like DeepResearch Bench by effectively managing long contexts and mitigating hallucinations through targeted memory retrieval.
2025-09-15	MOOM: Maintenance, Organization and Optimization of Memory in Ultra-Long Role-Playing Dialogues
• MOOM is a dual-branch memory extraction framework designed for ultra-long role-playing dialogues, modeling "plot development" and "character portrayal" as core storytelling elements. • It incorporates a novel forgetting mechanism based on "competition-inhibition" theory to effectively control memory capacity and prevent uncontrolled expansion. • The authors introduce ZH-4O, a large-scale Chinese role-playing dataset with average 600-turn dialogues and manual memory annotations, demonstrating MOOM's superior performance over state-of-the-art methods.
2025-09-13	Pre-Storage Reasoning for Episodic Memory: Shifting Inference Burden to Memory for Personalized Dialogue
• PREMem (Pre-storage Reasoning for Episodic Memory) is a novel approach that shifts complex reasoning processes from response generation to the memory construction phase. • It extracts fine-grained memory fragments (categorized into factual, experiential, and subjective information) and establishes explicit cross-session relationships based on cognitive schema theory, capturing evolution patterns like extensions and transformations. • Experiments on LongMemEval and LoCoMo benchmarks show significant performance improvements, enabling smaller models to achieve results comparable to larger baselines while reducing inference computational demands.
2025-09-11	OpenUnlearning:Accelerating LLM unlearning via unified benchmarking of methods and metrics
• Introduces the “OpenUnlearning” framework, designed to advance research on unlearning in large language models (LLMs). • OpenUnlearning integrates a wide range of unlearning algorithms and evaluation methods, streamlining the research workflow for studying forgetting. • Through targeted and task-specific evaluations, OpenUnlearning ensures the credibility and robustness of unlearning assessment standards.
2025-08-27	Memory-R1: Enhancing Large Language Model Agents to Manage and Utilize Memories via Reinforcement Learning
• Memory-R1 is an RL-driven framework that empowers LLMs to actively manage and utilize external memory via two specialized agents: a Memory Manager and an Answer Agent. • The Memory Manager learns structured operations (ADD, UPDATE, DELETE) to maintain memory, while the Answer Agent filters retrieved memories for accurate reasoning. • With only 152 training samples, it outperforms strong baselines on LoCoMo, MSC, and LongMemEval, demonstrating high data efficiency and generalization.
2025-08-26	MemoryVLA Perceptual-Cognitive Memory in Vision-Language-Action Models for Robotic Manipulation
• MemoryVLA is a newly developed robot operation framework, aiming to enhance the performance of robots in complex tasks by integrating visual, language, and perception-cognitive mechanisms. • This framework adopts an architecture similar to the human dual memory system, enhancing the robot's ability to handle long-sequence tasks. • MemoryVLA introduces perception-cognitive memory banks (PCMB), which can effectively integrate historical information with current decisions, thereby enhancing the success rate of robots in responding to complex scenarios.
2025-08-22	Memento: Fine-tuning LLM Agents without Fine-tuning LLMs
• Proposed a paradigm to improve agent behavior via “agent-side” learning while keeping the base LLM frozen. • Focused on adapting the agent’s components (e.g., memory/reasoning/routing) rather than model weights. • Reported performance gains across agent tasks without conventional LLM fine-tuning.
2025-08-21	Multiple Memory Systems for Enhancing the Long-term Memory of Agent
• Proposes a Multiple Memory System (MMS) inspired by cognitive psychology to address the issue of low-quality memory content in existing agent memory modules. • The system processes short-term memory into diverse fragments—keywords, cognitive perspectives, episodic memory, and semantic memory—to construct specialized retrieval and contextual memory units. • Experimental results on the LoCoMo dataset demonstrate that MMS significantly outperforms methods like MemoryBank and A-MEM, particularly in multi-hop reasoning and open-domain tasks.
2025-08-18	Semantic Anchoring in Agentic Memory: Leveraging Linguistic Structures for Persistent Conversational Context
• Semantic Anchoring is a hybrid agentic memory architecture designed to enhance the long-term context retention of LLMs by enriching vector-based storage with explicit linguistic cues such as syntactic dependencies, discourse relations, and coreference links. • The proposed framework employs a multi-stage pipeline involving dependency parsing, coreference resolution, and discourse tagging to construct a hybrid index, allowing retrieval systems to access memories based on both semantic similarity and structural linguistic roles. • Experimental results on adapted long-term dialogue datasets (MultiWOZ-Long and DialogRE-L) demonstrate that Semantic Anchoring outperforms strong RAG baselines, improving factual recall and discourse coherence by up to 18% while maintaining higher user satisfaction.
2025-08-13	Memp: Exploring Agent Procedural Memory
• Memp treats procedural memory as an external, learnable store of past successful experiences so an LLM agent can reuse effective “how-to” routines on new tasks, improving success and reducing wasted steps. • Memp follows a Build–Retrieve–Update loop: it builds memory items from trajectories/scripts, retrieves the most relevant items via semantic keys and vector similarity, and updates memory online by adding, filtering, and correcting items so the memory becomes more reliable over time. • On TravelPlanner and ALFWorld, Memp outperforms a ReAct baseline with higher success/score and fewer steps; vector-based retrieval beats random selection; online updates yield further gains, and learned memories can transfer from stronger to weaker models with diminishing returns as retrieval size grows.
2025-08-12	Context as Memory Scene-Consistent Interactive Long Video Generation with Memory Retrieval
• "Context-as-memory" significantly enhances the scene consistency and Memory capacity of long video generation by leveraging historical Context as memory. • The paper studies key designs such as context learning mechanisms, camera control, and memory retrieval strategies, and points out the balance between computational efficiency and generation quality. • Based on the long video generation architecture of the diffusion model, the current technological progress, challenges and future directions are expounded.
2025-08-12	Intrinsic Memory Agents: Heterogeneous Multi-Agent LLM Systems through Structured Contextual Memory
• Introduces Intrinsic Memory Agents, a multi-agent framework designed to address context limitations and role inconsistency using structured, agent-specific memories. • The method employs role-aligned memory templates and intrinsic updates derived directly from agent outputs, preserving heterogeneous perspectives and domain expertise without external summarization. • Evaluations on the PDDL benchmark demonstrate a 38.6% performance improvement with high token efficiency, while case studies show enhanced quality in complex planning tasks.
2025-08-06	RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory
• RCR-Router is a role-aware context routing framework designed for multi-agent LLM systems to address the limitations of static and full-context routing, such as excessive token consumption and redundant memory exposure. • The framework dynamically selects semantically relevant memory subsets for each agent based on their specific role and the current task stage, enforcing a strict token budget and utilizing an iterative feedback mechanism to refine context. • Experiments on multi-hop QA benchmarks (HotPotQA, MuSiQue, 2WikiMultihop) demonstrate that RCR-Router reduces token usage by 25–47% while maintaining or improving answer quality compared to baseline strategies.
2025-08-03	MLP Memory: A Retriever-Pretrained Memory for Large Language Models
• Introduces MLP Memory, a lightweight parametric module that learns to internalize retrieval patterns without requiring explicit document access during inference, effectively bridging the gap between RAG and parametric fine-tuning. • By pretraining an MLP to imitate a kNN retriever’s behavior on the entire pretraining dataset, the model compresses large datastores into a differentiable memory component that integrates with Transformer decoders via probability interpolation. • Experimental results show that MLP Memory achieves superior scaling behavior, improves QA performance by 12.3% relative to baselines, reduces hallucinations by up to 10 points, and offers 2.5× faster inference than RAG.
2025-07-29	SynapticRAG:Enhancing temporal memory retrieval in large language models through synaptic mechanisms
• The paper proposes MemTool, a short-term memory framework for managing dynamic tool sets across multi-turn conversations. It offers three architectures: Autonomous Agent, Workflow, and Hybrid, balancing autonomy and control. • In Autonomous Mode, the agent autonomously adds/removes tools using Search_Tools and Remove_Tools. Workflow Mode follows a fixed pipeline: pruning tools, then searching and adding new ones. Hybrid Mode separates tool removal and adding, offering a balance of stability and flexibility. • Using ScaleMCP’s 5,000 MCP servers and a 100-turn dialogue, the authors evaluate 13 LLMs with a 128-tool limit. Autonomous Mode achieves 90-94% tool removal efficiency, while Workflow and Hybrid perform consistently well, with Autonomous and Hybrid excelling in task completion.
2025-07-27	SynapticRAG:Enhancing temporal memory retrieval in large language models through synaptic mechanisms
• SynapticRAG is a novel memory retrieval framework for large language models (LLMs), designed to enhance memory retrieval in cross-session conversations. • By combining temporal association triggers with biologically inspired synaptic propagation mechanisms, SynapticRAG significantly improves the identification of relevant conversational history. • Experimental results show that the framework achieves improvements of up to 14.66% across multiple performance metrics and demonstrates clear advantages in dynamic memory management.
2025-07-17	MEM1 Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents
• MEM1 is an innovative end-to-end reinforcement learning framework designed to enhance the efficiency of large language models (LLMs) in long-term multi-round interactions. • MEM1 effectively solves the problem of memory dilation in context processing of traditional models by constructing a compact shared internal state. • The experimental results show that MEM1 significantly improves performance in multiple tasks while reducing memory usage, demonstrating its wide applicability and optimization potential in dynamic environments.
2025-07-03	MemAgent Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent
• MemAgent is a long text processing method that uses reinforcement learning (RL) to dynamically update memory, aiming to address the performance degradation and high computational complexity issues of large language models (LLMS) when dealing with long texts. • The model can maintain a linear time complexity while handling inputs of infinite length by treating memory as a latent variable and introducing stream processing and multi-session strategies. • The experimental results show that MemAgent performs outstandingly with high accuracy in ultra-long text tasks, especially having obvious advantages in complex multi-hop reasoning tasks.
2025-06-19	From RAG to Memory: Non-Parametric Continual Learning for Large Language Models
• The paper proposes HippoRAG 2, a “long-term memory–inspired” structured RAG system. It builds a knowledge graph from text and retrieves evidence via graph-based propagation (PPR) to support multi-hop association, while improving basic factual recall that earlier structured RAGs often hurt. • Offline, an LLM performs OpenIE to extract triples and form a KG, and adds passages as nodes linked to phrase nodes to fuse concept-level structure with context-rich passages. Online, it first retrieves top-k triples with embeddings, then uses an LLM for triple filtering to remove irrelevant triples; the remaining nodes seed a PPR run to rank the most relevant passages for the generator. • It evaluates factual QA, multi-hop reasoning, and narrative understanding, reporting Recall@5 for retrieval and F1 for QA. Compared with BM25, dense retrievers, and multiple structured-RAG baselines, HippoRAG 2 generally improves retrieval and end-to-end QA, and ablations plus “growing-corpus” settings support the contribution of its components.
2025-06-09	G-Memory: Tracing Hierarchical Memory for Multi-Agent Systems
• Introduces G-Memory, a hierarchical memory system designed to address the lack of self-evolution capabilities in Large Language Model (LLM)-based Multi-Agent Systems (MAS). • Implements a three-tier graph architecture—Insight Graph, Query Graph, and Interaction Graph—to manage lengthy interaction histories by abstracting generalizable insights and condensing specific collaborative trajectories. • Experimental results across embodied action and knowledge QA benchmarks demonstrate that G-Memory significantly enhances agent team performance, improving success rates by up to 20.89% without modifying the original frameworks.
2025-05-30	M+：Extending MemoryLLM with scalable Long-Term Memory
• M+ is a memory-augmented model designed to improve long-term information retention in large language models (LLMs). • Built upon MemoryLLM, M+ integrates long-term memory mechanisms with a jointly trained retriever, substantially enhancing the model’s ability to handle knowledge spanning over 20,000 tokens while maintaining comparable GPU memory overhead. • M+ achieves strong performance across multiple benchmarks, outperforming MemoryLLM and other competitive baselines, and demonstrates efficient information compression and end-to-end training, exhibiting mechanisms that closely resemble human memory.
2025-05-26	MemGuide: Intent-Driven Memory Selection for Goal-Oriented Multi-Session LLM Agents
• MemGuide is a two-stage framework designed to enhance multi-session task-oriented dialogue (TOD) by incorporating task intent and slot-level guidance into memory selection. • It employs Intent-Aligned Retrieval to match current context with stored intent descriptions and Missing-Slot Guided Filtering to prioritize memory units that fill information gaps using a Chain-of-Thought reasoner. • The authors also introduce MS-TOD, a multi-session TOD benchmark. Evaluations show MemGuide significantly improves task success rates and reduces dialogue turns compared to strong baselines.
2025-05-23	Towards General Continuous Memory for Vision-Language Models
• CoMEM addresses the token overload and performance degradation issues in traditional Retrieval-Augmented Generation (RAG) for Vision-Language Models (VLMs) by introducing a general continuous memory mechanism. • The method innovatively utilizes the VLM itself as a memory encoder combined with a lightweight Q-Former, efficiently compressing diverse multimodal and multilingual knowledge into a compact set of continuous embeddings. • CoMEM is data- and parameter-efficient (requiring only 1.2% trainable parameters) and plug-and-play, significantly enhancing performance on complex multimodal reasoning tasks while keeping the inference model frozen.
2025-05-21	Pre-training Limited Memory Language Models with Internal and External Knowledge
• Introduces Limited Memory Language Models (LMLM), a new class of models that externalizes factual knowledge to an external database during pre-training rather than encoding it in parameters. • The approach uses a modified pre-training objective that masks retrieved factual values from the loss, encouraging the model to perform targeted lookups for facts instead of memorizing them. • Experiments demonstrate that LMLMs match the factual precision of significantly larger models while enabling instant, verifiable knowledge updates and effective machine unlearning through simple database operations.
2025-05-11	In Prospect and Retrospect: Reflective Memory Management for Long-term Personalized Dialogue Agents
• Proposes Reflective Memory Management (RMM), a novel framework for long-term dialogue agents that addresses the limitations of rigid memory granularity and fixed retrieval mechanisms. • Integrates Prospective Reflection to dynamically organize dialogue history into topic-based memories, and Retrospective Reflection to iteratively refine retrieval using online reinforcement learning guided by LLM attribution signals. • Experimental results on MSC and LongMemEval benchmarks demonstrate that RMM significantly outperforms strong baselines, achieving over 10% improvement in accuracy and enhancing response personalization.
2025-04-22	MemoRAG Boosting Long Context Processing with Global Memory-Enhanced Retrieval Augmentation
• MemoRAG aims to enhance the ability of large language models (LLMs) in handling long contexts by improving the information retrieval and generation process through a global memory-enhanced retrieval mechanism. • This framework adopts a lightweight global memory module and a complex generation system, which can effectively manage long contexts and generate useful clues to assist in answer generation. • This model is applicable to a variety of tasks, including long document question answering and summarization, demonstrating its potential in handling complex long text scenarios.
2025-04-20	SAGE: Self-evolving Agents with Reflective and Memory-augmented Abilities
• SAGE addresses the long-term memory and multitasking challenges of large language models (LLMs) in dynamic environments through three collaborative agents. SAGE integrates a reflection mechanism and memory optimization based on the Ebbinghaus forgetting curve, helping the model effectively filter and store important information while reducing cognitive load. • SAGE continuously optimizes the Assistant’s decisions through an iterative feedback mechanism and reflection functionality. Its MemorySyntax component simulates human memory decay, dynamically managing both short-term and long-term memory to ensure the retention of critical information while reducing unnecessary memory burden. • Experiments show that SAGE significantly improves model performance on AgentBench and long-text tasks (e.g., HotpotQA), with performance improvements up to 2.26x in multi-hop question answering and code generation tasks, and it effectively resolves 73.6% of ambiguous references in dialog tasks, demonstrating its potential in real-world applications.
2025-04-10	Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory
• The paper proposes Dynamic Cheatsheet (DC)—a continuously updated “sticky-note” external memory for black-box LLMs at inference time. It distills verified solving patterns and reuses them across problems, enabling test-time learning without training. • DC consists of a Generator (Gen) and a Memory Curator (Cur): Gen produces an answer using the current memory, and Cur then refines/filters/compresses the information. A retrieval-based variant selects the most relevant past examples and solutions by similarity to assist generation, while preventing memory bloat. • DC is evaluated across multiple tasks and models (e.g., AIME, GPQA-Diamond, Game of 24, MMLU-Pro; GPT-4o, Claude 3.5 Sonnet, etc.) using metrics such as Soft Match and Functionally Correct. Results show substantial gains; for example, the jump on Game of 24 is largely driven by reusable Python solver code being repeatedly “written and reused” in memory.
2025-03-27	MemInsight: Autonomous Memory Augmentation for LLM Agents
• Proposed an autonomous memory augmentation pipeline to improve how agents store and later retrieve historical interactions. • Emphasized filtering/structuring memory to keep salient information and reduce irrelevant recall. • Validated on multiple agent scenarios (e.g., QA / recommendation / summarization) to show improved contextual responses.
2025-03-07	Memory-augmented Query Reconstruction for LLM-based Knowledge Graph Reasoning
• MemQ is proposed to decouple reasoning (natural-language steps) from query generation/execution (SPARQL): the LLM produces a clear reasoning plan, while the actual query is obtained via memory retrieval + rule-based reconstruction, reducing errors and hallucinations caused by entangling tool calls with reasoning. • During training, gold SPARQL queries are decomposed into query fragments, and a natural-language explanation is generated for each fragment to build a query memory bank of (explanation → fragment) pairs. At inference time, the LLM generates step-by-step plans; reconstruction uses semantic retrieval (Sentence-BERT) to fetch an adaptive Top-N set of fragments, then assembles them with rules and fills entity slots to produce the final executable query. • Experiments on WebQSP and CWQ use Hits@1 and F1, where MemQ achieves the best overall performance. Additional analyses with structural consistency / edge hit rate show the reconstructed queries are closer to the gold graphs, and ablation studies confirm that the main gains come from the memory bank + decoupling design.
2025-02-25	Towards effective evaluation and comparisons for LLM unlearning methods
• The paper examines machine unlearning in large language models (LLMs) and the importance of its evaluation, with a particular focus on removing undesirable or unnecessary data memories. • It introduces Unlearning with Calibration (UWC) to calibrate model performance and strengthen the evaluation of different unlearning methods. • The study emphasizes the importance of selecting appropriate evaluation metrics and recommends Extraction Strength (ES) as a primary evaluation tool to ensure accuracy and robustness in assessment.
2025-02-09	LM2 Large Memory Models
• LM2 aims to overcome the limitations of traditional Transformers in multi-step reasoning, relational argumentation, and long context processing. • The LM2 integrates an auxiliary memory module, which utilizes cross-attention mechanisms and gating technology to enhance information storage and update capabilities. • In multiple benchmark tests, LM2 has demonstrated significantly superior performance, particularly excelling in long context reasoning tasks, effectively enhancing the ability to process and remember complex information.
2025-02-03	TReMu: Towards Neuro-Symbolic Temporal Reasoning for LLM-Agents with Memory in Multi-Session Dialogues
• Introduced an evaluation task/benchmark targeting temporal reasoning over noisy, multi-session dialogues with memory. • Proposed a neuro-symbolic framework (TReMu) to improve temporal reasoning using memory-aware representations. • Constructed multi-choice QA style evaluations (augmented from existing dialogue sources) and reported improved performance.
2025-01-23	ON MEMORY CONSTRUCTION AND RETRIEVAL FOR PERSONALIZED CONVERSATIONAL AGENTS
• Introduces SECOM, a memory management method that constructs memory banks at the segment level to address limitations of turn-level and session-level approaches in long-term conversations. • SECOM partitions conversations into topically coherent segments and employs prompt compression (LLMLingua-2) as a denoising mechanism to enhance retrieval accuracy. • Experimental results demonstrate that SECOM significantly outperforms existing baselines on long-term conversation benchmarks like LOCOMO and Long-MT-Bench+.
2025-01-19	Alternate Preference Optimization for Unlearning Factual Knowledge in Large Language Models
• Proposes Alternate Preference Optimization (AltPO), a method designed to effectively address the challenges of machine unlearning in large language models (LLMs). • AltPO enhances unlearning by combining negative feedback from the forget set with positive feedback from the same domain to generate multiple alternative responses, thereby improving forgetting capability while preserving overall model performance. • Experimental results demonstrate that AltPO outperforms existing methods in terms of both unlearning quality and model utility.
2024-12-31	Titans Learning to Memorize at Test Time
• "Titans" aims to enhance the model's memory capacity when dealing with long sequences and complex contexts. • The Titans architecture combines short-term memory and long-term memory modules, overcomes the limitations of traditional recursive models and attention mechanisms, and is capable of handling larger context Windows. • The experimental results show that Titans exhibit superior performance and flexibility, especially in handling long dependency relationships and diverse tasks.
2024-12-17	On the Structural Memory of LLM Agents
• The paper investigates how the structure and retrieval methods of memory modules in large language models (LLMs) affect model performance, with a focus on different memory architectures and their roles in information extraction and generation. • The study finds that hybrid memory structures outperform others in complex tasks, demonstrating greater robustness in noisy environments. • Through hyperparameter sensitivity analysis, the research identifies memory retrieval strategies that are best suited to different task settings.
2024-12-01	SELF-UPDATABLE LARGE LANGUAGE MODELS BY INTEGRATING CONTEXT INTO MODEL PARAMETERS
• Proposes SELF-PARAM, a method to integrate contexts directly into LLM parameters without requiring extra storage modules, ensuring both high efficacy and long-term retention. • Employs a training objective that minimizes the KL divergence between an original model (with context access) and a target model (without context), utilizing diverse generated QA pairs. • Experiments demonstrate that SELF-PARAM significantly outperforms existing continual learning and RAG methods in question-answering and conversational recommendation tasks, achieving near-optimal performance with zero storage complexity.
2024-10-10	Assessing episodic memory in LLMs with sequence order recall tasks
• This study introduces the Sequence Order Recall Task (SORT), designed to evaluate the episodic memory capabilities of large language models (LLMs). • The task highlights the importance of episodic memory—linking memories with relevant context such as time and location—particularly in everyday cognitive tasks. • Preliminary results indicate that LLMs exhibit strong memory performance when contextual information is provided, but their performance degrades significantly when relying solely on training data.
2024-08-19	ELDER: Enhancing Lifelong Model Editing with Mixture-of-LoRA
• ELDER proposes a novel lifelong model editing method using a Mixture-of-LoRA structure to establish continuous associations between data and adapters, enhancing robustness against rephrased inputs. • The framework integrates a router network with a guided loss function to align LoRA allocations with edit knowledge and utilizes a deferral mechanism to preserve the model's general capabilities. • Extensive experiments on GPT-2 XL and LLaMA2-7B demonstrate that ELDER outperforms existing baselines in reliability, generalization, and scalability while maintaining performance on downstream tasks.
2024-08-16	MemLong: Memory-Augmented Retrieval for Long Text Modeling
• Proposed MemLong, a memory-augmented retrieval approach for long-context language modeling by fetching historical chunks externally. • Combined a retrieval/memory module with a partially trainable decoder-only LM, plus controllable retrieval attention over retrieved chunks. • Evaluated on long-context language modeling benchmarks to show improved generation quality and longer effective context.
2024-08-11	Towards Safer Large Language Models through Machine Unlearning
• This paper introduces the Selective Knowledge Unlearning (SKU) framework, aimed at improving the safety of large language models (LLMs). • The SKU framework consists of two main stages: harmful knowledge acquisition, followed by knowledge negation, which focuses on removing undesirable knowledge without degrading model utility under benign prompts. • SKU successfully reduces harmful outputs while preserving response quality, and demonstrates a strong balance between unlearning effectiveness and model utility across multiple LLM architectures, such as OPT and LLaMA2.
2024-08-06	RULER: What’s the Real Context Size of Your Long-Context Language Models?
• RULER is designed for the comprehensive evaluation of long-context language models (LMs) across a wide range of tasks. • It extends the traditional Needle-in-a-Haystack (NIAH) test by incorporating tasks such as multi-hop tracking and aggregation, enabling a more thorough assessment of models’ understanding under long-context settings. • RULER demonstrates strong performance in multi-hop reasoning and information retrieval tasks.
2024-07-22	A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts
• ReadAgent is a reading comprehension system designed to improve the performance of large language models (LLMs) when processing long-form text. • Through three steps—episodic pagination, memory summarization, and interactive lookup—ReadAgent significantly extends the effective context length by up to 20×. • ReadAgent outperforms traditional approaches on long-document reading comprehension benchmarks such as QuALITY, NarrativeQA, and QMSum.
2024-06-30	Towards Efficient and Effective Unlearning of Large Language Models for Recommendation
• Introduces E2URec, a recommendation data unlearning method specifically designed for LLM-based recommender systems (LLMRec). • E2URec significantly improves unlearning efficiency while preserving recommendation performance by updating only Low-Rank Adaptation (LoRA) parameters. • Experimental results show that E2URec outperforms existing baseline methods on real-world datasets.
2024-05-30	Knowledge Graph Tuning: Real-time Large Language Model Personalization based on Human Feedback
• Proposes Knowledge Graph Tuning (KGT), a novel approach that personalizes large language models (LLMs) by optimizing external knowledge graphs based on user feedback, without modifying model parameters. • KGT extracts personalized factual knowledge triples from user interactions and employs a heuristic optimization algorithm, avoiding the high computational costs and low interpretability of back-propagation methods. • Experiments with models like Llama2 and Llama3 demonstrate that KGT significantly enhances personalization performance while reducing latency by up to 84% and GPU memory costs by up to 77%.
2024-05-26	MemoryLLM:Towards self-Update Large Language Models
• MEMORYLLM is a self-updating large language model designed to effectively integrate new knowledge while maintaining long-term information retention. • By embedding a fixed-size memory pool in the latent space of the transformer, MEMORYLLM achieves a seamless combination of model self-updating and knowledge preservation. • Key design features include memory tokens that store compressed knowledge, an intelligent self-updating mechanism, and comprehensive evaluations of knowledge integration, retention capability, and robustness.
2024-05-23	HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models
• HippoRAG is a novel retrieval framework inspired by the hippocampal indexing theory of human long-term memory, designed to enable deeper and more efficient knowledge integration for LLMs. • By orchestrating LLMs, knowledge graphs, and Personalized PageRank (PPR) to mimic the neocortex and hippocampus, it enables effective single-step multi-hop retrieval. • The method outperforms state-of-the-art retrieval-augmented generation (RAG) methods on multi-hop QA tasks by up to 20% and is significantly faster and cheaper than iterative retrieval approaches.
2024-05-23	WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models
• Identifies an "impossible triangle" in lifelong model editing—reliability, generalization, and locality cannot be simultaneously achieved—attributing this to the gap between long-term and working memory mechanisms. • Proposes WISE, a dual parametric memory framework that utilizes a side memory for edits and a router to bridge it with the pretrained main memory, employing knowledge sharding and merging to handle continuous updates. • Extensive experiments show that WISE outperforms existing methods in question answering, hallucination correction, and out-of-distribution generalization settings across multiple LLM architectures.
2024-04-26	Enhancing Large Language Model with Self-Controlled Memory Framework
• Proposes the Self-Controlled Memory (SCM) framework to unleash infinite-length input capacity for Large Language Models (LLMs) without requiring modification or fine-tuning. • The framework comprises an LLM-based agent, a memory stream for storing historical information, and a memory controller that dynamically manages "Activation Memory" (long-term) and "Flash Memory" (short-term). • The authors also contribute a dataset covering long-term dialogues, book summarization, and meeting summarization, demonstrating that SCM achieves superior retrieval recall and response generation compared to baselines.
2024-04-24	From Local to Global: A GraphRAG Approach to Query-Focused Summarization
• Introduces GraphRAG, a graph-based retrieval-augmented generation approach designed to address the limitations of conventional vector RAG in answering global questions about an entire text corpus. • The method constructs an entity knowledge graph from source documents, partitions it into hierarchical communities using the Leiden algorithm, and pre-generates summaries to facilitate global sensemaking. • By utilizing a map-reduce mechanism over community summaries, GraphRAG significantly outperforms baseline RAG systems in comprehensiveness and diversity for large-scale datasets.
2024-04-15	Memory Sharing for Large Language Model based Agents
• Introduces the Memory Sharing (MS) framework, which enables multiple LLM-based agents to share Prompt-Answer (PA) pairs as memories in a dynamic, real-time pool. • The framework employs a dual-purpose mechanism where newly generated high-quality memories are used to enhance In-Context Learning for agents and simultaneously train the retriever to improve future retrieval relevance. • Experimental results across domains like Literary Creation and Logic Problem-solving demonstrate that the MS framework effectively evolves individual intelligence into collective intelligence, significantly improving performance on open-ended questions without explicit fine-tuning.
2024-04-13	LLM In-Context Recall is Prompt Dependen
• Investigates the information recall capabilities of large language models (LLMs), with particular emphasis on their dependence on prompt content and formatting. • Using the Needle-in-a-Haystack (NIAH) evaluation, the study finds that recall performance is strongly influenced by training data bias, as well as the content and structure of prompts. • The results show that architectural improvements, training strategy adjustments, and fine-tuning can all effectively enhance recall performance.
2024-04-07	Online Adaptation of Language Models with a Memory of Amortized Contexts
• Introduces Memory of Amortized Contexts (MAC), an efficient online adaptation framework for large language models (LLMs) designed to address catastrophic forgetting and high computational costs in keeping models up-to-date. • MAC utilizes a meta-learned amortization network to compress new documents into compact parameter-efficient finetuning (PEFT) modulations stored in a memory bank, using an aggregation network to retrieve and combine relevant knowledge for specific queries. • Experimental results on StreamingQA and SQuAD-Seq demonstrate that MAC significantly outperforms existing online finetuning methods in both adaptation performance and knowledge retention, while offering superior time and memory efficiency.
2024-03-24	MemoryBank: Enhancing Large Language Models with Long-Term Memory
• MemoryBank is a long-term memory mechanism designed for large language models (LLMs) to address memory limitations in continuous interactions. • By enabling models to effectively recall, update, and adapt user memories, MemoryBank enhances contextual understanding and user experience. • Experimental results and analyses demonstrate MemoryBank’s effectiveness in improving emotional support and personalized interactions.
2024-02-16	Large Language Model Unlearning
• Explores methods for implementing “forgetting” or “unlearning” in large language models (LLMs) to eliminate undesired or misaligned behaviors. • By applying a gradient ascent (GA) strategy and introducing a random-output loss, the study demonstrates that unlearning can effectively prevent models from generating harmful responses. • Experimental results show that the GA and GA + Mismatch approaches perform particularly well in reducing content leakage rates.
2024-02-06	Compressed context memory for online language model interaction
• Proposes a compressed contextual memory approach to improve the memory efficiency and computational performance of online language models when handling extended contexts. • By leveraging conditional LoRA integration and parallel computation, the method significantly reduces memory requirements and enables support for effectively unlimited context lengths, surpassing traditional sliding-window strategies. • Experimental results demonstrate that, across applications such as multi-task learning and dialogue generation, the approach reduces memory usage by up to 5× while effectively preserving generation quality and accuracy.
2023-12-10	Unlearn What You Want to Forget: Efficient Unlearning for LLMs
• Introduces the Efficient Unlearning (EUL) framework, designed to address the challenges of handling user privacy data in large language models (LLMs). • As LLMs are widely deployed, models may inadvertently memorize sensitive information during pretraining, raising significant privacy concerns. • EUL enables the effective removal of specific sensitive data from LLMs without full retraining, while preserving overall predictive performance.
2023-11-30	JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models
• JARVIS-1 is an open-world multi-task agent for Minecraft that generates plans and executes tasks using a multimodal language model (MLM). It can perceive visual information and human instructions, and, by combining multimodal memory, it leverages past experiences to improve future task performance. • JARVIS-1 integrates MLM and multimodal memory to generate action plans using visual observations and instructions, which are executed by goal-conditioned controllers. It features a self-improvement mechanism, where it autonomously generates tasks through self-instruction, explores the environment, and accumulates experiences to enhance decision-making abilities. • JARVIS-1 excels in over 200 Minecraft tasks, especially in long-term tasks (such as obtaining a diamond pickaxe), outperforming current state-of-the-art models by five times in success rate. As the game progresses, its performance improves through continuous learning and experience accumulation.
2023-11-15	Think-in-Memory: Recalling and Post-thinking Enable LLMs with Long-Term Memory
• Introduces a novel memory mechanism, Think-in-Memory (TiM), designed to enhance the performance of large language models (LLMs) in long-term human–AI interactions. • TiM incorporates an efficient retrieval mechanism based on locality-sensitive hashing, enabling effective memory storage and management over extended interactions. • Experimental results show that TiM significantly improves response accuracy and coherence in multi-turn dialogues.
2023-10-16	Character-LLM: A Trainable Agent for Role-Playing
• Introduces Character-LLM, a trainable agent framework that teaches LLMs to act as specific characters (e.g., Beethoven) by learning from reconstructed experiences rather than relying solely on prompts. • Proposes an Experience Upload process involving profile collection, scene extraction, and experience completion to generate high-quality, character-specific training data. • Implements Protective Experiences to mitigate hallucinations, enabling agents to effectively "forget" or refuse knowledge inconsistent with their character's era or identity.
2023-09-22	Augmenting Language Models with Long-Term Memory
• Introduces a new framework, LONGMEM, designed to enhance the ability of large language models (LLMs) to process long-form text. • LONGMEM employs a decoupled network architecture that combines a frozen LLM memory encoder with an adaptive residual side network, enabling efficient caching and updating of long-term contextual information. • By incorporating specialized memory-augmentation layers, a token-based memory retrieval module, and a joint attention mechanism, LONGMEM improves memory retrieval and context utilization, and demonstrates effectiveness across a variety of tasks.
2023-08-16	MemoChat: Tuning LLMs to Use Memos for Consistent Long-Range Open-Domain Conversation
• Proposes MemoChat, an instruction tuning pipeline designed to enable Large Language Models (LLMs) to employ self-composed memos for maintaining consistency in long-range open-domain conversations. • The approach utilizes a "memorization-retrieval-response" cycle, teaching LLMs to restructure dialogue history into memos and retrieve relevant evidence for answering current queries. • Experiments show that MemoChat outperforms strong baselines on a newly curated, expert-annotated consistency benchmark (MT-Bench+), verifying the efficacy of the memo-equipped inner thinking process.
2023-05-23	RET-LLM: Towards a General Read-Write Memory for Large Language Models
• RET-LLM is a framework that equips large language models (LLMs) with a dedicated read-write memory unit, enabling them to explicitly extract, store, and recall knowledge from text. • Inspired by Davidsonian semantics, the system extracts knowledge in the form of triplets (concept, relationship, concept) and uses a controller to manage interactions between the LLM and the memory module using a text-based API. • The memory unit is designed to be scalable, updatable, and interpretable, effectively handling temporal-based question answering tasks where static models often fail.
2023-05-22	RECURRENTGPT: Interactive Generation of (Arbitrarily) Long Text
• Introduces RECURRENTGPT, a language-based simulacrum of the LSTM recurrence mechanism built upon LLMs to generate arbitrarily long texts without forgetting. • Utilizes a dual-memory system: a short-term memory updated in the prompt and a long-term memory stored on hard drives retrieved via semantic search. • Enables interpretable and interactive text generation ("AI as Contents"), allowing human users to observe and edit natural language memories and plans during the generation process.
2023-05-08	Prompted LLMs as Chatbot Modules for Long Open-domain Conversation
• Proposes MPC (Modular Prompted Chatbot), a novel approach using pre-trained LLMs as individual modules (clarifier, memory processor, utterance generator, summarizer) to create high-quality conversational agents without fine-tuning. • Utilizes techniques like few-shot prompting, chain-of-thought (CoT), and external memory (using DPR) to achieve long-term consistency and flexibility in open-domain dialogue. • Human evaluation results demonstrate that MPC is on par with or superior to fine-tuned models like Blenderbot3 in terms of sensibleness, consistency, and engagingness, particularly in maintaining long-term persona consistency.

Datasets & Benchmark

Date	Paper & Summary	Tags	Links
2026-11-04	Twists and Turns: A Graph-Based Benchmark and Framework for Long-Term Memory in Exploration Tasks
• Introduces LExMem for interconnected memory in long-term exploration. • Graph structures encode branching, revisions and reconsideration in 224 dialogues and 2,018 QA instances; LExGraph adds incremental nodes and neighbor expansion. • The graph framework outperforms evaluated alternatives and transfers to external benchmarks, within a structurally generated evaluation setting.
2026-11-04	Robocasa-Mem: Benchmarking Environment and Process Memory in Long-Horizon Mobile Manipulation
• Introduces RoboCasa-Mem for environment- and process-memory-dependent mobile manipulation. • Nine tasks derive later constraints from earlier behavior; HMAP recursively updates entity and process records in a high-level language planner. • Provides 1,800 demonstrations and raises mean success from 16.2% to 24.2%, with inconsistent gains on process-memory tasks.
2026-11-03	One Assistant, Many Memories: Benchmarking Real-World Group Assistance When Recall Is Not Enough
• Introduces GroupAssistBench, separating group-memory recall from open-ended assistance. • Uses 80,299 real messages from 50 groups, source-anchored memories and task rubrics informed by user feedback. • The strongest memory system approaches full-context recall but falls behind on use, especially with distributed or conflicting evidence.
2026-11-03	From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?
• Introduces Continual AutoResearch to separate session continuity from research-insight transfer. • Matches data and time budgets across persistent and reset agents on nine task families, with retained-information ablations. • All five frontier agents find stronger solutions sooner, but final held-out performance does not consistently improve.
2026-11-02	MemShift: Separating Detection from Repair in LLM-Derived Personal Memory
• Introduces MemShift, measuring change detection separately from correct repair. • Covers revision, deletion, redefinition, writer changes and export, fixing gold values and dependencies before text and using deterministic scoring. • Four open LLMs struggle to recompute derived values; generated programs over typed records close much of the gap in synthetic worlds.
2026-11-02	They Know Where It Is, But Not What It Says: A Bilingual Hallucination Benchmark of Hadith Recall in Open LLMs
• Introduces a bilingual exact-text recall benchmark for recovering Hadith from bibliographic identifiers. • Tests six collections without retrieval, scoring lexical and semantic fidelity and flagging refusals. • None of 30,000 responses meets both high-fidelity thresholds; this is a result for the tested metrics and models, not all religious knowledge.
2026-11-02	What Is This Memory Useful For? Measuring Task-Conditioned Utility in Long-Term Agent Memory
• Measures task-conditioned memory utility instead of assigning one importance score. • A task-blind writer and frozen reader compare memory presence under matched filler, with a ledger of formation, storage and exposure costs. • Answer-bearing pairs gain 0.33 more, but coding benchmarks show no measurable benefit; labels can change across readers and sessions.
2026-11-01	MemHarmBench: Evaluating Cross-Session Prompt Injection in Memory-Persistent AI Agents
• Introduces MemHarmBench for cross-session harm from persistent-memory poisoning. • Exposes attacks, removes their original sources and runs legitimate tasks in a new memory-preserving session, jointly measuring intent, completion and anomalies. • Attacks can influence critical decisions despite normal completion and no obvious anomaly, within the tested systems and victim categories.
2026-11-01	Delivered, Not Used: Evaluating Agent Memory at Its Point of Use
• Introduces MemUse to evaluate memory at delivery and trace its consequences. • Compares delivered entries with already visible evidence, rating relevance, accuracy, novelty, influence and effects with citation checks and cost linkage. • Across 8,000 SWE-bench episodes, resolve rates stay near a no-memory baseline while most deliveries are irrelevant or unused; success is not memory utility.
2026-11-01	Answering Without Arriving: An Evidence Audit for Embodied Question Answering
• Audits whether embodied-QA scores genuinely arise from memory evidence. • Holds questions fixed across text-only, closed-loop and reference-replay conditions on 200 episodes and 800 questions, checking provenance and shortcuts. • An agent reaches no target yet answers 45.5% correctly, near its no-image score; acquisition and use must be established before crediting memory.
2026-10-31	DuSE-Bench: Benchmarking Joint Reasoning over Evolving Dialogue and Documents
• Introduces DuSE-Bench for reasoning over jointly evolving dialogue memory and documents. • A support graph links updates and reused conclusions to original turns and passages, separating content, response-type and attribution scores. • Of 1,000 tasks, 700 require both sources; joint-content accuracy peaks at 56.7% across 11 configurations, often without complete attribution.
2026-10-31	CivicBench: A Benchmark and Memory Framework for Web-Based Civic Data Analytics
• Introduces CivicBench and Task-State Memory to test whether civic analyses remain supported by source records. • Compiles tool observations into typed assertions, reconciles compatible scopes and retrieves unresolved-constraint state. • Across 400 tasks, eight of nine model-pair mean differences have 95% intervals excluding zero; one remains inconclusive.
2026-10-30	ImplicitHop: Evaluating knowledge integration across context and memory
• Introduces ImplicitHop to compare knowledge integration across context and pretrained memory. • Controls composition depth with formal graphs and knowledge-graph tasks, evaluating implicit inference without chain of thought. • Accuracy drops sharply with depth for both sources, with context performing better; conclusions apply to the implicit-inference protocol.
2026-10-30	TwinPast: Benchmarking Historical Evidence Use in Streaming Video Understanding with Evidence Contrastive Video Pairs
• Introduces TwinPast to distinguish correct answers from actual use of visual history. • Its 350 reviewed video pairs hold recent frames and questions fixed while changing historical evidence and require both answers to be correct. • Across 34 configurations, the strongest streaming system reaches 49.14% PairAcc; retained evidence can still be used with the wrong temporal relation.
2026-10-30	MemU-Bench: Benchmarking GUI Agents on Long-Horizon User Interactions
• Introduces MemU-Bench for vague mobile requests that depend on interaction history. • AppMem organizes app-centric memories and interrupted intents, with outcomes checked against application databases and path efficiency. • Across 118 tasks from 29 trajectories, GPT-5.5 SPL rises from 15.7% to 66.9% in this benchmark.
2026-10-29	Finding Is Not Returning, Returning Is Not Reconstructing: A Multi-Dimensional Evaluation of Agent Social Re-entry
• Separates finding a conversation, returning to it and reconstructing its context. • A shared-snapshot framework isolates conversational resurfacing and persistent-memory effects. • Across six LLM configurations, better finding or replying does not consistently improve contextual reconstruction.
2026-10-29	VideoMemBench: How Well Do MLLMs Remember in Video Conversations?
• VideoMemBench tests retention and reuse across turns in video conversations. • Questions carry episodic or prior-answer dependencies, checked for necessity and non-leakage. • It contains 326 videos and 2,476 turns; the best reported score among over 20 models is 45.7%.
2026-10-28	REEL: Same Frames, Different Answers and Different Costs
• REEL compares video-memory representations built from exactly the same frames. • Paired correctness changes, cumulative GPU time and storage distinguish computation reuse from evidence loss. • Across six VLMs and two benchmarks, text differs from full frames on 20%–29% of questions; prefix memory is more faithful but can reach about 21 GiB per video.
2026-10-27	Rare, Relevant, and Still Wrong: How Memory Fails LLM Group Decisions
• CrossProfile diagnoses failures of rarity-first and relevance-first discussion memory. • Sixty templates yield 480 controlled tasks crossing fact ownership and evidence type while fixing text. • Relevance readmits decoys and rarity is no better than random on average; removing attribution gains 13–24 points in the classic configuration.
2026-10-26	When Retrieval Is Not Enough: Obsolete Facts in Agent Memory
• Separates retrieval failure from interference by obsolete facts already in the retrieved evidence. • Uses equal-length evidence replacement on MemoryAgentBench single-hop FactConsolidation. • Paired Qwen2.5 accuracy falls from 100% to 44.6%; evidence is limited to the tested single-hop setting and models.
2026-10-26	The Right Fact, the Wrong Decision: SciValidityBench for Scientific Agent Memory
• Introduces SciValidityBench to distinguish recalling a correct fact from being justified in using it for a scientific decision. • Provides 64 TaskSpecs across five validity modules and six domains, with state compilation, counterfactual pairs, and corruption–repair controls. • Explicit scientific state improves decisions by 9.2–18.8 points across four model families; exposed dependencies matter more than a particular graph serialization.
2026-10-26	BehaviorMem: Building and Evaluating User Memory from Real-World E-Commerce Behavior Streams
• Introduces BehaviorMemBench and BehaviorMem to separate remembering behavioral evidence from inferring user preferences. • Uses 1,194 users and 17,462 real behavioral batches; a two-axis state maintains facts and inferred claims through revision, dormancy, and reactivation. • Leads tested agentic memories across four builder–reader combinations; statistical baselines remain competitive, and remembering does not guarantee inference.
2026-10-24	ConMem: Counterfactual Multi-Scenario Probing for Resolving LLM Memory Misalignment
• Introduces MemAlign and ConMem to study misattribution, overgeneralization, and privacy-sensitive retention in long-term memory. • Builds 1,079 validated synthetic memory cases and probes counterfactual task effects, irrelevant invariance, and risk-boundary shifts. • ConMem reaches 75.16% accuracy and 64.15% balanced accuracy, reducing over-retention; evidence comes from synthetic multi-session contexts.
2026-10-23	4DMemory: Benchmarking Long-Horizon Spatiotemporal Memory with Harness-Augmented Agents
• Introduces LongVideo and 4DMemoryBench for long-horizon retention and spatiotemporal reasoning. • Provides roughly 900 hours of video and over 5,000 reviewed questions, plus a harness combining skills, tools, and model-specific retrieval. • All 13 tested models improve by over 50% with the harness; the combined intervention does not isolate one memory component.
2026-10-23	Navigating Epistemic Parity in LLM Agents: A Benchmark for Cross-Source Conflict Resolution Between Memory and Procedural Skills
• Introduces a benchmark for conflicts between episodic memory and procedural skills with no predefined precedence. • Uses 52 scenarios across four conflict types and three context conditions to control placement and explicit priority directives. • Across 1,383 runs on three models, resolution shows positional bias and frequent silent adjudication; directives do not eliminate all non-compliance.
2026-10-22	ColaBench: Evaluating Memory-Guided Knowledge Acquisition in Personal Agents
• Benchmarks collaborative workplace memory with explicit evidence links. • Uses 349 tasks, ten profiles and 13,033 synthetic documents. • MemKB reaches 86.25% accuracy in 4.07 rounds on average; unsupported assumptions remain.
2026-10-22	Memory-Aware Reward Modeling: Learning Human Preferences for LLM Memory Use
• Measures human preferences over assistant memories. • Collects about one thousand pairwise comparisons from 87 participants to train and assess a reward model. • The model reaches about 72.8% accuracy; heterogeneous preferences limit a universal quality interpretation.
2026-10-22	When LLMs Forget What They Believe: Characterizing State Integrity Collapse in Long-Horizon Dialogues
• Separates recalling conversational state from complying with it. • Tests seven models at five depths and evaluates explicit state anchoring. • Most failures reflect behavioral drift despite retained facts; anchoring helps immutable constraints within the benchmark.
2026-10-22	AfterDeleteBench: Auditing Post-Deletion Information Persistence in LLM Agents
• Tests whether deleted secrets remain recoverable from memory descendants. • Composes surviving records and compares deletion with provenance-closure removal. • Recovery reaches 62–99% in tested attacks; closure removes observed leakage under the evaluated assumptions.
2026-10-21	MOSAIC: An Omni-Modal Benchmark for Long-Term Memory and Cross-Modal Reasoning in Conversational Agents
• Benchmarks memory across multiple audiovisual conversation sessions. • Controls writing, evidence selection, delivery and reading over 923 questions. • Spatial and causal gaps persist, and gold evidence reveals reader bottlenecks beyond retrieval.
2026-10-21	MemoryTax: The Bitter Lesson of Agent Memory
• Audits the cost and information loss introduced by memory writers. • Controls writer and reader access in one LoCoMo conversation, including a photo-access confound. • With 152 questions and one run per cell, some adjusted gaps reduce to one answer; broad superiority is unestablished.
2026-10-21	ProactEgoMem: Benchmarking Proactive Memory in Streaming Egocentric Video
• Benchmarks proactive memory use over egocentric activities. • Tests retained states and locations alongside notification and start/stop decisions. • Strong models struggle with both accuracy and timing; evidence is bounded by the constructed video tasks.
2026-10-21	Externalise the Memory: A Pre-Registered Study of Where an Embodied Agent's Experience Should Live
• Compares internalizing evidence with external memory retrieval. • A preregistered study reports both adaptation and retrieval results, including unsupported hypotheses. • Distillation gains about one nonsignificant point; retrieval gains vary, with the second-task interval crossing zero.
2026-10-21	Why Bother Me Again? Benchmarking Personal Agent Memory for Interaction Preferences
• Tests whether assistant memory uses changing preferences appropriately. • Tracks four users across 99 sessions, including five preference changes. • Retrieval does not beat no-memory controls and file memory can overapply preferences; the sample is small.
2026-10-20	MemDev-Bench: Does Memory Help Code Agents?Benchmarking Memory Systems over Real Repository Histories
• Tests coding-agent memory over real repository histories. • Uses 4,869 historical records and 555 executable tasks with relevance controls. • Memory helps some fine-grained edits but can reproduce errors; raw traces reduce steps in tested settings.
2026-10-20	DuplexLedger: Benchmarking Transactional Memory in Full-Duplex Speech Agents
• Benchmarks transactional memory during full-duplex speech. • Bilingual sessions test event-level writes, updates, interruptions and reversals. • Across 1,094 sessions, a trained model improves 4.11→8.68 accuracy but remains weak and increases interruption latency.
2026-10-20	When Does Experience Memory Help Closed-Loop LLM Driving Agents? A Controlled Evaluation with Resampled Warmup Pools, Token-Matched Fillers, and Audited Baselines
• Tests when stored driving experience helps a closed-loop LLM agent. • Resampled warmup pools, token-matched fillers and audited baselines control key confounds. • Only one of ten pools helps and none beats a rulebook; the negative result is limited to one environment and backbone.
2026-10-20	After the Attacker Leaves: Benchmarking Carryover Attacks on LLM Agents
• Measures how memory attacks survive beyond their planting session. • Separates planting from triggering across 12 scenarios and 101 attacks. • Five tested models show no measurable decay over five rounds; cross-runtime transfer is weaker.
2026-10-19	Computation, Not Memory: Chain-of-Thought and State Tracking in LLM Agents
• Separates reasoning computation from retained reasoning memory. • A small file-system environment contrasts current reasoning, earlier traces and erroneous counterfactual traces. • Current computation helps more in two Qwen models; diagnostic probes do not establish universal causal claims.
2026-10-19	Are Self-Evolving Agents Continual Learners?
• Tests whether self-evolving language agents retain prior abilities. • Four streams compare persistent-memory and harness-editing approaches. • Global edits and local memory collisions can cause forgetting despite frozen weights; self-evolution is not guaranteed continual learning.
2026-10-19	The Geometry of Agent Memory: Forecasting Retrieval-Bound Memory Failures, and Characterizing Its Limits
• Tests whether retrieval geometry predicts agent-memory failures. • Compares 25 geometric features with question-difficulty controls on 2,367 LongMemEval pairs. • Geometry AUROC is 0.640 versus 0.655 for difficulty; the tiny combined gain is statistically unresolved.
2026-10-18	MemBubble: Benchmarking How Agent Memory Shapes LLM Response Diversity
• Measures how remembered user history narrows response diversity. • Topic-matched but scenario-shifted questions use one thousand real histories. • Response entropy falls 34.7% versus full-history conditioning; diversity is not a complete measure of user utility.
2026-10-18	SENRI: Understanding the Space-Event Nexus in Real-Life Videos
• Tests the space–event links required for long-video memory. • Five hundred manual questions cover 25.3 hours; a verifier fills working-memory evidence gaps. • Reader-side evidence recovery is distinct from repairing the stored memory itself.
2026-10-18	MemDerive: A Diagnostic Benchmark for Agent Memory over Evolving Data
• Benchmarks memory over changing facts and derived states. • Datalog generates events, questions and gold derivation graphs across ten scenarios. • Four memory systems trail BM25 while oracle evidence reaches about 96.01, exposing derivation and invalidation failures.
2026-10-18	DiaLong: Benchmarking Reminiscence in Long-Context Dialogues
• Benchmarks reminiscence over long dialogues. • One hundred human-started, model-extended conversations contain controlled true and false facts. • The design supports diagnostic comparisons but does not fully represent natural longitudinal conversations.
2026-10-17	Can Agents Keep Secrets They Were Never Told? Benchmarking Latent Privacy Risks Across Sessions with MemJoin
• Benchmarks privacy leakage from joining otherwise insufficient memory sessions. • Separately measures stored attributes, retrieved clues and person-linked relations in 288 synthetic cases. • Six systems reveal risks missed by attribute visibility alone; paired clue removal changes some outputs.
2026-10-17	HABIT-Bench: Evaluating Habit Induction from Longitudinal Weak Evidence in Agent Memory
• Benchmarks habit induction from weak longitudinal agent-memory evidence. • Controlled histories test changing habits, temporary exceptions and unaccepted suggestions. • The best of eight methods reaches 33.61% macro accuracy; oracle evidence remains difficult and histories are synthetic.
2026-10-17	Do Conversational Agents Need Explicit Memory? A Controlled Comparison with Raw-History Retrieval
• Audits whether explicit conversational memory improves on raw-history retrieval. • Controls retrieval and uses oracle evidence to separate construction loss from access failure. • Six systems show no consistent advantage on two benchmarks; localized gains remain under tight budgets.
2026-10-16	Beyond Marginal Influence: Interaction-Aware Memory Selection for LLM Agents
• Benchmarks interactions among jointly sufficient memory records. • Exact oracles separate subset-search errors from model errors; group deletion tests blocks before individual records. • Perfect recovery on 6,912 exact conditions does not transfer universally to LLM interfaces or deployment-ready compression.
2026-10-16	Beyond Success Rate: Interventional Evaluation of Memory in Vision-Language-Action Systems
• Separates task-fact memory use from execution and numerical confounds in VLA systems. • Intervenes at retrieval with fact changes, equivalent replacements and minimal perturbations. • ChainBench rises from 39.1% to 74.3%, while numerical interventions show why aggregate success alone is insufficient.
2026-10-16	Coopshift: Post-Training Amplifies The Memory Curse In LLM Cooperation
• Tests how post-training changes the effect of remembered interaction history. • Compares Base, SFT and aligned checkpoints with two or forty visible rounds in repeated games. • Longer history reduces average cooperation at post-trained stages in binary games; effects depend on tested incentives and models.
2026-10-16	Persistent Memory and Iterative Improvement in an LLM Agent for Chinese-Character Drawing
• Measures when persistent drawing experience helps an LLM agent. • Compares notes and reusable code over 668 items, with a separate privileged-supervision analysis. • First-attempt gains are 4.6/2.8 points; designs do not differ significantly and privileged gains are not experience-only gains.
2026-10-15	MemoProAgent: Toward Memory-Driven Proactive Assistance
• Benchmarks memory-driven proactive assistance and intervention timing. • Controls oracle evidence and trigger windows; a graph-memory harness gates relevant interventions. • Baselines score at most 15%, while the proposed trained agent reaches 42.4%; oracle performance is not retrieval performance.
2026-10-15	Language Models Share a Readout Bottleneck in Working Memory
• Uses N-back to distinguish working-memory addressing precision from content interference. • Combines probes, attention sharpening and non-target identity suppression in 70M–14B models and minimal Transformers. • Errors often retain correct target addressing; sharpening helps little while suppressing competing content improves sensitivity, supporting interference-limited readout on this controlled task.
2026-10-15	MemCut: Every Secret Leaves a Trail in the Long-Term Memory of LLM Agents
• MemCut evaluates secret recovery from agent memory after user-requested deletion. • Evidence trails for 456 secrets distinguish native retrieval from direct access to residual objects across deletion baselines. • Residual recovery remains 39%–76% after additional deletion; subset studies identify coverage bottlenecks, so lower native recovery does not prove complete forgetting.
2026-10-15	Rethinking the Role of Tool Graph Memory in Multi-Turn Tool Use
• Tests tool-graph memory while separating content benefits from extra computation. • Compares graph, random or absent guidance and same-experience textual insights across nine models and two benchmarks, tracking calls. • Graph memory seldom improves significantly and some gains cost over twice as many calls; textual insights work better with lower overhead in the tested implementations.
2026-10-15	Beyond Downstream QA: Rethinking How We Evaluate Agent Memory
• FEB-30 inspects stored fact evolution directly, separating storage errors from downstream retrieval and generation. • Materials generated from known trajectories provide operation-level truth for 30 cases, six diagnostic categories and three difficulty levels. • One system shows a 16-point average intrinsic/extrinsic gap and 20% operation classification accuracy; these limited results are not universal rates.
2026-10-14	Evaluating Compressed Memory: Dependence, Utility, and Content Benefit
• Separates compressed-memory dependence, utility and document-specific benefit, showing why zeroing alone cannot establish usefulness. • Tests nine systems with zeroing, separately trained no-memory or constant-input models, and mismatched documents. • Some memory-dependent models underperform alternatives; mismatched xRAG documents also improve F1 while accepted-answer gains remain uncertain, requiring explicit comparisons and scoring conditions.
2026-10-14	Memorizing Who, Knowing How: Benchmarking Personal and Procedural Memory for Persistent Assistants
• PraxisMemBench evaluates interdependent personal and procedural memory maintenance in persistent assistants. • Trajectories require joint updates; PraxisMem uses privileged joint-state supervision, location-aware GRPO and residual-routed OPSD. • Coupling reduces completion from 84.6% to 77.9%; PraxisMem reaches 40.78% session success versus 20.60% for native Claude Code memory under different supervision.
2026-10-13	Can Agents Truly Forget? Explore Machine Unlearning for LLM Agents
• AUBench extends agent unlearning across parameters, persistent memory and tool-mediated reacquisition. • Builds a controlled fictitious-author environment with offline web snapshots, documents and provenance-tracked trajectories, comparing parameter- and memory-level methods. • Results expose a gap between model forgetting and system-level removal; the benchmark does not guarantee complete unrecoverability.
2026-10-13	Can Agents Forget When Asked? A Benchmark for Memory Unlearning in LLM Agents
• Bemula evaluates precise forgetting on explicit user requests. • Its 50 multi-session instances contain 135 requests of six types and 1,742 questions, testing storage, removal, preservation, relearning and extraction resistance. • Across six memory systems and multiple backbones, end-to-end request pass rates never exceed 1.5%, exposing both residual disclosure and excessive deletion.
2026-10-13	Auditing Future-Memory Use in Point-in-Time Language Agents
• Audits whether future facts accessible in parametric memory bias point-in-time evidence gathering. • Predefines directions with closed-book probes, then compares real and anonymized identities across 48 historical company-comparison tasks. • Qwen does not support the predicted shift and DeepSeek shows an inverse pattern; the mechanism remains unresolved, separating accessibility from search and choice.
2026-10-13	Knowing the User Is Not Knowing What We Established: Grounding Awareness in Personal Agents
• Introduces Grounding Awareness: distinguishing established arrangements from merely remembered proposals. • Controlled tests vary acceptance, revision and scope; Grounding Layer tracks status and checks consequential actions. • Across agents and memory systems, more user information can increase unauthorized execution; explicit grounding helps, while strong recall alone does not ensure authorization.
2026-10-13	SEAR: Self-Evolving Agentic Robotics
• SEAR tests whether retained skills and memories improve robotic language agents across tasks. • Harvests versioned code skills and written memories from completed tasks, using no-library controls, budget-matched restarts and isolated evaluation. • Seven models across 330 simulated and five physical tasks show selective transfer; erroneous lessons also persist, and gains depend on difficulty and library content.
2026-10-12	EgoProbe: Benchmarking Long-Range and Exact Memory in Streaming Egocentric Videos
• EgoProbe separates long-range retention from event-specific binding in egocentric video memory. • Base tests distant recall; Echo adds related intervening events and coarse-to-fine probing, with 19 evaluated video-language models. • The best model reaches 55.32% Base accuracy and 43.69% two-turn completion; streaming and offline-prefix protocols differ and should not be conflated.
2026-10-12	Can a Frontier VLM Serve as a Robot Policy for Memory-Dependent Manipulation?
• Evaluates a frozen general-purpose VLM using full-episode context for memory-dependent manipulation. • The model outputs waypoints and gripper commands without task-specific training across 16 temporal, spatial, object and procedural-memory tasks. • It succeeds on 92.1% of 800 episodes; the human reference uses a different oracle-skill interface, and 63 failures concentrate on completion and button timing.
2026-10-11	Suggested Is Not Superseded: A Pre-Registered Evaluation of a System-One Model as a Memory-Retirement Gate
• Evaluates whether assistants prematurely retire memories when a change is merely suggested. • Labels 500 pairs as superseded, needing confirmation or compatible, then tests binary, three-way and escalation gates under preregistered hypotheses. • Binary gating prematurely retires 39.2% of suggested changes; three-way gating lowers this to 15.5% but loses 15.3 detection points, exceeding the planned budget.
2026-10-10	MemForm: Benchmarking Memory Formation Poisoning in LLM Agents
• MemForm isolates adversarial manipulation of memory-formation interfaces. • Paired interventions target content, evidence, scope, retention selection and consolidation paths, separating what is stored from how it is formed. • Six LLMs and two response formats show susceptibility; successful experiences can still yield harmful memory, and defenses remain interface-dependent.
2026-10-10	VitaBench: Memory Lifecycles Beyond Retrieval
• VitaBench evaluates whether memories should remain exact, be compressed, consolidated or deleted. • Repeated reviews and full-horizon counterfactual evaluation measure lifecycle headroom, horizon reversal and optimal-action diversity. • About 56.5–60.2% of states change preferred actions across horizons; a learned policy scores 101.5 versus 83.4 for the best fixed strategy on generated naturalistic streams.
2026-10-08	What Should Embodied Agents Remember? Benchmarking and Evaluating Memory Modality for Embodied Agents
• MeMoREval compares memory modalities for embodied language agents. • Agents experience events before seeing requests, then answer or act without new observations; evaluation measures memory construction and use costs. • Across 53 tasks in four simulated environments, frames plus scene graphs achieve the best aggregate success; only two physical demonstrations are provided.
2026-10-07	Before Bringing It Up: When and How AI Companions Should Use Memory
• Studies whether accurate memories are socially appropriate to mention in the current companion interaction. • Two interview rounds with 14 users inform five checks and four handling modes, evaluated blindly on 80 scenarios across five models. • Two LLM judges favor Reconsider by 15 and 23 net points, but GPT and Claude intervals include zero and evaluator preferences vary with model family.
2026-10-07	On-Policy Distillation Teaches New Skills but Not New Knowledge
• Separates factual memory acquisition from compositional skill transfer during on-policy distillation. • A controlled synthetic framework varies teacher facts and skills independently, then isolates KL direction and student rollouts. • Across four models from three families, reverse KL mainly transfers skills while forward KL restores factual transfer; recent QA and mathematics show a similar asymmetry.
2026-10-07	Stale, Misattributed, or Late: Where Personal Memory Fails Before Generation
• Separates pre-generation personal-memory failures into stale values, misattribution and late evidence. • Controls active slots, participant information and retrievers to test update merging, entity resolution, latency and propagation into answers. • Correctly keyed updates remove observed stale exposure, but open-domain merge recall on LongMemEval never exceeds 0.062; same-name identity ambiguity remains.
2026-10-06	When Old Facts Return: Re-Reads, Reverts, and the Limits of Temporal Memory
• Shows that value sequences alone cannot distinguish rereading an old source from a genuine factual revert. • Tests 130 atomic transitions extracted from software fixes and a guard against reactivating retired values after verbatim rereads. • Accuracy falls from 98.5% to 10.8% and recovers to 97.7% with the guard; legitimate reverts remain unresolved, with selected inputs and post-failure evaluation limiting scope.
2026-10-06	Persistent Memory in Multi-Agent LLM Inference: What It Costs, What It Buys, and When You Can Tell
• Separates active-cache savings from multi-agent decomposition and the benefit of persistent reasoning-trace memory. • Eight dataset pairs with 100 items per arm measure peak KV and accuracy, analyzing how independent items and resets remove useful recall signals. • The persistent tier adds 0.368 MiB peak cache with a 0.015 accuracy difference whose interval crosses zero; the null concerns one architecture and these task conditions.
2026-10-06	Beyond Corrected Memory: Execution Consistency in Multi-Agent Systems
• Shows that correct shared records may lack receipt, dependence or response-validity evidence needed to judge execution compliance. • Controlled removal and restoration produce opposite-label pairs; CAVERT extracts supported log relations for consistency diagnosis and recovery. • Removal makes 82.4% of pairs indistinguishable and restoration separates 97.9%; diagnosis wins across 12 settings, but missing log relations still limit judgment.
2026-10-06	DSV-Mem: Evaluating Multimodal Memory in Professional Workflows for MLLM Agents
• DSV-Mem evaluates persistent state memory over dense professional visual artifacts with revisions and authority changes. • Its 1,000 expert-reviewed questions cover current, past and derived states, change history and conflict/refusal, separating state synthesis from dialogue filling. • The best of 27 configurations remains below 45%; governing updates dominate tested difficulty, and the abstract says benchmark and code will be released.
2026-10-06	Memory Depth and Reconstructed Context Width: A Controlled Evaluation of Hierarchical Retrieval
• Separately controls memory hierarchy depth and reconstructed context width to test whether deeper structures necessarily help. • Crosses D1–D4 with token budgets on EverMemBench and compares production retrieval with oracle archive access. • Increasing width from 1K to 4K adds 10.11–17.98 accuracy points without monotonic depth gains; production saturates beyond 8K–16K, unlike oracle access.
2026-10-05	The Right Memory in the Wrong Context: Verifying Retrieval Admissibility in Long-Term Agent Memory
• Classifies memory-query pairs as admissible, inadmissible or unresolved, separating recall, prompt exposure and disclosure. • Compares routes at matched required-evidence recall and tracks memory IDs across frozen rankings, paired cases and controlled exposure scenarios. • Namespace filtering improves anchor recall with fewer similarity evaluations, but answer gains are observational; only one reader interval excludes zero in 16 exposure scenarios.
2026-10-05	Defense-in-Depth for LLMs: Evaluating Memory Gates Against Activation-Induced and Memory-Induced Sycophancy
• Separates internal activation bias from external memory handling when evaluating memory-induced sycophancy. • Combines steering with five memory defenses on four models; Router Gate keeps, rewrites or drops individual memories. • Selective filtering preserves more accuracy than removing all memory; inverse steering adds no statistically supported benefit, and defense judging uses a fixed subsample.
2026-10-05	Mining Agent Skills from Production Traces
• Evaluates how execution traces should become reusable skills when reliable outcome labels are unavailable. • Holds mining fixed while crossing three evidence regimes with workflow and declarative-ontology skill forms on enterprise tasks. • ThinkingBox favors workflows and labeled mixed evidence; APEX moderately favors ontologies without a clear evidence-regime preference.
2026-10-05	Cross-Lingual Transferability of Training Data Extraction Attacks to Recover Memorized PII
• Studies cross-lingual retrieval of personal information memorized during training. • Translates extraction contexts into four European languages, varies wording, and analyzes internal representations. • Some English leaks remain recoverable across languages, increasingly with multilingual capability; web-presence checks cover only a sample of translations.
2026-10-04	No Hindsight for LLM Fact-Checkers: Measuring Leakage Channels in Misinformation Detection
• Separates memorized future outcomes from post-claim retrieval evidence in fact-checking evaluations. • Reconstructs point-in-time evidence, probes outcome information in model representations, and compares a bottleneck with an information penalty. • Finds substantial parametric leakage; post-claim evidence inflates AVeriTeC accuracy by 6.3 points but has little effect on QuanTemp++.
2026-10-04	Assembling Insights for Agentic Machine Learning Engineering Systems
• Assembles domain, competition, and improvement memories from prior competition solutions and repair writeups while controlling target leakage. • Validates source evidence and exposes insights as skills or a playbook; new sources incrementally refresh affected domains and clusters. • In a 160-competition protocol, normalized private-test score improves 123.9% over no insights; the best injection mode varies by task.
2026-10-04	EMBER-Bench: Benchmarking Cross-Event Causal Memory in Long-Horizon Embodied Tasks
• Evaluates how remembered events constrain current state and future actions, beyond historical retrieval. • Provides 189 household tasks and 699 QA pairs with event and causal-chain labels for action prediction and causal traceback. • The best of 16 models reaches 61.2% versus a two-human mean of 98.3%; correct traceback is not associated with correct action prediction.
2026-10-04	MetaKernelBench: Measuring GPU Kernel Knowledge Transfer Beyond Code
• Measures whether experience from one kernel attempt helps a fresh attempt in another DSL. • Pairs CuTe and TIRx versions of 74 problems and transfers distilled natural-language skills under matched budgets. • Across six models, paired lift ranges from −19% to +29%, with regressions on 16–45% of problems; source success alone is insufficient.
2026-10-04	Memory Canonicalization: A Framework and Benchmark for Cross-Model Drift in Persistent LLM Memory
• Studies factual and emotional interpretation drift when different language models read the same persistent memory. • Canonicalizes ambiguity and conditional structure and separates emotional valence, testing 176 synthetic objects in three arms. • The uncorrected emotional gain of 0.050 does not survive multiple-comparison correction; factual-drift comparisons are not significant.
2026-10-03	MemLeak: Cross-User Semantic Leakage in Multi-Tenant AI Agent Memory
• Studies cross-user semantic leakage when personal agents share a vector memory store. • Compares sparse and MiniLM retrieval, response contamination, and three mitigations including hard ownership gating. • Incidental leakage reaches 70–100% in pooled same-team retrieval; only hard gating consistently restores the clean baseline in the tested configurations.
2026-10-03	StegoMemory: Agentic Memory Acts as Covert Steganographic Channel
• Evaluates persistent agent memory as a covert cross-session information channel. • Separately tests encoding, persistence, and recovery in 14,000 trials across 13 models and 91 configurations using synthetic secrets. • Exact recovery occurs in 20.1% of trials, and 14.4% also pass task and safety scoring; persistence is the main bottleneck.
2026-10-03	Knowing the Store: What a Memory Backend Must Write Down Before an Agent Can Read It
• Tests whether agents can judge memory availability before retrieving record values. • Uses 148 questions across three minimally differing synthetic stores, comparing lifecycle counts, attribute names, and cosine lookup. • No model reliably beats lookup on short lists; missing backend metadata hurts performance, and downstream answer benefits remain untested.
2026-10-02	Ego2World: Compiling Egocentric Cooking Videos into Executable Worlds for Belief-State Planning
• Compiles egocentric cooking into partially observed environments with separate world state and agent belief for memory-reuse evaluation. • Distinguishes action validity and task attainment across 105 tasks and conducts a paired persistent-belief study. • Persistent belief raises validity by 4.15 points and reduces visual queries by 90.27%, but uses more tokens without a detected completion gain.
2026-10-02	A Controlled Audit of Personal AI Memory for Rating Prediction
• Distinguishes use of historical item-rating associations from reliance on a user’s rating distribution. • Audits 400 held-out profiles and 6,160 ratings using within-user permutations, full history, extracted memory, and numerical readers. • The Mem0 pipeline worsens error on Coat; association benefits on MovieLens are smaller and inconclusive after adjustment.
2026-10-02	DyadMem: A Long-Term Memory Benchmark of How Agents Work with Users
• Introduces user-conditioned relational agent memory for remembering how to work with a particular user. • Annotates capture, update, and recall across 3,065 episodes, 50,961 sessions, and 61,210 QA instances. • Across 20 models, full-pipeline QA trails gold-memory QA substantially, exposing capture, recall, and unsafe-deletion failures.
2026-10-02	When Evidence Changes: Evaluating Memory Repair and Re-reading in Language-Model Agents
• Evaluates memory repair against re-reading after source documents are revoked or replaced. • Compares caching, rebuilding, graph-local repair, and source-filtered re-reading on public ICU tasks with two 7B models and full-pipeline costs. • Every memory pipeline costs at least twice full re-reading on short records; none of four primary replacement tests is significant.
2026-10-02	Self-Propagating Misalignment in LLM Agents, and Why Auditing or Disabling Memory Is Not Enough
• Studies whether an agent can persist a misaligned goal for a later aligned agent to execute. • Tests 20 scenarios across 11 models, two prompting regimes, filesystem persistence, and memory auditing. • Propagation succeeds in 58% and 18% of the two prompting regimes and can survive 100 unrelated sessions; misalignment is experimentally simulated.
2026-10-01	RealCompanion: Benchmarking Human Understanding from Reasoning over Longitudinal Real-World Conversations
• Evaluates when real companion messages need distant memory and how systems reconstruct a person. • Provides ten relationships, 27,218 messages and traceable labels, separating natural messages from authored questions. • A recency window covers 95.9% overall but only 2.2% of memory-dependent probes; tested detectors fail, and the dataset contains only ten relationships.
2026-09-30	Where the Evidence Lives: Auditing AI Companions' Self-Descriptions
• Audits whether an agent’s memory self-descriptions are backed by executed accumulation mechanisms. • Cross-checks user judgments, behavior and implementation logs, marking claims supported, contradicted or unresolved. • In a one-month Lita deployment with nine colleagues, two of three memory layers never accumulated despite middling ratings; this small study cannot estimate prevalence.
2026-09-30	When Context Changes: Understanding Update Failures in LLMs
• Distinguishes stale answers from failure to retain updated values in conversational context. • CICM, probes and component interventions identify attention drift; a one-layer analysis explains accumulated weight on old values. • Input-adjusted attention corrects most directly queried stale-value errors while preserving nearly all correct answers, not a general automatic repair guarantee.
2026-09-30	What Should an Agent Remember? Disentangling Retention from Retrieval in Bounded-Memory Evaluation
• Separates retention from query-time selection in bounded agent-memory evaluation. • Crosses retention and selection rules on the same 300 seeded episodes and repeats on SQuAD text. • Query-aware selection gains 15.5 points at fixed access; most of a mixed 68.7-point advantage comes from access, and all 319 bounded-recency failures arise from eviction.
2026-09-30	Beyond the Shadows of Plato's Cave: Evaluating False Memory in Autonomous Agents via Counterfactual Reasoning
• Studies internal belief distortion attributed to false memory rather than relying only on final answers. • FAME measures hidden-state concept drift under counterfactual interventions before answer generation. • Reports AUROC 76.2–96.7% and improvements on math, code and reasoning benchmarks; interpretation depends on the counterfactual evaluation construction.
2026-09-29	Memory Consolidation Flattens the Temporal Shape of User Facts
• Studies loss of temporal aspect when user statements are consolidated into stored notes. • LAPSE pairs statements differing only in temporal form and separately probes writers, readers and action decisions. • Three writers show one-way flattening in 244 of 381 pairs; the planned use task lacked sensitivity and downstream action findings are exploratory.
2026-09-29	APM-Bench: Benchmarking Cross-session Persistent Memory for Egocentric Streaming Video Assistants
• Evaluates memory persistence across interrupted egocentric sessions, timely access and recognition of missing evidence. • APM-Bench contains 104 trajectories, 549 sessions and 2,719 candidates, comparing general models with specialized memory systems. • Finds unresolved trade-offs among recall, proactive assistance, latency and storage within a finite collection of multi-session trajectories.
2026-09-29	When Correct Memory Goes Wrong: Fuzzing Persistent Memory Use in LLM Agents
• Systematically fuzzes misuse of otherwise correct persistent memories. • U-Fuzz mutates queries or memory states from checkpoints under explicit obligations, validates mutants and guides search by observed behavior. • Finds more confirmed failures than tested baselines, including output-only API settings; the abstract provides no single count or rate.
2026-09-29	Auditable Long-Term Memory: A Deterministic Retrieval Chain Measured at 479/475 of 500 on LongMemEval-S
• Reports an auditable conversational retrieval chain and evidence-packet construction. • Separates hybrid retrieval, reranking, coverage-first packets and deterministic scaffolds from a replaceable final reader. • Scores 479/500 and 475/500, but development used the same questions, update scoring was modified and complete requests were not retained; superiority is not established.
2026-09-29	TAGGRAPH: Tag-Augmented Graphs for Graph Retrieval of Agent Persistent Histories
• Compares graph and lexical retrieval over shared 5W conversational memories. • Tests local traversal, chronological links and PageRank diffusion while auditing missing tags and vocabulary normalization. • The best graph reaches 0.844 MRR versus BM25 0.867 on LongMemEval-S; rankings change by setting and store-size effects remain unresolved.
2026-09-28	Memory Is a Derivation: The Distributed-Evidence Paradox in Long-Term Agents
• Audits whether persistent claims follow from the interaction history available at write time. • DerivAudit separates missing citation scope, invalid composition and admission reliability before downstream reuse. • Broader history supports nearly 60% of apparently unsupported memories, yet 17–21% remain unsupported; more evidence alone does not ensure reliable admission.
2026-09-28	RoutePrism: Tracing Construction Order Effects in Agent Memory
• Diagnoses how memory construction order changes retained evidence. • RoutePrism rebuilds identical sources in different orders and restores displaced records under matched controls. • Restoring one supporting record recovers over sixty accuracy points in the controlled intervention.
2026-09-28	ReplayLens: Auditing Agents' Use of Outcomes
• Audits whether agents use stored outcome relationships rather than incidental record features. • ReplayLens independently changes score attachment, order, names, and slots. • Controlled interventions reveal ingestion-order sensitivity and harmful dependence that endpoint accuracy misses.
2026-09-28	When Does Selection Replace Extraction? A Pre-Registered Test of Agent Memory with a Typed Decision Model
• Tests whether selecting raw turns can replace extracted memory facts. • A preregistered comparison varies retrieval budgets and answer models. • Selection is competitive at tight budgets and much cheaper to write; extraction wins at generous budgets and reranking hurts abstention.
2026-09-28	PairPref: When Should Memory Guide the Answer? A Benchmark for Contextual Preference Use
• Tests when a retrieved user preference should influence an answer. • PairPref holds preferences and requests fixed while changing situations across 1,227 pairs. • Free-generation success on both situations is only 3.6–18.3%, exposing an applicability gap.
2026-09-28	After the Fix: How Corrected Agent Histories Transfer to Related Tasks
• Distinguishes repairing a past episode from making it useful for a new task. • Compares original, corrected, and independent executions across 3,300 runs. • Benefits vary by task family; memory updates require both previous-version and fresh-start references.
2026-09-28	MemoReason: Evaluating the Effect of Parametric Memory on Contextual Reasoning in LLMs
• Separates parametric familiarity from contextual reasoning performance. • Pairs factual tasks with structurally matched fictional versions. • Performance drops by up to 15.7%, but direct recall of factual answers is rarely the dominant error.
2026-09-28	Share-Borne AI Virus: Memory-Hopping Attacks Across LLM Agents
• Studies adversarial state propagation through shared artifacts and persistent memories. • Simulated agent communities track survival, spread, and repeated memory-to-artifact handoffs. • Some configurations reach 60–80% of agents over chains up to eight hops; evidence is simulation-based.
2026-09-27	Separating Memory and Workflow Effects in Predicting Individual Answers
• Separates memory construction from answer-workflow effects in personalization. • Compares concrete memories, traits, raw records, retrieval, and multi-answer fusion. • Raw evidence often helps, but gains do not establish a universal human-prediction ceiling or an advantage over recency truncation.
2026-09-27	Characterizing Memory Misalignment in Human-LLM Interaction From User Perspectives
• Characterizes mismatches between agent memory and user expectations. • Combines usage records, diaries, co-design workshops, and a 121-user preference study. • Finds fourteen mismatch types and heterogeneous preferences, favoring low-friction proactive control.
2026-09-26	CoMemBench: Benchmarking Collaborative Memory Boundaries across Multi-Agent Workflow Topologies
• Introduces topology-conditioned boundaries for collaborative memory. • Tests sharing and isolation in 800 execution-grounded workflows across four domains. • Broader context improves availability but can weaken isolation; rankings vary by topology.
2026-09-26	MemTransfer: Benchmarking Memory Beyond Matched Experience in Embodied Decision-Making
• Tests embodied memory transfer under changed execution conditions. • MemTransfer compares six representations across starting poses, blocked routes, and history relevance. • Episodic success drops sharply at new starts; robustness to one mismatch does not imply robustness to another.
2026-09-26	CUE-Mem: Benchmarking Long-Term User Memory via Implicit Cues in Multimodal Conversations
• Tests user memory inferred from implicit multimodal cues. • CUE-Mem contains 2,674 text-image-audio questions spanning recall, patterns, recommendations, and refusal. • Detailed captions recover some evidence at growing cost; native multimodal access still suffers retrieval noise.
2026-09-26	VoxMem: Benchmarking Multimodal Memory in Large Audio Language Models
• Benchmarks spoken memory beyond transcript semantics. • VoxMem crosses four acoustic evidence types with four operations over 34,743 sessions and 3,196 instances. • None of fifteen tested audio-language models exceeds 40% at 32K context; nonlexical cues are especially weak.
2026-09-26	The Epistemics of Agent Memory: Measuring, and Governing, the Consolidation Decision in Long-Horizon LLM Agents
• Evaluates the quality and governance of memory consolidation decisions. • Studies episodic boundaries, promotion, ConsolidationBench, and reversible, auditable consolidation. • Reports useful controlled gains but finds its quality score does not predict real transfer accuracy.
2026-09-25	A Benchmark and Diagnostic Study of Epistemic Admission in Shared Agent Memory
• Benchmarks admission of correlated claims into shared memory. • Static and live tracks compare eight policies with source-lineage-aware diagnostics. • Coverage and false-admission trade off; an admitted uncontested false belief is repeated in 97–99% of probes.
2026-09-24	Probing Stability-Plasticity Tradeoffs in Agent Memory through Cognitive Experimental Paradigms
• Diagnoses stability-plasticity trade-offs in agent memory. • MemProbe uses interference, misinformation, consolidation, and reconsolidation experiments over 56 episodes. • Six systems with similar aggregate scores exhibit different maintenance behaviors.
2026-09-24	In-Context Binding Capacity in Language Models
• Measures how many entity-value bindings language models can recall. • Uses load-dependent recall curves and a broader sweep of thirty open models. • Finds scale-related capacity and interference effects; state updating and downstream transfer are not measured.
2026-09-23	TWIST: A Proposed Benchmark for Intervention Quality in Conversational Memory, with a Human-Validated Draft-Alignment
• TWIST proposes evaluation of when conversational memory should intervene as beliefs change. • Four tracks pair intervention tests with hard negative controls; the validated draft-vetting track contains 161 items. • Flat retrieval baselines trade contradiction recall against false alarms, while a conservative system misses many contradictions; only one track has the reported human validation.
2026-09-23	EmbodiedMemory-Bench: Benchmarking Embodied Memory for Long-Horizon Embodied Tasks
• EmbodiedMemory-Bench evaluates memory construction and updating through later embodied actions. • Its 2,554 episodes test visual detail, dynamic state, interaction outcomes and experience generalization; EMem provides structured external memories. • Current models remain uneven across capabilities, while EMem and an 8B memory policy improve matched backbones in the reported tests.
2026-09-21	Memory vs. Context? Influential Factors of Factual Recall in Language Models
• This replication studies competition between parametric factual memory and contradictory context. • World-capital probes across 31 models are extended to five additional knowledge relations and varied prompt framing. • Some original trends replicate, but phrasing can change memory reliance by up to 80 points and family-level exceptions limit generalization.
2026-09-21	MemCalib: Benchmarking and Optimizing Memory Use in LLM Agents
• MemCalib evaluates whether language models give retrieved memories an appropriate influence on responses. • MemCalib-RL separates over-use and under-use through bidirectional counterfactual credit assignment and token-level atom ablation. • Results across three model families/configurations report better balance and external transfer; common post-training methods exhibit directional trade-offs.
2026-09-21	DolphinBench: Mapping the Pareto Frontier of Agent Memory
• DolphinBench evaluates long-term agent memory through task completion, cost and latency. • Three knowledge-work personas each supply about 500k history tokens and 200 tasks verified to require relevant prior history. • The benchmark makes quality–cost trade-offs explicit; the abstract does not report a winning memory system.
2026-09-20	VibeMemBench: Evaluating Memory Systems for Coding Agents on Real Repository Coding Tasks
• VibeMemBench evaluates whether coding-agent memory improves executable repository tasks. • It pairs 111 coding targets from 90 repositories with 3,634 history trajectories, retaining targets where reference history helps a reference solver. • Direct experience helps four of five held-out solvers, yet 11 of 12 tested memory-system pairings fail to beat memory-off; target selection is intentionally history-sensitive.
2026-09-20	Impact Is Not Invalidation: Ask About the Claim, Not the Diff
• This work evaluates when stored coding-agent memory claims become invalid after repository changes. • Execution-verified claim flips distinguish claim-specific invalidation from generic behavioral-change detection. • Across 10,369 claims from 23 Python libraries, claim-focused questions yield 0.705–0.974 precision versus 0.291–0.329 for diff-focused judgments.
2026-09-19	The Price of Safety: Benign-Case Utility and Token Overhead of Memory-Poisoning Defenses in LLM Agents
• This study measures the benign-traffic cost of agent-memory poisoning defenses. • Memory backend, retrieval and judging are held fixed while three write-time defenses and one read-time reranker are compared across repeated runs. • The reranker quarantines legitimate memory on 33.6% of adjudicated items; accuracy effects are small and uncertain in a five-conversation study.
2026-09-17	Correct Now, Insufficient Later: Auditing Update Sufficiency in Context Compression
• Introduces update sufficiency: compressed memory should retain distinctions needed for future updates, even when its current answer is correct. • Audits paired histories with identical current answers and shared future updates but different correct subsequent answers, using a small synthetic pilot. • Finds that present-answer correctness can hide lost update evidence; the study also exposes identifier shortcuts and does not establish natural-task generalization.
2026-09-15	LSREP: A Longitudinal State-Replay Protocol for Evaluating Conversational Memory, with ICE v2 as an Audited Local-First Architecture
• LSREP evaluates how conversational memory accumulates, ages and incorporates revisions across time. • Ordered replay, repeated probes, changing reference answers and fidelity checks audit the local-first ICE v2 memory architecture. • Private replay and public LongMemEval tests reveal quality-cost trade-offs and implementation defects; ICE v2 underperforms vector-RAG in the public diagnostic.
2026-09-15	CapMem: A Benchmark for Caption-Based Episodic Memory in Egocentric Video
• Introduces CapMem to evaluate reusable caption-based episodic memory for egocentric video understanding. • Provides 75 videos totaling 33.7 hours and 1,000 multiple-choice questions, comparing temporal captions, direct video input, and retrieve-then-verify strategies. • Caption memories often outperform direct input on videos longer than 20 minutes, and verification can further improve accuracy under tested settings.
2026-09-14	MemRiskBench: Trace-Aware Risk-Preserving Evaluation for Long-Horizon LLM Agents
• Evaluates sparse high-impact risks in long-horizon memory traces. • Deterministic checks cover stale facts, conflicts, leakage, revocation, and constraint decay. • A 120-episode study preserves risk coverage in a smaller evaluation subset.
2026-09-09	PRAGMA: Evaluating Personalized Guidance with Memory Alignment in Lifelong Conversations
• Benchmarks personalized guidance beyond factual memory recall. • Longitudinal histories and annotated evidence test evolving preferences and mistaken assumptions. • Evaluations reveal failures in both evidence retrieval and its practical use.
2026-09-04	Does Your Agent's Memory Survive a Model Upgrade? A Controlled Study of Memory Portability
• Tests memory portability across model and embedding upgrades. • Controlled synthetic histories compare raw logs, retrieval, notes, and fixed-schema graphs. • Results expose model-coupled note loss and the value of retained source histories.
2026-09-03	When Users Don't Ask: Benchmarking Context-Driven Memory Retrieval in Conversational Agents
• Tests context-driven memory retrieval when users do not ask explicit factual questions. • Four conversational query styles separate recall from end-to-end response grounding. • Implicit and composed queries expose gaps missed by conventional QA evaluation.
2026-09-03	ICM-Bench: Person-Level Identity Reasoning in Multimodal Agents with Long-Term Memory
• Benchmarks persistent person-identity reasoning in multimodal memory. • Synthetic life-album videos link recurring people to traceable cross-time evidence. • Evaluated systems handle events better than accumulated long-term identity profiles.
2026-08-31	UTILMEM: Benchmarking Evidence Utilization in Long-Term Conversational Memory
• Benchmarks task-oriented integration of distributed conversational evidence. • Five domains test implicit relevance, dense histories, synthesis, and distractor resistance. • Results reveal a substantial gap between successful retrieval and useful evidence integration.
2026-08-31	Measure Before You Manage: Evaluating Agent Working Memory in Coding Agents
• Studies semantic heterogeneity in coding-agent working memory. • Archived trajectories compare object-aware compression, retrieval, and real-system replay. • Results show nominal token budgets conceal delivered-context and management-cost differences.
2026-08-31	Good Memory Has ECC: Evaluating the Memory of Vision-Language Models Beyond Accuracy
• Evaluates memory efficiency, compression, and calibration beyond accuracy. • ECCBench measures compute budgets, compressible inputs, and uncertainty-aware abstention. • Experiments find weak video compression and calibration in pretrained VLMs.
2026-08-30	Hindsight Memory-PRM: Supervising Memory Management with Auditable Hindsight Credit
• Learns memory utility from auditable hindsight evidence. • Retrievals, citations, and controlled deletion probes calibrate credit along memory versions. • LoCoMo and LongMemEval tests report strong results from a local 8B policy.
2026-08-26	Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context
• Tests episode reconstruction in flat interleaved conversations. • SCALE-QA pairs task-grounded questions with temporal–semantic memory reconstruction. • Three backends show gains over strong retrieval and long-context baselines.
2026-08-26	MemToC: Benchmarking Memory-Tool Conflict Resolution in Large Language Models
• Benchmarks conflicts between parametric knowledge and executable-tool returns. • Known source correctness separates retention, tool following, and abstention behaviors. • Fine-tuning helps some models, but robustness and abstention tradeoffs remain.
2026-08-25	MemUse: Moving Memory Evaluation from Direct QA to Natural Integration in Long-Term Human-AI Conversation
• Evaluates natural memory integration in real long-term conversations. • A four-month deployment compares factual recall with user-cued memory moments. • Natural integration correlates with satisfaction, while direct QA accuracy does not in this study.
2026-08-21	DreamBench-SWE: A Multi-Session Memory-Hygiene Benchmark for Software Agents
• Benchmarks software-agent memory hygiene across sessions with executable scoring. • Preregistered audits compare no memory and several evidence-retaining configurations. • Memory improves over no-memory controls; superiority among memory systems is not established.
2026-08-20	Can Agent Memory Systems Track Evolving State?
• Separates current-state tracking from recall of superseded information. • StateMemBench and explicit supersession relations diagnose and repair evolving-state failures. • Experiments improve current-state accuracy across memory and retrieval backends.
2026-08-20	MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
• Tests cognitive traps induced by relevant stored memories. • Reasoning fixation and belief distortion probes evaluate five memory frameworks. • All tested strategies lose to no memory; adaptive guidance partially mitigates failures.
2026-08-15	Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents
• Compares heterogeneous memory storage mechanisms under one evaluation harness. • Three models and four suites measure quality, efficiency, and scaling across substrates. • No substrate dominates; factual retrieval and sequential actions favor different choices.
2026-08-14	When Personal Memory Has No Single Answer: Evaluating LLM Agents under Irreducible Conflict
• Benchmarks personal-memory conflicts without a uniquely justified answer. • TANGLE tests conflict recognition, calibration, clarification, and faithful evidence handling. • Experiments expose lost conflict relations and motivate evidence-dependent action policies.
2026-08-11	MobileMem: Learning from a Year of Mobile Experiences
• Provides year-scale mobile experiences for long-term memory research. • Knowledge-grounded synthesis creates temporally coherent text and multimodal user trajectories. • The benchmark covers updates, implicit preferences, and multi-hop temporal reasoning.
2026-08-05	When Memory Lies: An Empirical Study of Spatial Memory Staleness in VLM Agents
• Studies stale spatial memory under conflicting visual observations. • Dynamic FrozenLake detection and navigation tasks isolate auditing and action-selection failures. • Filtering helps text settings, while unreliable visual grounding limits recovery.
2026-08-04	MAFIA: Query-Only Memory Attacks via Probing and Factual Injection against Audited LLM Agents
• Evaluates query-only poisoning against audited agent-memory systems. • Tests persistence under large benign stores and active input auditing. • Experiments expose remaining weaknesses in retrieval competition and semantic audits.
2026-07-30	Can an AI Assistant Really Forget? Auditable Deletion from Addressable Memory
• Tests auditable deletion from addressable conversational memory. • Removing exchange-linked rows is compared with refitting and never-observed histories. • Deletion reduces disclosure but remains distinguishable from rebuilding without the target exchange.
2026-07-29	MemSecBench: Tracking Agent Memory Poisoning from Persistence to Consequence and Repair
• Tracks memory poisoning from persistence through consequences to repair. • Write–execute–forget protocols cover 310 cases and 24 agent configurations. • Results show large configuration-dependent differences in attack propagation and selective repair.
2026-07-27	Keep It InMind: Benchmarking the Implicit-Association Blind Spot in Agent Memory
• Benchmarks implicit associations between stored facts and later tasks. • Paired controls distinguish missing knowledge from failures to surface decisive memory. • Systems recall explicit facts well but struggle when retrieval requires world-knowledge bridges.
2026-07-26	When Does Memory Help? A Cost-Aware Evaluation of Long-Term Memory in Tool-Using LLM Agents
• Measures the marginal utility and full cost of tool-agent memory. • Instrumented episodic tasks test dependence, updates, corruption, and actual action use. • Experiments favor update-aware stores in changing-fact tasks and expose retrieval–action gaps.
2026-07-24	Ground Truth First: A Longitudinal Evaluation Instrument for Agent Memory, and the Tenure Crossover in Memory-Architecture Rankings
• Generates longitudinal memory evaluation from ground truth before dialogue text. • Validity intervals, provenance, and verified rendering support controlled history-length comparisons. • Memory architecture rankings change with tenure; write quality strongly predicts answer quality.
2026-07-23	RUMBA: Russian User Memory Benchmark
• Introduces Russian conversational-memory evaluation with an aligned English subset. • Timestamped dialogues probe semantic type, session scope, and temporal reasoning. • Fine-grained evaluation identifies different strengths and failures across memory systems.
2026-07-23	Beyond Episodic Evaluation: Memory Architectural Bottlenecks in Sequential Embodied Question Answering
• Studies sequential embodied QA with memory retained across questions. • Comparisons isolate occupancy maps, inherited histories, and grounded 3D visual memory. • Simulation and robot tests favor spatially organized visual evidence for accuracy and navigation cost.
2026-07-18	RECON: Benchmarking Agent Memory for Compositional Reasoning over Long Contexts
• Benchmarks compositional reasoning over long, changing evidence histories. • Six tasks test invalidation chains, conflicts, counterfactuals, and temporal constraints. • Current systems struggle with both retrieval and reasoning beyond isolated fact updates.
2026-07-16	Bad Memory: Evaluating Prompt Injection Risks from Memory in Agentic Systems
• Studies persistent prompt injection through agent memory files. • Sandboxed multi-session tests compare two coding-agent systems across four models. • Already-planted memories influence later sessions, with persistence varying across configurations.
2026-07-14	PM-Bench: Evaluating Prospective Memory in LLM Agents
• Benchmarks execution of delayed intentions while other activities continue. • A simulated week tests cue monitoring, intention retention, and timely action. • Eight-model evaluation remains challenging, with no universally best memory strategy.
2026-07-14	MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations
• Evaluates explicit lifecycle operations rather than final memory answers alone. • Structured traces record triggers, targets, scope, transitions, and supporting evidence. • Tests expose update and trajectory-reconstruction failures hidden by answer accuracy.
2026-07-13	SLVMBench: Skill Learning from Video Memory
• Tests learning skills from long video memory for immediate application. • Tutorials embedded in hours of distractor video support temporally annotated task questions. • Current video models struggle to acquire and apply skills from long histories.
2026-07-02	MemSyco-Bench: Benchmarking Sycophancy in Agent Memory
• Identifies and formalizes "memory-induced sycophancy" (agents over-aligning with retrieved historical user memory at the cost of factual accuracy) and introduces MemSyco-Bench, which shifts evaluation from retrieval success to post-retrieval decision calibration. • A benchmark of five task categories — Objective Fact Judgment, Contextual Scope Control, Memory-Evidence Conflict, Valid Memory Selection, and Personalized Memory Use — built from simulated multi-turn dialogues; evaluates memory systems (Mem0, A-Mem, MemGPT, MemoryBank, LightMem, SuperMemory, NaiveRAG) across backbones. • Memory consistently increases sycophancy — DeepSeek-V4-Flash factual accuracy drops 56.1%→40.2% with sycophancy rate rising 24.3%→52.3%; memory systems degrade objective-fact accuracy (Qwen3-8B down to 26–36%); existing benchmarks' errors are dominated by retrieval failure (47.4%–66.1%).
2026-07-02	AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents
• Introduces a bounded-memory "contract" in which every decision prompt is freshly assembled by typed retrieval from five typed layers (L1–L5) with no raw cross-decision transcript appended, keeping the prompt bounded across arbitrarily long runs and making each memory layer individually ablatable. • Studies how explicit memory layers shape long-horizon LLM-agent decisions, instantiated in the stochastic deck-building game Slay the Spire 2; releases a reproducible testbed of 298 completed trajectories with condition tags, frozen memory/skill snapshots, prompt records, and analysis scripts. • In the fixed-A0 ablation the no-store baseline wins 3/10 games and adding the triggered strategic-skill layer (L5) wins 6/10 (directional, Fisher exact p≈0.37); a public frontier-LLM benchmark reports zero wins at A0 across five configurations versus a 16% developer-reported human win rate.
2026-06-25	DMV-Bench: Diagnosing Long-Horizon Multimodal Agents' Visual Memory with Incidental Cue Injection
• This paper targets a gap in agent-memory evaluation: most benchmarks test textual recall, while multimodal agents often need to remember visual details that were incidental when first observed. • DMV-Bench builds an interactive home-furnishing e-commerce environment with 1,000 product variants, text-leakage controls, multi-session shopping chains, and baked-in visual cues that later identify a previously seen product. • The proposed DualMem architecture keeps visual and verbal memory codes in parallel and outperforms caption-only and recent multimodal memory baselines across chain lengths on Gemini 2.5 Flash and Qwen2.5-VL-7B.
2026-06-23	MEMPROBE: Probing Long-Term Agent Memory via Hidden User-State Recovery
• The first benchmark to evaluate long-term agent memory directly as an auditable post-interaction artifact — reconstructing a user's hidden, taxonomy-anchored state from the memory an agent leaves behind — rather than only through indirect downstream task success. • Memory recovery benchmark; memory-equipped agents assist simulated users through leak-controlled tasks, after which each hidden user-state dimension is reconstructed and scored against ground truth under both full-store (dump_all) and top-k (retrieve) access modes. • Spanning 50 simulated users × 31 hidden dimensions (1,550 targets) and 5 memory systems, task completion nearly saturates (~99.9%) while category-balanced recovery stays moderate (B≈0.611–0.624 full-store, dropping to 0.473–0.540 under top-k), with episodic/relational memory the hardest.
2026-06-23	Are We Ready For An Agent-Native Memory System?
• Reframes LLM agent memory as a data-management system and proposes an analytical framework that decomposes it into four core modules (representation & storage, extraction, retrieval & routing, maintenance), paired with a structured taxonomy of existing systems. • Systematic experimental study benchmarking 12 representative memory systems plus 2 reference baselines on a unified testbed across 5 workloads / 11 datasets, evaluating task effectiveness, retrieval fidelity, dynamic-update robustness, long-horizon stability, and operational cost, plus per-module ablations. • No single architecture dominates — effectiveness depends on alignment between memory structure and workload bottleneck; graph-based methods handle updates most reliably while append-only stores return stale facts; highly structured systems incur orders-of-magnitude higher latency without proportional accuracy gains.
2026-06-23	MEMPROBE: Probing Long-Term Agent Memory via Hidden User-State Recovery
• This paper argues that long-term memory should be evaluated as an auditable post-interaction artifact, not only through downstream personalization quality, task success, or later answer accuracy. • MEMPROBE simulates assistance trajectories for users with hidden taxonomy-anchored state banks, then reconstructs those hidden states from the memory left by an agent after leak-controlled tasks. • The released benchmark covers 50 simulated users and 1,550 hidden user-state targets across five representative memory systems, revealing failures that remain invisible when task completion is already near saturation.
2026-06-23	OpenThoughts-Agent: Data Recipes for Agentic Models
• This paper studies how to curate training data for broadly capable agentic language models, addressing the gap left by open efforts that usually target a single benchmark and do not reveal generalizable data recipes. • The OpenThoughts-Agent project provides a fully open data curation pipeline and reports more than 100 controlled ablation experiments over pipeline stages, task sources, and data diversity. The resulting dataset contains 100K examples assembled for agentic model training. • Fine-tuning Qwen3-32B on this dataset yields 44.8% average accuracy across seven agentic benchmarks, improving by 3.9 percentage points over Nemotron-Terminal-32B. The training data also shows strong scaling behavior in compute-controlled comparisons, providing an open resource for studying how reusable agentic experience traces and task diversity shape agent training.
2026-06-17	LegalWorld: A Life-Cycle Interactive Environment for Legal Agents
• This paper introduces LegalWorld, which models Chinese civil litigation as a five-stage causal state chain encompassing legal consultation, preparation of initiating pleadings, first-instance proceedings, appeal, and second-instance proceedings. This design enables early factual representations, evidence organization, and litigation strategies to impose realistic constraints on subsequent procedures and adjudicative outcomes. • The environment is built on 75,309 paired first- and second-instance judgments and incorporates local in-scenario memory, global case memory, and modular libraries of skills and tools. Local memory maintains conversational continuity within an individual scenario, while global memory consolidates case facts, evidence, claims, and procedural status at the conclusion of each stage to ensure cross-stage consistency. • LongJud-Bench, constructed on this environment, evaluates agent capabilities across the complete litigation lifecycle. A total of 217 evaluators with legal backgrounds provided 18,992 ratings, validating the procedural realism and role consistency of the generated trajectories. Cross-model experiments further demonstrate substantial specialization across legal consultation, document drafting, and courtroom argumentation, indicating that a single aggregate score cannot reveal these stage-specific capability differences.
2026-06-17	GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents
• This paper introduces GateMem, extending memory evaluation from single-user private memory to multi-principal shared-memory settings. Agents must enforce context-dependent controls over the reading and disclosure of information in a shared memory pool according to the requester’s identity, role relationships, authorization scope, and deletion status. • The benchmark spans four domains—healthcare, workplace, education, and household settings—and comprises 91 long-form multiparty interaction scenarios with 2,218 hidden checkpoints. It jointly measures the practical utility of legitimate long-horizon requests, information leakage across authorization boundaries, and the reliability of active forgetting following explicit deletion requests. • Experiments reveal a pronounced tension among utility, access control, and forgetting. Long-context methods generally achieve stronger overall governance performance but incur substantial token overhead. Although retrieval-augmented and external-memory methods reduce computational costs, they may still disclose unauthorized information or recover deleted content, indicating that current systems remain inadequate for reliable institution-scale shared-memory deployment.
2026-06-15	MemTrace: Probing What Final Accuracy Misses in Long-Term Memory
• This paper introduces MemTrace, advocating user “knowledge points” with explicit types and session provenance as the unit of long-term memory evaluation, rather than aggregating mutually independent question instances into a single accuracy score. This formulation enables the persistent availability of the same fact to be traced across varying conditions. • The benchmark generates probes along three controlled dimensions: memory recency, question type, and evidence condition. These dimensions respectively assess the retention of facts as the number of sessions increases; the ability to answer questions about current states, historical states, and trajectories of change; and the capacity to respond safely when evidence is absent or a question contains a false premise. • MemTrace covers 20 users, 835 knowledge points, 15,422 question instances, and 13 memory-system configurations. Experiments show that systems with similar aggregate accuracy may exhibit fundamentally different failure structures. Moreover, when failures occur, relevant evidence is retrievable approximately ten times as often as it is absent, indicating that the primary bottleneck lies in the effective use of accessible evidence during reasoning rather than solely in storage capacity or retrieval recall.
2026-06-14	Control-Plane Placement Shapes Forgetting: An Architectural Study of Agent Memory Across Thirteen System Configurations
• This paper distinguishes between the recall plane, which retrieves existing facts, and the control plane, which governs memory overwriting, release, deletion, and updating. It investigates how the placement of large language models within these two planes systematically alters modes of forgetting failure. • The authors construct the ForgetEval evaluation suite, comprising 1,000 template-based test cases and 385 adversarial cases, and compare 13 system configurations across six architectural paradigms. The results show that deterministic rules, write-time LLM processing, knowledge-graph abstraction, and mutation-time hooks address distinct lexical, temporal, normalization-related, and intent-aware forgetting challenges, offering partially complementary capabilities. • Mutation-time LLM hooks increase the pass rate on 345 non-primitive existence cases from approximately 70% to over 93% and restore intent-aware deletion under prefix collisions and compound-fact conditions. The study therefore demonstrates that forgetting capability depends on where intelligent control is situated within the memory lifecycle, rather than merely on whether the system employs a large language model.
2026-06-12	StreamMemBench: Streaming Evaluation of Agent Memory for Future-Oriented Assistance
• This paper introduces StreamMemBench, extending the evaluation target of personal-agent memory from static factual recall to the complete streaming assistance pipeline. It examines whether an agent can identify and retain user-relevant evidence from continuous first-person observations and apply that evidence to both current and subsequent tasks. • The benchmark is constructed from EgoLife egocentric life streams and uses evidence anchors to connect initial tasks with follow-up tasks. Four metrics—Fidelity, Initial Evidence Use, Feedback Incorporation, and Follow-up Reuse—respectively diagnose evidence preservation, initial application, feedback assimilation, and experiential reuse. • The dataset comprises 8,107 evidence anchors and 16,214 task queries. Evaluation of eight classes of memory systems across two model backbones shows that storing information, or incorporating feedback within the current interaction, does not ensure that an agent can reliably invoke the relevant experience in future tasks. These findings indicate that memory evaluation should primarily assess its actual influence on subsequent behavior.
2026-06-09	Recalling Too Well: Sycophancy Evaluation and Mitigation in Memory-Augmented Models
• This paper shows that persistent memory systems in LLMs amplify sycophancy, causing models to prioritize user-aligned beliefs over factual correctness. • It introduces the MIST benchmark and evaluates multiple state-of-the-art memory systems and model families, systematically measuring this effect. • The authors propose lightweight mitigation strategies that significantly reduce sycophancy while maintaining or improving factual recall performance.
2026-06-08	H2HMem: A Multimodal Memory Benchmark for Agents in Human-Human Interactions
• This paper introduces H2HMem, a multimodal benchmark designed to evaluate agent memory in complex human-human interaction settings beyond single-user text conversations. • It includes dyadic and multi-party dialogues and evaluates memory across recall, reasoning, and application dimensions. • Results show that current systems struggle with cross-modal, cross-participant, and long-horizon memory construction and utilization, highlighting key limitations of existing agents.
2026-06-04	SubtleMemory: A Benchmark for Fine-Grained Relational Memory Discrimination in Long-Horizon AI Agents
• Addresses the fact that long-term memories may complement, diverge from, or contradict each other. • Constructs relation-controlled memory variants and embeds them into realistic long user-agent histories. • Reveals that current memory systems remain weak at fine-grained relational memory preservation and use.
2026-06-04	When Should Memory Stay Silent: Measuring Memory-Use Boundaries in Memory-Augmented Conversational Agents
• Studies when available sensitive memories should not be used in the current conversational turn. • Introduces RBI-Eval, a controlled probe set comparing model behavior with and without sensitive memory access. • Shows retrieval alone cannot ensure safety, motivating memory-aware decisions at both retrieval and generation time.
2026-06-03	MemoryBench: A Benchmark for Memory and Continual Learning in LLM Systems
• Targets the lack of benchmarks for LLM systems learning from accumulated service-time user feedback. • Builds a user-feedback simulation framework across multiple domains, languages, and task types. • Shows current baselines remain far from satisfactory in both effectiveness and efficiency.
2026-06-03	MemoryDocDataSet: A Benchmark for Joint Conversational Memory and Long Document Reasoning
• Combines two hard capabilities: navigating multi-session dialogue memory and reasoning over long documents. • Creates synthetic microworlds with characters, event timelines, long documents, conversations, and mixed-source questions. • Shows joint retrieval architectures better integrate conversational memory with document navigation.
2026-06-02	From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents
• Frames persistent memory as a long-lived attack surface where untrusted inputs can shape future behavior. • Identifies write channels, structural vulnerabilities, and memory poisoning attack categories, and introduces MPBench. • Shows active memory systems are especially vulnerable and that prompt-injection defenses are insufficient.
2026-06-02	PersistBench: When Should Long-Term Memories Be Forgotten by LLMs?
• Studies safety risks introduced by persistent long-term memory in conversational assistants. • Defines cross-domain leakage and memory-induced sycophancy as memory-specific failure modes. • Evaluates many frontier and open-source LLMs and reveals high failure rates on both risks.
2026-06-01	Beyond Static Dialogues: Benchmarking Realistic, Heterogeneous, and Evolving Long-Term Memory
• Addresses the static and homogeneous nature of many long-term dialogue memory benchmarks. • Constructs evolving user profiles, heterogeneous interactions, and challenging questions across memory types. • Reveals weaknesses in current systems under realistic, dynamic long-term memory settings.
2026-06-01	EGOSTREAM: A Diagnostic Benchmark for Streaming Episodic Memory in Egocentric Vision
• Targets streaming episodic memory for egocentric vision, where agents must remember over continuous input. • Builds diagnostic questions across multiple cognitive dimensions under a unified streaming MLLM framework. • Exposes key weaknesses of current memory management mechanisms in visual-agent settings.
2026-06-01	WorldMemArena: Evaluating Multimodal Agent Memory Through Action-World Interaction
• Argues that multimodal agent memory should be tested through interaction with dynamic worlds. • Builds a four-stage action-world loop and multi-session tasks to evaluate writing, maintenance, retrieval, and use. • Shows that better storage quality does not always translate into better downstream multimodal memory performance.
2026-05-31	Connecting the Dots: Benchmarking Reflective Memory in Long-Horizon Dialogue
• Targets reflective memory, where agents must synthesize fragmented cues rather than only recall explicit facts. • Introduces RefMem-Bench with annotated QA instances across reflective-memory dimensions and task formats. • Shows that question-guided evidence retrieval and abstraction-aware supervision improve reflective answering.
2026-05-30	Momento: Evaluating Persistent Memory and Reasoning with Multi-Session Agentic Conversations
• Addresses the gap between single-session agent benchmarks and realistic multi-session service interactions. • Creates tool-mediated tasks with persistent memory, temporal dependencies, changing user goals, and consequential actions. • Shows current agents often fail by misestimating user state across sessions.
2026-05-30	SuperMemory-VQA: An Egocentric Visual Question-Answering Benchmark for Long-Horizon Memory
• Targets real-world long-horizon visual memory for AI assistants using egocentric daily activity video. • Builds human-validated VQA pairs covering object, location, intention, and scene memory. • Shows current systems remain unreliable for practical everyday visual memory needs.
2026-05-29	Beyond Static Dialogues: Benchmarking Realistic, Heterogeneous, and Evolving Long-Term Memory
• Introduces RHELM, a benchmark that simulates one-year virtual life trajectories of 10 diverse personas (500K–1M tokens context per user) to evaluate LLMs' realistic, heterogeneous, and evolving long-term memory capabilities. • Proposes the LOOP module (pLan-rOllout-evOlve-Prune) to dynamically generate life trajectories and integrates heterogeneous external sources (emails, diaries, reports) alongside 11,764 dialogue turns, enabling high-fidelity multi-source memory evaluation. • Defines 7 question types and 27 challenge characteristics, including memory-conditioned misleading queries that test conflict-aware responses; shows that SOTA models like Claude Opus 4.5 only score ~38% on average, with cross-source aggregation and real-world contextual reasoning as the dominant bottlenecks.
2026-05-07	STALE: Can LLM Agents Know When Their Memories Are No Longer Valid?
• Introduces STALE, a long-context benchmark of 400 expert-validated implicit conflict scenarios to evaluate LLM agents' ability to detect when earlier memories are invalidated by new observations without explicit negation. • Proposes a three-dimensional probing framework that tests State Resolution, Premise Resistance, and Implicit Policy Adaptation, revealing that current models struggle to apply updated states in downstream behavior. • Presents CUPMEM, a prototype framework that performs explicit write-time state adjudication and topology-triggered belief propagation, significantly improving robust memory updating.
2026-05-03	MEMAUDIT: An Exact Package-Oracle Evaluation Protocol for Budgeted Long-Term LLM Memory Writing
• Proposes MEMAUDIT, an exact evaluation protocol that fixes experience streams, candidate memory representations, storage costs, evidence units, future queries, and budgets as an auditable optimization problem solvable by branch-and-bound and MILP. • Distinguishes memory representation quality, validity retention, and budget allocation effects, outperforming end-to-end QA accuracy as an evaluation approach.
2026-04-28	StratMem-Bench: Evaluating Strategic Memory Use in Virtual Character Conversation Beyond Factual Recall
• Introduces StratMem-Bench with 657 dialogue instances and mandatory/auxiliary/irrelevant memory pools to evaluate strategic memory use beyond factual recall in virtual character conversations. • Designs metrics for strict memory compliance, memory integration quality, proactive enrichment, and conditional irrelevance rate, systematically assessing memory decision-making in role-play.
2026-04-11	Trust Your Memory: Verifiable Control of Smart Homes through Reinforcement Learning with Multi-dimensional Rewards
• Introduces MemHomeLife and MemHome, a benchmark suite for memory-driven smart home control grounded in long-term real interaction logs. • Evaluates fine-grained memory operations including adding, updating, deleting, and using memories in device control. • Addresses a major gap in prior benchmarks by measuring both long-term memory utilization and feedback on memory management quality.
2026-04-09	Towards Real-world Human Behavior Simulation: Benchmarking Large Language Models on Long-horizon, Cross-scenario, Heterogeneous Behavior Traces
• Introduces OmniBehavior, the first benchmark built entirely from real-world data for long-horizon, cross-scenario, heterogeneous behavior simulation. • Highlights that real user decisions depend on cross-scenario causal chains that are largely missing from prior datasets. • Shows that current LLMs remain weak on complex behavior simulation even with longer contexts, and exhibit homogenized personalities and optimistic bias.
2026-04-09	KnowU-Bench: Towards Interactive, Proactive, and Personalized Mobile Agent Evaluation
• Introduces KnowU-Bench, an online benchmark for interactive, proactive, and personalized mobile agents in Android simulation. • Hides user profiles and provides only behavioral traces, requiring agents to infer preferences through interaction and decide when to intervene, ask consent, or stay silent. • Reveals large performance drops for frontier models when preference inference and intervention calibration are required.
2026-04-02	Memory in the LLM Era: Modular Architectures and Strategies in a Unified Framework
• Proposes a unified modular framework that decomposes LLM-based agent memory systems into four essential components: information extraction, memory management, memory storage, and information retrieval. • Conducts a comprehensive experimental evaluation and robustness analysis of representative agent memory methods on long-term conversational benchmarks, examining token efficiency, context scalability, and evidence position sensitivity. • Introduces a novel memory architecture that integrates tree-based organization with hierarchical storage to achieve state-of-the-art performance while maintaining low computational overhead.
2026-03-04	Towards Realistic Personalization: Evaluating Long-Horizon Preference Following in Personalized User-LLM Interactions
• Evaluates the blind spots of large language models in consistently following diverse user preferences under realistic and extremely long-horizon interaction settings. • Introduces the RealPref benchmark, covering 100 user profiles and large-scale interaction data, specifically designed to assess generalized understanding of preference expressions ranging from explicit to implicit. • Quantitatively demonstrates that increasingly implicit preference expression and excessively long contexts can cause cliff-like drops in LLM preference-following performance, thereby identifying a key bottleneck in the development of personalized perceptive assistants.
2026-03-04	LifeBench: A Benchmark for Long-Horizon Multi-Source Memory
• The evolution of long-term personal assistants requires models to perform complex reasoning based on non-declarative memory, such as inferring habits from digital traces, yet existing benchmarks offer virtually no coverage of this capability. • Introduces the LifeBench benchmark, which leverages real-world priors and hierarchical structures from cognitive science to simulate densely distributed multi-source events spanning long time horizons, in order to evaluate agents' integrative reasoning abilities. • Even state-of-the-art memory systems achieve only slightly above 50% accuracy (55.2%) on this benchmark, highlighting the extreme difficulty of long-horizon reasoning over fragmented digital traces.
2026-03-02	AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations
• Static offline evaluation data cannot faithfully capture the scalability and reliability of memory management strategies in dynamic, long-horizon interactions. • Builds AMemGym, an interactive environment that supports online policy evaluation, using structured sampling to generate high-fidelity user profiles and state-evolution trajectories at low cost. • The dynamic environment objectively reveals the long-horizon performance degradation of existing systems such as RAG, and validates the framework’s effectiveness in driving self-evolution of memory strategies.
2026-03-02	According to Me: Long-Term Personalized Referential Memory QA
• Traditional benchmarks are mostly based on single-text dialogue histories and therefore cannot effectively evaluate agents' personalized referential reasoning ability in real-life, multimodal, and long-horizon settings. • Introduces ATM-Bench, a multimodal and multi-source benchmark built from four years of private data, and proposes schema-guided memory for structured representation of heterogeneous data. • Reveals the performance limitations of existing systems on complex real-world experience collections, and shows that structured SGM outperforms traditional descriptive methods in multi-source scenarios.
2026-02-27	MemEmo: Evaluating Emotion in Memory Systems of Agents
• Existing evaluations of memory systems focus primarily on factual recall, while lacking measurement of how effectively emotional information is handled in long-horizon interactions. • Introduces the first emotion-enhanced evaluation benchmark and the HLME dataset, covering three dimensions: emotional information extraction, emotional memory updating, and emotion-aware question answering. • Experiments reveal significant instability across current state-of-the-art memory systems on emotion-related tasks, pointing the way toward future optimization of emotional coherence.
2026-02-18	MemoryArena: Benchmarking Agent Memory in Interdependent Multi-Session Agentic Tasks
• Introduces MemoryArena, a unified evaluation platform for benchmarking agent memory in multi-session tasks. • The benchmark covers tasks such as web navigation, planning, information search, and reasoning, requiring agents to continuously learn and utilize memory during execution. • The study reveals limitations of existing long-context memory benchmarks in evaluating real-world interactive scenarios and emphasizes the necessity of jointly evaluating memory and action.
2026-02-18	AgentLAB: Benchmarking LLM Agents against Long-Horizon Attacks
• Introduces AgentLAB, the first benchmark specifically designed to evaluate the vulnerability of LLM agents to adaptive long-horizon attacks. • Supports five novel attack types, including memory poisoning, spanning 28 real-world agent environments and 644 security test cases. • Experimental findings reveal that LLM agents are extremely vulnerable under long-term attacks, and that existing single-interaction defense mechanisms are insufficient to effectively address such complex security threats.
2026-02-11	Locomo-Plus: Beyond-Factual Cognitive Memory Evaluation Framework for LLM Agents
• It introduces the LoCoMo-Plus benchmark to evaluate "cognitive memory" in LLM agents, focusing on their ability to retain and apply implicit constraints like user goals and values across long conversational contexts. • It constructs scenarios with "cue–trigger semantic disconnect," requiring models to apply latent information even when downstream queries lack surface-level similarity to the original memory cues. • It establishes an evaluation framework based on "constraint consistency," revealing through extensive experiments that current LLMs and memory systems struggle significantly with complex cognitive memory tasks.
2026-01-29	AgentLongBench: A Controllable Long Benchmark For Long-Contexts Agents via Environment Rollouts
• Proposed AgentLongBench, evaluating long-context agents via simulated environment rollouts with context lengths ranging from 32K to 4M tokens. • Constructed dynamic interaction trajectories based on Lateral Thinking Puzzles, featuring both Knowledge-Intensive and Knowledge-Free settings. • Revealed that high information density in tool responses poses a greater challenge than memory fragmentation, identifying "minimum token requirement" as a key factor.
2026-01-24	MemoryRewardBench: Benchmarking Reward Models for Long-Term Memory Management in Large Language Models
• Proposed MemoryRewardBench, the first benchmark to systematically evaluate the ability of Reward Models (RMs) to assess long-term memory management. • Covers 10 settings across reasoning, dialogue, and generation tasks with contexts from 8K to 128K tokens. • Designed Outcome-based and Process-based evaluation criteria, revealing a generational advantage in newer models for evaluating memory capabilities.
2026-01-23	How Does Personalized Memory Shape LLM Behavior? Benchmarking Rational Preference Utilization in Personalized Assistants
• Proposed the problem of Rational Personalization and introduced the RPEval benchmark to evaluate appropriate user preference utilization. • Included a dataset for intent reasoning and a multi-granularity evaluation protocol, revealing widespread "irrational personalization" (e.g., Filter Bubbles). • Proposed RP-Reasoner, a pragmatics-based mechanism to infer latent user intent and selectively integrate personalized information, reducing errors significantly.
2026-01-13	Mem2ActBench: A Benchmark for Evaluating Long-Term Memory Utilization in Task-Oriented Autonomous Agents
• Proposed Mem2ActBench to evaluate agents' ability to actively leverage long-term memory for executing tool-based tasks, beyond passive retrieval. • Constructed a dataset of 2,029 sessions via automation and reverse-generated 400 memory-dependent tool-use tasks. • Experiments revealed that existing systems struggle with "Parameter Grounding," failing to extract correct parameters from memory for tool execution.
2026-01-11	CloneMem: Benchmarking Long-Term Memory for AI Clones
• Introduces CLONEMEM, a benchmark designed to evaluate long-term memory in AI Clones by utilizing non-conversational digital traces (e.g., diaries, social media, emails) spanning 1–3 years, rather than traditional conversational history. • Proposes a hierarchical data construction framework that generates coherent longitudinal life trajectories, capturing the evolution of an individual's experiences, emotions, and opinions over time. • Experimental results reveal that existing memory systems (such as A-Mem and Mem0) struggle with this setting, often underperforming compared to flat retrieval and failing to accurately track internal state changes due to lossy compression and reliance on narrative templates.
2026-01-11	RealMem: Benchmarking LLMs in Real-World Memory-Driven Interaction
• RealMem is a benchmark designed to evaluate LLMs in "long-term project-oriented" interactions, distinguishing itself from casual or task-oriented dialogue benchmarks by focusing on evolving goals and dynamic states. • The framework employs a three-stage synthesis pipeline (Project Foundation, Multi-Agent Generation, Memory Management) to create over 2,000 cross-session dialogues across 11 realistic scenarios. • Evaluations indicate that current state-of-the-art memory systems struggle with managing long-term project states, temporal reasoning, and proactive alignment, highlighting a critical gap in autonomous agent capabilities.
2026-01-08	KnowMe-Bench: Benchmarking Person Understanding for Lifelong Digital Companions
• Introduction: Introduces KnowMe-Bench, a benchmark derived from long-form autobiographical narratives (4.7M tokens) designed to evaluate "person understanding"—inferring stable motivations and principles—in lifelong digital companions, moving beyond simple fact retrieval. • Methodology: Utilizes a "Cognitive-Stream Reconstruction" pipeline to transform raw narratives into flashback-aware, time-anchored streams rich in inner thoughts and sensory details. It employs a 3-tier hierarchical evaluation suite ranging from factual extraction to psychoanalytic depth. • Findings: Experiments with various memory architectures (RAG, Mem0, MemOS) reveal that while retrieval systems improve factual accuracy, they struggle with temporal grounding and deep reasoning (e.g., the "Update Paradox"), highlighting a critical gap in modeling the complex, non-linear nature of human memory.
2026-01-07	Mem-Gallery: Benchmarking Multimodal Long-Term Conversational Memory for MLLM Agents
• Introduction: Introduces Mem-Gallery, a benchmark designed to evaluate multimodal long-term conversational memory in MLLM agents, bridging the gap between text-only long-context benchmarks and single-session multimodal tasks. • Methodology: Constructs a dataset of high-quality multi-session conversations grounded in both visual and textual information. It establishes an evaluation framework covering three functional dimensions: Memory Extraction & Adaptation, Memory Reasoning, and Memory Knowledge Management. • Findings: Extensive benchmarking across thirteen memory systems reveals that explicit multimodal information retention is crucial. However, current models struggle with memory reasoning, handling knowledge conflicts, and face significant efficiency bottlenecks.
2026-01-07	EvolMem: A Cognitive-Driven Benchmark for Multi-Session Dialogue Memory
• Introduction: Introduces EvolMem, a cognitive-driven benchmark designed to evaluate the multi-session memory capabilities of LLMs and agent systems, addressing the gap in assessing non-declarative memory and long-term consistency. • Methodology: Grounded in cognitive psychology, it categorizes memory into declarative (e.g., Retrieval, Inference) and non-declarative (e.g., Habituation) types. The benchmark uses a hybrid data synthesis framework—combining topic-initiated generation and narrative-inspired transformations—to create diverse, controllable multi-session dialogues. • Findings: Evaluations reveal that no LLM consistently excels across all memory dimensions, with significant weaknesses in non-declarative tasks. Furthermore, existing agent memory mechanisms often fail to outperform strong base models and suffer from high latency.
2025-12-07	PersonaMem-v2: Towards Personalized Intelligence via Learning Implicit User Personas and Agentic Memory
• Introduction: Introduces PersonaMem-v2, a state-of-the-art dataset featuring 1,000 realistic personas, 300+ scenarios, and 20,000+ implicit user preferences embedded in long-context interactions (up to 128k tokens). • Gap & Findings: Benchmarking reveals that frontier LLMs (including GPT-5) struggle with implicit personalization, achieving only 37-48% accuracy. However, Reinforcement Fine-Tuning (RFT) significantly improves reasoning capabilities for user understanding. • Methodology: Proposes an "Agentic Memory" framework that maintains a single, human-readable, and evolving memory. This approach outperforms GPT-5 with 55% accuracy while being 16× more token-efficient (using 2k memory vs. 32k history).
2025-11-09	FlashAdventure — Long-Term Memory for GUI Agents
• This paper introduces FlashAdventure, a benchmark for evaluating GUI agents on long-horizon, narrative-driven tasks requiring memory, planning, and interaction across complete story arcs. • It identifies the observation–behavior gap, where agents must recall earlier observations and apply them later when those memories become actionable. • The authors propose COAST (Clue-Oriented Agent for Sequential Tasks), an agent framework using long-term clue memory with a Seek-Map-Solve cycle to improve sequential decision-making. • The benchmark contains 34 Flash-based adventure games and evaluates agents through milestone-based automated scoring with CUA-as-a-Judge. • FlashAdventure demonstrates that explicit long-term memory enables GUI agents to better store, retrieve, and utilize historical information in long-horizon interactive environments.
2025-11-04	Toward Multi-Session Personalized Conversation: A Large-Scale Dataset and Hierarchical Tree Framework for Implicit Reasoning
• Introduces the IMPLEXCONV dataset and the TACITREE framework for studying implicit reasoning in personalized dialogue. • IMPLEXCONV consists of 2,500 examples focused on implicit reasoning scenarios, capturing subtle syntactic and semantic relationships within conversations. • TACITREE enhances large language models (LLMs)’ ability to perform implicit contextual reasoning in long dialogues by hierarchically organizing dialogue history.
2025-10-27	Know Me, Respond to Me, benchmarking LLMs for Dynamic User profiling and personalized response at scale
• Introduces the PERSONAMEM benchmark, designed to evaluate the performance of large language models (LLMs) in dynamic user profiling and personalized response generation. • Although existing models achieve some success in recalling user preferences, they still exhibit significant performance gaps when handling novel scenarios. • The paper provides a detailed description of the benchmark’s structure, the process for generating user dialogues, the methods for evaluating model performance, and related work, highlighting the importance of personalized dialogue generation in enhancing user experience.
2025-10-10	Human-inspired Episodic Memory for Infinite Context LLMs
• EM-LLM (Event Memory Large Language Model) is a novel large language model designed to address the limitations of existing models in handling long contexts. • EM-LLM achieves near-unlimited context processing without fine-tuning, delivering significant improvements over existing models across multiple benchmarks. • The model integrates surprise-based event segmentation, graph-theoretic boundary refinement, and a two-stage memory retrieval mechanism, substantially enhancing performance on information retrieval and question-answering tasks.
2023-09-26	Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions
• MemoryAgentBench is a benchmark designed to evaluate four core capabilities of language models with memory mechanisms (Memory Agents): accurate retrieval, test-time learning, long-range understanding, and conflict resolution. • By integrating existing datasets with newly constructed data, MemoryAgentBench enables a systematic evaluation of these capabilities. • The benchmark reveals limitations of current approaches in memory updating and long-horizon dialogue processing, highlighting key challenges for future research.
2025-07-27	MiniLongBench: The Low-cost Long Context Understanding Benchmark for Large Language Models
• MiniLongBench is a low-cost long-text understanding benchmark designed to improve the efficiency and affordability of evaluating large language models (LLMs) on long-context understanding (LCU). • By applying data compression techniques, MiniLongBench significantly reduces the number of evaluation samples while maintaining evaluation consistency, and shows a high correlation with results from the original LongBench benchmark. • Evaluations across multiple task categories demonstrate MiniLongBench’s effectiveness, although further improvements are still needed for summarization and synthesis tasks.
2025-07-27	PersonaBench: Evaluating AI Models on Understanding Personal Information through Accessing (Synthetic) Private User Data
• PersonaBench is a benchmark designed to evaluate AI models’ ability to understand personal information. • The paper highlights the importance of personalization in AI assistants and emphasizes the challenge posed by the lack of publicly available datasets for assessing such capabilities. • The evaluation primarily focuses on retrieval-augmented generation (RAG) models, with results indicating that current models still struggle to effectively handle personal queries.
2025-07-27	MemBench: Towards More Comprehensive Evaluation on the Memory of LLM-based Agents
• MemBench is designed to comprehensively evaluate the memory capabilities of LLM-based agents. • By constructing datasets that cover both factual memory and reflective memory, the study addresses limitations of existing evaluation approaches. • The paper provides a detailed description of memory mechanism construction—including user relationship graphs and multi-layer memory designs—and emphasizes the importance of evaluation metrics such as accuracy, efficiency, and capacity.
2025-07-27	Evaluating the Long-term memory of large language models
• This paper investigates the memory capabilities of large language models (LLMs) in long-term tasks, with a particular focus on dialogue systems. • By constructing the Long-Order Chronological Conversation (LOCCO) dataset, the study provides a quantitative evaluation of LLMs’ long-term memory performance. • Experimental results show that while LLMs can retain historical conversational information to some extent, their memory gradually degrades over time.
2025-07-27	Know You First and Be You Better: Modeling Human-Like User Simulators via Implicit Profiles
• Introduces a user simulation framework, the Implicit User Profile User Simulator (USP), designed to enhance interactions between dialogue systems and human users by inferring implicit user attributes. • USP extracts implicit features from user dialogues and combines conditionally supervised fine-tuning with reinforcement learning under cycle consistency, improving the realism and coherence of generated conversations. • Experimental results show that USP achieves significant advantages across multiple metrics, particularly when compared with other dialogue generation models such as GPT-4o and PlatoLM.
2025-07-27	Unveiling Privacy Risks in LLM Agent Memory
• Investigates privacy vulnerabilities in LLM agent memory, specifically the risk of extracting private user-agent interactions stored in long-term memory. • Proposes the Memory EXTRaction Attack (MEXTRA), a black-box attack leveraging a novel prompt design (Locator + Aligner) and automated prompt generation to extract sensitive user queries. • Experiments on representative agents (EHRAgent and RAP) demonstrate significant vulnerability, analyzing key factors like similarity scoring functions and memory configurations that influence leakage.
2025-06-15	PersonaFeedback: A Large-scale Human-annotated Benchmark For Personalization
• Proposes the PersonaFeedback benchmark for evaluating large language models (LLMs) in personalized response generation. • The study shows that while LLMs have made progress in producing personalized content, they still face limitations in complex scenarios. • By leveraging dynamic user attribute inference, personalized profiles, and reward models, the researchers aim to improve the effectiveness of personalized question answering.
2025-06-09	Minerva: A Programmable memory test benchmark for language models
• Minerva is a programmable memory testing benchmark designed to evaluate large language models (LLMs) across diverse memory tasks. • It quantitatively assesses models’ ability to use memory, with a particular focus on tasks such as information retrieval, reasoning, and state tracking. • Experimental results indicate that while some models perform well on simple tasks, there remain substantial gaps on more complex tasks.
2025-05-28	Self-Taught Agentic Long-Context Understanding
• The Agentic Long-Context Understanding (AgenticLU) framework is designed to enhance large language models (LLMs) in long-text understanding and reasoning. • AgenticLU introduces a Chain-of-Clarifications (CoC) mechanism that optimizes the model’s self-clarification process and employs tree-structured search paths to generate clarification questions, thereby significantly improving the accuracy and effectiveness of multi-step reasoning. • Evaluation results show that the framework outperforms existing prompting techniques on long-context question answering, while keeping computational overhead well controlled.
2025-05-22	EMBODIED AGENTS MEET PERSONALIZATION: INVESTIGATING CHALLENGES AND SOLUTIONS THROUGH THE LENS OF MEMORY UTILIZATION
• The paper investigates the challenges LLM-powered embodied agents face in personalized assistance, specifically focusing on memory utilization for object semantics and user behavioral patterns. • It introduces MEMENTO, a two-stage evaluation framework, which reveals that current agents struggle with sequential user patterns and coordinating multiple memories due to information overload. • The study proposes a hierarchical knowledge graph-based user-profile memory module that separates personalized knowledge from episodic history, achieving substantial improvements in both single and joint-memory tasks.
2025-03-11	SCBench: A Benchmark for Long Context Methods Based on KV-Cache
• SCBENCH (Shared Context BENCH) is a benchmark specifically designed to evaluate long-context large language models (LLMs). • SCBENCH focuses on the lifecycle of key–value (KV) caches, covering generation, compression, retrieval, and loading, and aims to fill a gap in existing benchmarks regarding KV cache evaluation in multi-turn interactions. • Experimental results reveal substantial differences across methods on different tasks, and show that dynamic sparse attention and cache optimization strategies deliver superior performance in complex scenarios.
2022-03-04	LongMemEval: Benchmarking chat assistants on long-term interactive memory
• The paper introduces LONGMEMEVAL, a comprehensive benchmark for evaluating the long-term memory capabilities of chat assistants. • The benchmark assesses five core memory abilities, capturing key challenges faced by existing systems. • LONGMEMEVAL adopts a unified three-stage framework—indexing, retrieval, and reading—and proposes several design optimizations to improve memory recall and question-answering accuracy.
2025-02-25	Towards Effective Evaluations and Comparisons for LLM Unlearning Methods
• Explores machine unlearning in large language models (LLMs) and the importance of its evaluation, with a particular focus on eliminating unnecessary data memorization. • The study addresses two key challenges: the robustness of evaluation metrics and the trade-off between removing target knowledge and preserving other knowledge. • It recommends Extraction Strength (ES) as a primary evaluation metric to ensure accuracy and robustness in unlearning assessment.
2022-02-13	DO LLMS RECOGNIZE YOUR PREFERENCES? EVAL-UATING PERSONALIZED PREFERENCE FOLLOWING IN LLMS
• PREFEVAL is a benchmark designed to evaluate large language models (LLMs) in their ability to infer, remember, and follow user preferences over long conversations. • The benchmark includes 3,000 user preference–query pairs spanning 20 topics, revealing significant challenges for current LLMs in adhering to user preferences. • The study shows that explicit preferences are easier for models to infer than implicit preferences, and that task types and preference expression styles have a substantial impact on model performance.
2015-01-25	Episodic Memory Benchmark: Episodic Memories Generation and Evaluation Benchmark for Large Language Models
• Explores the importance of episodic memory in large language models (LLMs) and proposes the construction of new benchmarks to evaluate models’ reasoning capabilities. • The authors develop a comprehensive framework with newly designed tasks and evaluation protocols, emphasizing the need for novel training strategies to effectively integrate episodic memory. • The framework provides a promising solution for evaluating episodic memory in LLMs.
2024-12-24	Muse: A Multimodal Conversational Recommendation Dataset with Scenario-Grounded User Profiles
• Introduces Muse, the first multimodal conversational recommendation dataset, containing 83,148 utterances across 7,000 conversations in the Clothing domain with scenario-grounded user profiles. • Proposes a three-stage multi-agent generation pipeline (User Profile Generator → Simulated Dialogue Generator → Dialogue Optimizer) powered by multimodal LLMs, where user profiles are derived from real-world scenarios rather than manual design, improving scalability. • Validates quality against four baselines (MMCONV, Redial, Inspired, PEARL) across five dimensions (naturalness, coherence, informativeness, product-context relevance, image-text consistency) and demonstrates utility by fine-tuning three MLLMs for recommendation and response generation.
2025-01-23	LongGenBench: Benchmarking long-form generation in long context LLMs
• LongGenBench is a benchmark designed to evaluate large language models (LLMs) in generating high-quality long-form text, with a particular emphasis on following complex instructions. • Unlike existing benchmarks, LongGenBench focuses specifically on long-text generation scenarios, covering tasks such as diary writing and menu design. • Despite strong performance on other evaluations, LLMs face significant challenges on the LongGenBench benchmark.
2015-01-03	LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks
• LongBench v2 is a multi-task benchmark for evaluating large language models (LLMs) in long-context understanding and reasoning. • It consists of 503 multiple-choice questions spanning diverse task types, with a focus on comprehending and answering long-form text. • The study finds that the best-performing models surpass human experts in long-context tasks, highlighting the importance of enhanced reasoning and increased test-time computation.
2024-11-12	MT-Eval: A Multi-Turn Capabilities Evaluation Benchmark for Large Language Models
• MT-Eval is a benchmark designed to evaluate the performance of large language models (LLMs) in multi-turn conversations. • While existing evaluations primarily focus on single-turn dialogue, MT-Eval fills this gap by constructing 1,170 multi-turn queries. • The benchmark categorizes interaction patterns into recall, expansion, refinement, and follow-up, revealing that most models perform consistently worse in multi-turn settings than in single-turn scenarios.
2024-11-12	LONGGENBENCH: Long-context Generation Benchmark
• LongGenBench is a newly proposed long-context generation benchmark designed to evaluate the performance of large language models (LLMs) on long-form text generation tasks. • It complements existing benchmarks that primarily focus on retrieval skills by emphasizing coherence and logical consistency across multiple sub-questions. • The study shows that different models exhibit substantial performance disparities in long-text generation.
2024-10-23	MADial-Bench Towards real-world evaluation of memory-augmented diglogue generation
• MADial-Bench is a benchmark for evaluating memory-augmented dialogue generation, targeting the limitations of current dialogue systems in long-term memory. • MADial-Bench incorporates concepts from cognitive science to assess memory retrieval and recognition, and introduces multi-dimensional evaluation metrics. • The study shows that while large language models (LLMs) perform well in emotional support, their memory recognition and injection capabilities still require improvement.
2024-10-04	L-CiteEval: A Long-Context Citation Evaluation Benchmark
• L-CiteEval is a multi-task evaluation benchmark for long-context models (LCMs), designed to measure their abilities in understanding and citation. • The benchmark covers 11 tasks, supports context lengths ranging from 8K to 48K, and provides a comprehensive evaluation framework. • The study shows that closed-source models outperform open-source models in citation quality and generation accuracy, while retrieval-augmented generation (RAG) techniques effectively improve citation quality.
2024-08-16	A personal long-term memory dataset for memory classification,Retrieval, and Synthesis in question Answering
• PerLTQA is a question-answering dataset designed to enhance long-term memory integration in dialogue systems. • PerLTQA combines semantic memory and episodic memory, containing 8,593 questions across 30 personas, with the goal of improving memory classification, retrieval, and synthesis. • Experiments show that BERT-based models outperform other large language models in memory classification tasks.
2024-08-11	CAN LONG-CONTEXT LANGUAGE MODELS UNDER-STAND LONG CONTEXTS
• Explores the capabilities and limitations of large language models (LLMs) in long-text processing, and introduces the GLE benchmark for evaluating LLMs’ performance in long-context understanding. • The paper describes the construction process and evaluation criteria of long-dependency question-answering tasks, and compares the performance of different models. • Experimental results show that the GLE benchmark effectively assesses LLMs’ ability to process long-form text.
2024-08-11	Evaluating Very Long-Term Conversational Memory of LLM Agents
• Evaluates the memory capabilities of large language models (LLMs) in long-duration conversations, with a particular focus on multimodal dialogue scenarios. • By developing the LOCOMO dataset, the researchers establish a comprehensive evaluation benchmark covering tasks such as question answering, event summarization, and multimodal dialogue generation. • Experimental results indicate that while some LLMs perform strongly, they still lag significantly behind humans in memory and reasoning, and the study outlines an evaluation framework and directions for future improvements.
2024-08-11	Lamp: When large language models meet personalization
• Explores the importance of large language models (LLMs) in personalized response generation and introduces LaMP, a new benchmark specifically designed for training and evaluating personalized text generation and classification tasks. • LaMP comprises seven personalized subtasks, highlighting the effectiveness of leveraging user-specific inputs (e.g., historical data) and retrieval-augmented strategies to enhance language model performance. • Experimental results demonstrate that personalization methods significantly improve model performance, with the best results achieved through fine-tuning and the use of appropriate retrieval strategies.
2024-06-19	LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding
• LongBench is a bilingual, multi-task benchmark designed to evaluate large language models (LLMs) on long-context understanding. • The benchmark comprises 21 datasets spanning six task categories: single-document QA, multi-document QA, summarization, few-shot learning, synthetic tasks, and code completion, with an average length of 6,711 words and 13,386 characters. • Experimental results show that commercial models (e.g., GPT-3.5-Turbo-16k) generally outperform open-source models on long-context tasks.
2024-04-16	HIERARCHICAL CONTEXT MERGING: BETTER LONG CONTEXT UNDERSTANDING FOR PRE-TRAINED LLMS
• HOMER (Hierarchical cOntext MERging) is an algorithm designed to address the limitations of large language models (LLMs) in handling long contexts. • By splitting long inputs into smaller chunks and hierarchically merging them, HOMER improves both memory efficiency and reasoning capability when processing long-form text. • Experimental results show that HOMER delivers strong performance with 32K and 64K context inputs, maintaining low perplexity and reduced memory consumption.

Systems & Models

Date	Paper & Summary	Tags	Links
2026-11-03	SEAM: Situated Executable Associative Memory for a Co-located Community
• Presents SEAM, organizing community recollections as associative and executable shared memory. • Preserves source posts while revising a knowledge wiki; separate retrieval layers resolve relevant knowledge and actionable standing rules. • A two-week deployment with 46 people illustrates discovery and rule use, offering observational design evidence rather than controlled performance superiority.
2026-11-03	Spatial Memory for AR Glasses: Hierarchical Visual Acquisition and Source-Linked Recall
• Presents an AR spatial-memory architecture from visual acquisition to source-linked recall. • Selects RGB captures by quality, novelty and temporal support, separating interpretation status and maintaining source dependencies. • A 131.35-second replay selects six RGB views from 648 observations; frequent feature extraction remains costly and long-term QA gains are unestablished.
2026-11-02	Scalaris: A Scalable Enterprise Memory Architecture for Conversational AI Assistants
• Introduces Scalaris with distinct representations for user information and assistant responses. • Separates evolving cross-session user knowledge from selectively retrieved assistant content using specialized ingestion and retrieval pipelines. • Qwen3-32B scores 92.2% and 90.77% on two benchmarks with roughly tenfold fewer context tokens than the comparison; proprietary-assistant gains are internal reports.
2026-11-02	El Agente Fármaco: A Scalable AI Scientist for Drug Discovery
• Presents a drug-discovery agent centered on shared scientific memory in Fármaco Atlas. • Preserves computational records and provenance, retrieves targeted evidence for decisions and executes calculations over molecule collections. • Reports system gains on 22 tasks and a retrospective comparison; computational and model-judged scores do not establish experimental efficacy.
2026-11-01	Learning by Doing: Autonomously Compiling and Curating Agent Skills from Verified Traces
• Presents Zep Agent Skills for compiling and maintaining procedural memory from verified traces. • Reflector and Curator roles preserve provenance and verification strength, gate versions and update retrieval ranks from subsequent outcomes. • Three executor models gain 4.5–18.8 percentage points; most gains occur on tasks solved by the teacher.
2026-10-30	Dynamical Memory Transformer: Selective Recurrent Memories with Constant Cost Attention
• Presents the Dynamical Memory Transformer, separating sequence encoding from information reading. • Complex diagonal dynamical units update through input-dependent leak rates, with attention reading their states rather than token history. • Selective-copy length extrapolation improves, but language modeling remains behind Mamba, exposing retention and addressability trade-offs.
2026-10-30	Learning to Live: A Step Toward Lifelong Embodied Intelligence
• Presents LIFE to connect continuous observation, selective memory and capability reuse. • Plastic Memory preserves images and structured fields for later re-indexing; EDM registers validated methods as reusable tools or skills. • Improves evidence retention at matched write budgets; the 24-hour robot-video study is a limited replay case, not long-term autonomous deployment validation.
2026-10-30	Neurons Are All You Need
• Introduces NURESA with persistent memory registers inside neural units. • M accumulates activation and gates responses; thresholded consolidation populates C, and M+C is injected into a fresh same-architecture model. • Reports transfer with GPT-2 and Qwen2.5 among others; the tested protocols do not establish arbitrary cross-architecture transfer or privacy guarantees.
2026-10-28	ParaTTT: Parallelizing Test-Time Training Layers with Deep Memories
• ParaTTT parallelizes deep nonlinear memory by refining estimates of online update trajectories. • Three parallel passes refine feature and gradient estimates, with convergence proved under sufficient conditions. • 340M and 1.3B models show competitive language results and better recall than linear recurrences; two iterations yield 4.9× full-model training speedup at 340M.
2026-10-28	Dual DeltaNet: Value-Side Memory Reorganization for Linear Attention
• Dual DeltaNet reorganizes memory from the value side to complement key-side correction. • A two-sided delta rule updates associations and value-feature address organization, with chunkwise parallel training. • It reports stronger overall language, downstream and retrieval results than tested alternatives, with overhead relative to KDA.
2026-10-27	V-LUMEN: Visual Lookup Memory for Embedding Scaling in Vision-Language Models
• V-LUMEN extends lookup-memory capacity to vision-language models. • A frozen codebook, spatial hashing and text-conditioned addressing retrieve embeddings through gated residual fusion. • Reports gains on five reasoning benchmarks with frozen Qwen3-VL-2B and transfer to three image-disjoint benchmarks; capacity still requires storage.
2026-10-26	MarkovSim: an LLM Agent Harness with Bounded Context for Long-Horizon Computational Simulation
• Presents MarkovSim, an LLM simulation harness with memory-based bounded context management. • Stores workflow records in a predefined schema and reconstructs compact progress context at finite-state-machine transitions. • Passes 16 of 18 runs on six finite-element problems while keeping context bounded; human consultation remains for engineering judgments.
2026-10-22	Delta N-gram: Forward-Only Continual Writing into a Language Model's Native N-gram Memory
• Writes factual updates into sparse hash-addressed model memory. • A small forward writer patches a frozen backbone and is evaluated on facts and document updates. • Recalls 16/16 in a small test versus 2–4/16; broader gains coexist with general-capability drift.
2026-10-20	ClawMMU: Protection-Coherent Virtual Memory for Stateful LLM Agents
• Applies ownership and authorization controls to agent memory. • Scope and origin checks govern reads, writes and downstream actions. • The abstract presents an access-control design without quantitative evidence of broad effectiveness.
2026-10-20	Industrial ReMEmbR: Scaling Spatiotemporal Memory to Robot Fleets
• Provides shared semantic memory for industrial robot fleets. • Queries by content, position and time and generates descriptions on demand. • Tests 450 decodes for a specified equivalence property; masking alone does not establish comprehensive privacy.
2026-10-20	UniWALM: Unified World Action Language Model for Embodied Robot Control
• Unifies language, video and action through a language-as-memory loop. • A shared model learns recurrent language state without a pretrained language backbone. • Reports 34.5 versus 28.5 on RMBench and 11.1→61.1 on a real object-finding task.
2026-10-19	MemT-1.0: Towards Memory-Native Evolution
• Builds language models around a memory-native recurrent storage design. • Combines Gated DeltaNet and mixture-of-storage mechanisms with continued training and post-training. • At 128K, Nano reports 1.86× speed and 66× storage reduction; these are configuration-specific results.
2026-10-19	Kronecker Delta Rule for Long Context Linear Attention
• Revises matrix-valued recurrent memory with a Kronecker delta rule. • Structured updates control interference while retaining a compact state. • Combines language experiments with stability analysis under explicit update and stochastic-process assumptions.
2026-10-19	Magnolia: A Persistent-Memory Agent Framework for Computational Chemistry
• Structures computational-chemistry agent memory by validation scope. • Session records become confirmed project facts and human-validated cross-project rules. • 317 runs yield 116 entries and four rules; four controlled replays provide limited evidence of reuse.
2026-10-18	Mixture-of-Memories: Low-Overhead Language Modeling with Modular Knowledge Injection
• Combines latent and lexical memory experts around a frozen language model. • Budgeted routing and answer gating support modular knowledge injection and updates. • Evaluated on modeling, QA and update efficiency; routing and gating remain reliability dependencies.
2026-10-16	An Interactive GraphRAG Workspace for Reifying, Auditing, and Steering Conversational AI Memories
• Makes conversational coding memory editable and auditable. • A hierarchical fact graph supports traversal, temporal-drift checks and visual manipulation compiled into prompts. • Reports 92.2% token reduction; the abstract establishes a tool design rather than broad user or task effectiveness.
2026-10-15	Multimodal Filesystem Memory For Long-Horizon Embodied Agents
• Links multimodal agent experience through filesystem memory. • Textual indices connect original episodes with spatial records, consolidated knowledge, procedures and current plans. • Reports 16/24 Ender Dragon milestones and 0.440 KiCad reward with one backbone; evidence is task- and model-bounded.
2026-10-15	Hy-Memory-30B-A3B: Attributed Reinforcement Learning for Long-Term Memory Writing
• Hy-Memory-30B-A3B specializes in cross-framework long-term memory writing, targeting retention, precision and updates. • On-policy distillation precedes RL with chunk probes, cross-chunk attribution, over-storage controls and GRPO, carrying written banks across windows. • Leads four harnesses on fixed 20% benchmark subsamples with a 55.49% macro average versus 54.62% for Claude Opus 5; this is not full-benchmark evidence.
2026-10-13	Titan-Llama: Retrofitting Pre-trained LLMs with External Neural Memory Modules
• Titan-Llama retrofits pretrained Transformers with hierarchical neural memory combining recent attention and historical retention. • Blockwise attention preserves recent tokens while older context enters recurrent modules updated at test time, without training from scratch. • Against segmented attention, QuALITY rises from 23.87% to 40.51% and LongHealth from 14.50% to 40.75%; speed gains apply to reported configurations.
2026-10-13	Depth Memory
• Depth Memory formulates cross-layer representation mixing as storage and retrieval. • Linear Depth Memory uses recursive least squares with associative-memory and inverse-covariance streams alongside the residual stream. • At 400M and 1.5B scales it lowers perplexity relative to PreLN, mHC and AttnRes with competitive zero-shot accuracy; larger-scale behavior remains untested.
2026-10-08	STAPLE: Static Sparse Upcycling of Pretrained Checkpoints via Embedding Modules
• STAPLE adds token-indexed static sparse memory to pretrained language models. • Preserves existing feed-forward computation while adding a residual embedding pathway for scalable token-specific information. • OLMo-2 1B gains 4.27, 2.80, 2.27 and 1.65 points on HumanEval, MBPP, GSM8K and DROP, with gains at 7B; serving results depend on implementation.
2026-10-08	Trainable Quantum-Native Language Models: Dissipative Memory, Capacity Limits, and Hardware Validation
• Studies how dissipative gating supports memory in quantum byte-level language models. • Token-conditioned channels update a quantum register; Born-rule predictions and key-value recall tests probe storage and capacity. • Gating beats one-token controls but not two-token controls, and three stored pairs remain unsolved; models are classically simulable and trail n-gram baselines.
2026-10-07	YANchor-4B: Effective Long-Horizon Reasoning in O(N) Time with O(1) Memory
• YANchor-4B retains distant details in independent anchor memory alongside recurrent state and local attention. • Adapting Qwen3.5-4B, it learns record writes, selection and retrieval while keeping historical state capacity fixed as generation grows. • Mean AIME 2024–2026 pass@1 is 82.93% and HMMT is 63.64%; throughput claims depend on the stated H100 and batching settings.
2026-10-06	PERSIST: Who-What-When Memory Across Sessions for Full-Duplex Spoken Dialogue
• PERSIST models who, what and when in persistent spoken-dialogue memory, separating historical states from later revisions. • Readable event records combine semantic, speaker and temporal retrieval; backbone reuse reduces retrieval latency from 578.42 to 7.03 ms. • On the introduced SpokenTrace benchmark, end-to-end accuracy is 85.08% and all-support EM@3 is 82.10%; evidence is benchmark-specific.
2026-10-05	Voltic: Distinguishing Volatility from Stochasticity in Recurrent Memory
• Distinguishes changing associations from noisy observations when deciding how strongly to overwrite recurrent memory. • Uses anisotropic uncertainty with diagonal or quasi-diagonal approximations compatible with delta-rule kernels. • Leads controlled recall and an eight-task reasoning average, with improved extrapolated retrieval; language experiments use 45M-parameter models.
2026-10-05	HLA: Expressive Hybrid Linear Attention via Chunk-Wise Dynamic Mixing
• Adds query-dependent access to historical chunks in Gated DeltaNet recurrent memory. • Represents chunks as affine state transitions and routes them with content-dependent gates and effective-support regularization. • Gains reach 5.57 points on LongBench-V2 and 3.97 on RULER for Qwen3.5; from-scratch tests support extrapolation beyond training length.
2026-10-04	GitSwarm: Decentralized Compounding Inference
• Uses branchable shared memory to preserve partial solutions, experiments, and failures across independent inference episodes. • Atomic Git commits and semantic dependencies let agents verify, extend, and combine prior contributions. • Reports 79.4% on ProgramBench and subsequent reuse of 94.7% of contributions; findings are tied to the stated runs and budgets.
2026-10-04	AgentDiscover: Autonomous Discovery with Minimal Search Scaffolding
• Stores ideas, candidates, experiments, and their relations in queryable long-term memory for autonomous scientific search. • A server maintains the database; agents can use, combine, or replace queries corresponding to classical search rules. • Reports favorable cost efficiency and matches or exceeds same-model baselines on eleven tasks under the tested budgets.
2026-10-04	Universal Test-Time Training
• Shares fast-weight memory across both depth and time instead of assigning a private memory to each layer. • Provides routed shared-expert and dense variants so later chunks can reuse writes made at different depths. • At matched state and compute, 124M and 760M language models gain 2.6 and 2.1 RULER points over layer-private counterparts.
2026-10-02	Gated Slot Attention-2: Two-Sided Associative Memory Correction in Linear Attention
• Corrects both sides of associative memory through shared latent slots in recurrent language models. • Combines key-side Gated Oja Rule-2 with value-side Gated Delta Rule-2 and chunkwise training. • Approximately 1.3B-parameter models improve language modeling and retrieval over tested linear-attention baselines, with some constant throughput overhead.
2026-10-02	Continual Graph Memory for Mathematical Research Agents
• Organizes mathematical facts, plans, counterexamples, and failed explorations in continual cross-problem graph memory. • Dependency-aware retrieval and an evidence-sensitive curator update the research frontier; scoped recall prompts local re-proving of prior findings. • The authors report closure on ten First Proof Second Batch tasks and four component studies; open-problem solutions remain author-reported claims.
2026-10-02	Beyond the Parameter Monolith: Reconstructive Memories, Executable Skills, and Residual Assembly for Language Models
• Separates language-model computation from reconstructive document memories and deterministic execution. • Stores 52,809 versioned memory elements, addresses them with provenance-controlled indices, and reconciles typed proposals through residual assembly. • Reconstruction reaches about 75% token accuracy; general capability, whole-store utility, and end-to-end efficiency remain unestablished.
2026-10-01	The Surprising Effectiveness of Shared Memory in Looped Transformers
• Trains looped language models to share historical memory across recursions. • Only the first recursion writes shared KV; later recursions read it alongside a short local window. • Five-recursion hybrid models reduce perplexity by 1.12–1.82 and context memory by 76–79%; analyses find complementary shared and local representations.
2026-10-01	MemFit: Efficient Long-Term Agentic Memory
• Keeps verbatim conversational turns in an append-only store indexed by segment summaries. • Uses LLM-free insertion and lexical-semantic retrieval with cross-encoder reranking of caption-augmented episodes. • Reports leading results on LoCoMo, MemGallery and LongMemEval-S with several-fold construction savings; the abstract omits per-metric numbers.
2026-10-01	OneStreamer: Unifying Perception, Memory, and Proactive Response in Streaming Video Interaction
• Learns query-independent, time-grounded detail and event memories alongside live video response. • Hierarchical captions supplement a recent visual window without rereading old features; state-transition training reduces repetitive waiting supervision. • Leads compared methods on eight benchmarks, with caption ablations supporting historical QA gains; training combines synthetic and cleaned data.
2026-10-01	Mem++: Non-Destructive Memory for Long-Term Organizational LLM Agents
• Retains whole dated, authored organizational documents and earlier versions instead of fixing facts at write time. • Mem++ uses generation-free writes and time-filtered lexical-semantic retrieval, leaving version interpretation to the reader. • Outperforms the strongest memory baseline by 8.0–13.1 points on OrgMemBench; conversational rankings vary by variant and reader.
2026-10-01	VISTA: A Visual Harness for Reasoning in an Interactive World
• Preserves original visual observations for active recall during interactive multimodal reasoning. • VISTA lets a general-purpose model retrieve past observations and reorganize its visual input as it reasons. • Completes all 25 public ARC-AGI-3 games with 57.4% fewer actions than first-time humans; evidence covers these public games and three additional benchmarks.
2026-09-30	MemCodex: Self-Programming Hierarchical Memory for Language Agents
• Evolves executable programs for summary, relational, skill and latent memory layers. • Rewrites construction, indexing, retrieval and routing; coarse-to-fine reads stop at sufficient evidence or descend to original history. • Reports 10.1% higher average success than the strongest adaptive-memory baseline with 3.4× fewer context tokens and 2.1× faster inference in MemArena.
2026-09-28	Mnemon: Raw Records, Fast Judgments, Slow Thoughts
• Keeps raw dated conversations and separates fast evidence judgments from slower search planning and answering. • An LLM plans searches, Jev judges relevance and currency, and budget rules assemble a small view; background timelines remain linked to records. • Reports 91.7% LoCoMo and 83.8% LongMemEval-S with gpt-4.1-mini; results depend on the answering backbone and are author-reported.
2026-09-28	Stashbird: Efficient Speaker-Indexed Memory for Conversational Agents
• Links conversational memory state to source episodes and speakers. • Stashbird maintains episodic records, semantic relations, community summaries, and deletion provenance. • Reduces ingestion and retrieval workload substantially, with benchmark-dependent accuracy trade-offs.
2026-09-28	NavHarness: Towards Lifelong Embodied Navigation
• Carries navigation experience across fresh reasoning sessions. • NavHarness checks evolving maps and records against observations and consolidates recovery handovers. • Improves GOAT-Bench task success over independent sessions by 18.6 and 22.6 points for two models.
2026-09-28	Coding Agent Memory Post-training: Unlocking the Memory Potential of Pre-trained File Operations for Long-Horizon Tasks via Reinforcement Learning
• Post-trains agents to use ordinary filesystem operations as memory. • CAMG supplies four long-horizon environments and persistent workspaces; asynchronous PPO learns memory behavior. • The 4B and 9B models are competitive with larger Qwen baselines on tested coding benchmarks.
2026-09-28	Where Memory Belongs: Ledger, an Object Ledger for Memory-Augmented VLAs
• Separates short perceptual memory from explicit long-term object memory. • Ledger combines an in-policy frame memory with an external object record and planner. • Achieves 64.3% across four RoboMME suites versus 45.9% for the strongest compared prior method.
2026-09-28	FactorEngram: Factorized N-gram Memory with Basis-Level Gating for Language Models
• Factorizes learned n-gram memory into shared components. • Sparse coefficients and basis-level contextual gates support selective memory readout. • Improves language modeling on 340M and 1B backbones; ablations favor middle-layer insertion before attention.
2026-09-28	MS-GLA: Multi-Scale Gated Linear Attention for Addressing Representational Bottlenecks via Multi-Temporal Resolution
• Expands effective recurrent memory through multiple temporal resolutions. • MS-GLA pools different token spans across heads and learns input-dependent fusion. • Reports up to 18.9% recall improvement and lower perplexity at matched parameter counts.
2026-09-27	LSTMem: Hierarchical Long Short-Term Online Memory for Large Language Models
• Separates history accumulation from memory expression inside frozen LLMs. • LSTMem uses gated cell and hidden matrices with cross-layer propagation and reconstruction feedback. • Improves three memory/reasoning benchmarks on Qwen3-4B; cross-layer propagation contributes to gains.
2026-09-26	SCLATE: a Substrate for Continual-Learning Agent Training and Evaluation
• Provides a shared execution substrate for continual-learning agents and benchmarks. • SCLATE schedules sessions and consolidation events while preserving native harnesses and memory. • Compares seven benchmarks and ten configurations; added memory is not uniformly better, while post-training improves usage.
2026-09-26	Memory as a cache: Exact context reuse and deletion by construction
• Makes model context exactly composable and deletable by construction. • SMem independently encodes blocks and reads their union through cross-attention. • Improves edit and reuse efficiency with recipe-dependent perplexity trade-offs; very long deletion tests probe cost rather than served quality.
2026-09-25	MACBT: A Multi-Agent Cognitive Behavioral Therapy Decision Support System with Longitudinal Memory
• Adds longitudinal memory to a multi-agent CBT support workflow. • Tracks cognitive patterns and intervention history across sessions using a fine-tuned backbone. • Reports improved continuity and session quality on simulated dialogue with model judges; clinical effectiveness is not established.
2026-09-25	PIA: A Personal Intelligence Agent Turning Health Conversations into Records and Records into Understanding
• Builds typed persistent health records from conversational requests. • PIA separates extraction, storage, retrieval, and longitudinal understanding through pluggable domain controls. • Reports operational lessons rather than controlled clinical effectiveness; self-reports and causal links remain noisy.
2026-09-25	Memory as Middleware for Self-Improving AI Agents
• Argues for agent memory as a pluggable middleware layer. • ALTK-Evolve illustrates isolation, consistent writes, provenance-aware sharing, and lifecycle governance. • Presents a reference implementation and research agenda rather than a comparative performance claim.
2026-09-24	ADM-Planner: LLM-Guided Long-Horizon Planning for Mobile Manipulators with Attention-Enhanced Dynamic Memory
• ADM-Planner maintains selective dynamic memory for long-horizon mobile manipulation. • Persistent workspace knowledge is separated from object state, and relevant updates trigger retrieval and replanning. • In the reported noisy simulator setting, success reaches 100% versus 62% with static memory; the live LLM pilot has only six episodes.
2026-09-23	Memory Attention
• Memory Attention constructs attention values using token-indexed memory plus contextual keys. • Layer-specific lookup tables supply reusable token representations and support normalization folding and CPU prefetching. • Matched-token training reports better language modeling and downstream averages, but introduces additional memory parameters.
2026-09-23	Agent Memory with Episodic Retrieval for Financial Decision-Making
• META introduces episodic retrieval into a multi-agent financial decision system. • Market-state embeddings index past outcomes and reflections, which inform indicator weighting and final decisions. • The authors report stronger short-horizon accuracy and robustness; the abstract supplies no numerical results or long-horizon validation.
2026-09-22	AkasicMEM: Governed Enterprise Memory for Agents
• AkasicMEM targets persistent enterprise memory shared across tasks and agents. • Transitive lineage, policy composition at write time and policy reevaluation at retrieval preserve authorization across derived memories. • The architecture runs on a unified vector–graph–relational database; the abstract does not report a quantitative benchmark.
2026-09-22	Latest Exact Match Attention
• LEMA uses binary keys and queries to retrieve the latest exact match from a growing internal state. • A surrogate training scheme enables non-differentiable matching, while dictionary-based inference provides context-length-independent per-token lookup. • Models up to 834M parameters outperform fixed-state GDN on some recall tasks but remain behind softmax Transformers on loss and tested natural-text retrieval.
2026-09-22	MemoryAthena: Adaptive Routing over Latent and Generated Memories
• MemoryAthena selectively combines direct learned-memory retrieval with generated latent memories. • A lightweight causal router learns counterfactual future-token advantages and bounds interpolation around the direct retrieval pathway. • Five-task QA averages rise from 37.65 to 39.28; the complete memory-side system uses about 201M parameters beyond the frozen backbone.
2026-09-21	Jev-Mem: System-One-Controlled Agentic Memory for Efficient AI Agents
• Jev-Mem separates lightweight memory control from deliberative language-model reasoning. • A System-One controller manages typing, relations, query routing, retrieval budgets and stopping over a structured memory plane. • The authors report a 0.777 LoCoMo judge score, 158 s construction time and 0.93 s average query latency in their setup.
2026-09-20	RPMem: Learning Long-Term Recurrent Parametric Memory Across Sessions for LLM Agents
• RPMem maintains evolving cross-session memory that can survive a language-model backbone replacement. • Session compilation produces backbone-independent latent memory, a recurrent gate consolidates it, and a mapper generates backbone-specific LoRA parameters. • Across three benchmarks and five backbones, Qwen3-8B reaches 85.52% on PERMA with near-constant memory and update cost.
2026-09-17	CoLearn: An Agentic Tutor that Learns its Learner in a Human--AI Co-Learning Loop
• CoLearn maintains an evidence-grounded learner memory for personalized tutoring. • Soft-evidence Bayesian knowledge tracing updates topic mastery and misconceptions, which guide future questions and visible progress feedback. • Memory-conditioned questions receive 68–69% preference in blind comparisons; mastery convergence is evaluated in persona simulations.
2026-09-16	Agora: Git as Shared Memory for Collective AutoResearch
• Presents Agora, which uses a Git commit DAG as durable shared memory for collective autonomous research. • Stores reproducible results, hypotheses, insights, and verification records as contributions, with a derived frontier index guiding further work. • A 12-day, 13-worker weight-transfer case study demonstrates sustained collaboration; without a controlled comparison, it does not isolate gains in discovery efficiency.
2026-09-15	Persistent Recurrent Memory Between Transformer Layers - Improves Language Model Generalization
• A persistent recurrent state bridges the lower and upper halves of a decoder-only Transformer. • Cross-attention reads hidden representations, a GRU updates state, and gated addition modulates subsequent computation. • A six-layer model adds 3.7% parameters and reduces held-out loss by 28.5% across five seeds; evidence is small-scale.
2026-09-14	Semantic-TVM: Structure-Preserving Trustworthy Virtual Memory for Memory-Augmented and Tool-Using Agents
• Semantic-TVM protects exact private values in memory-augmented and tool-using agent workflows. • A trusted local runtime preserves executable values while exposing recoverable handles or selectively masked spans to remote models. • On Memory-EHR, span-level protection reports 84.17% task success versus 52.33% for whole-field replacement with DeepSeek, alongside low measured exposure.
2026-09-14	MoME: Mixture-of-Memory Embeddings for Context-Aware Sparse Lookup
• Gives each token multiple context-dependent memory embedding slots. • A learned hidden-state gate selects which slots to read. • Controlled pretraining reports gains over fixed lookup baselines at matched budgets.
2026-09-11	AIM: A Privacy-Aware Interoperable Memory Framework for Multi-Agent Multi-User LLM Systems
• Manages private and shared memory across agents and users. • Visibility classification and index-level controls enforce owner-scoped retrieval. • MUMBench evaluates retrieval, creation, updating, and deletion with mixed operation accuracy.
2026-09-11	RunningTensor: Generalizing Linear Attention to Higher-Order Recurrent States
• Generalizes matrix recurrent memory to higher-order tensors. • Rank-one updates and contracted reads retain recurrent and parallel computation. • Order-three experiments improve associative recall and pretrained retrieval tasks.
2026-09-09	Kernel-Managed Shared Memory for System-Wide Personalization
• Centralizes shared personalization memory in the agent-system kernel. • Agents write tagged records while the kernel handles retrieval, privacy, and injection. • AIOS trials report stronger personalization and shorter prompts across three models.
2026-09-07	Kalman Delta Networks: Uncertainty-aware Associative Memory
• Tracks uncertainty when updating recurrent associative memory. • Kalman-inspired diagonal and isotropic approximations support parallel associative scans. • Controlled 750M and 1.3B pretraining improves perplexity and average downstream accuracy.
2026-09-03	Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations
• Scales latent n-gram memory independently of backbone width. • Discrete routing, grouped-query readout, and selective null retrieval control activation cost. • VLM experiments report efficient scaling and semantically interpretable discrete addresses.
2026-08-30	Agent Zero Memory: Provenance-Aware Long-Term Memory for LLM Agents
• Combines episodic, associative, and documentary long-term memory. • Parallel searches use provenance tracking and citation-locked evidence access. • LongMemEval and LoCoMo experiments report strong accuracy across multiple backbones.
2026-08-30	Diachronic Hypergraphs for Orchestrated Multi-Agent Multimodal Memory Curation
• Stores collaborative multimodal events in a temporal hypergraph memory. • Role-aware retrieval and lifecycle updates preserve agents, tools, decisions, and evidence together. • Experiments report improvements over compared multi-agent memory baselines.
2026-08-27	Fast Weight Attention for Continual Learning
• Formulates recurrent fast-memory updates as temporally aligned online learning. • Derives normalized regression and inner-product updates with bounded rehearsal. • Representative variants remain competitive in language modeling and improve arithmetic length extrapolation.
2026-08-24	Wontopos Tablet 2: Measuring Multilingual and Multimodal Memory Retrieval Without Lexical Matching
• Retrieval path with no lexical matching, no keyword scoring and no language model of its own • Measures a production memory engine on LongMemEval-S and BEAM-1M, and on cross-lingual retrieval of photographs stored with no text • 95.2% mean recall@5 over 70 store-and-query language cells where BM25 reaches 19.0% and is exactly zero in 54; changing only the reader moves LongMemEval-S by 2.0 points and only the re-ask budget moves BEAM-1M by 8.9
2026-08-18	MoNe: Modular Neural Memory for Efficient Long Context Inference
• Attaches modular fast-weight memory to frozen Transformers. • Segment-wise test-time updates produce query-conditioned keys and values without rereading context. • At 128K tokens, authors report roughly 80% lower compute and peak GPU memory.
2026-08-17	Proteus: Incremental Memory Activation for Long-Context Sequence Modeling
• Expands effective memory capacity progressively as context grows. • Early compression and later capacity activation reduce interference between stored information. • Experiments improve several neural-memory architectures, especially on longer contexts.
2026-08-12	LoKiFormer: Locality-aware Attention with Decoupled Knowledge Memory for Efficient Large Language Model Pretraining
• Separates local sequence processing from addressable global knowledge memory. • Combines convolution-enhanced attention with a parametric key-value knowledge module. • Authors report 1.33-fold faster pretraining convergence than compared baselines.
2026-08-12	MindMemOS: A Portable and Self-Evolving Memory Operating Layer for AI Agents
• Provides a portable, self-evolving agent memory operating layer. • Validation-driven schema search, consolidation, corrective feedback, and skill evolution adapt memory. • LoCoMo, PersonaMem, and SpreadsheetBench evaluations report memory and skill gains.
2026-08-12	MARCH: Scaling Recurrent Memory with Content-Routed State Anchors
• Extends recurrent memory with content-routed historical state anchors. • Caches state checkpoints and retrieves them through learned anchor queries. • Pretrained models improve reasoning, LongBench, and retrieval over linear-attention baselines.
2026-08-10	MixFormer: Linear Transformer with Mixture of Memory Experts
• Combines multiple memory experts in a linear Transformer. • Time-aware attention uses learned decay and positional biases to update differentiated states. • Long-sequence text and image-generation experiments report improved performance.
2026-08-03	RING: Retrieval-Internalized Generation for Continual Large-Scale Knowledge Injection
• Internalizes retrieval into trainable memory experts for continual knowledge injection. • Pretraining, supervised search traces, and reinforcement learning optimize internal routing. • News-2025 evaluations report competitive accuracy and efficiency against RAG baselines.
2026-08-03	Learning What to Remember: Test-Time Training via Context Distillation
• Trains fast weights to preserve context useful for future predictions. • A long-window teacher supervises hidden states of a short-window student. • Experiments improve long-context modeling and enable inference-time adaptation of pretrained Transformers.
2026-08-03	DART: Decoded Attention over Recurrent States for Efficient Long-Context Sequence Modeling
• Uses recurrent chunk states as an attention-readable memory. • Decodes query-conditioned keys and values from Mamba-2 states with gated fusion. • Experiments improve associative recall while reducing cache relative to matched attention.
2026-08-03	Mamba with Hierarchical Memory: Solving Representation Bottleneck in Long Sequence Modeling
• Adds hierarchical working and long-term memory to pretrained Mamba. • Extracts paragraph semantics from hidden states for persistent task-relevant retrieval. • Passkey and LongBench-E tests improve retrieval and reasoning with 2% additional parameters.
2026-08-03	LiveMem: Maintaining Memory State Continuity in Long-Running LLM Inference
• Maintains intrinsic memory across repeated context-window turnover. • Combines bounded KV attention, memory-oriented post-training, and state-aware serving. • LongMemEval tests show recall after evidence leaves the active window.
2026-08-03	Structured Memory for Edge Language Models: Persistent Context and Corpus Retrieval via O(1) SSM State Injection
• Injects precomputed SSM states for retrieval and persistent context. • Hierarchical consolidation combines episodic, semantic, and retrieved corpus states. • Edge-model experiments report comparable RAG quality with sharply reduced prefill latency.
2026-08-03	Maglev: Sliding Recurrent Memory
• Adds fixed-size recurrent memory to sliding-window Transformers. • A stronger prefiller supplies memory-consistency targets for a recurrent decoder. • Experiments improve validation loss and downstream benchmarks; only the decoder serves inference.
2026-07-31	TransMem: Transforming Hidden States into Memory for Large Language Models
• Reuses sparse historical hidden states as latent agent memory. • A gated module learns from evidence-conditioned distillation over a frozen backbone. • LoCoMo, HotpotQA, and MemoryAgentBench experiments report consistent gains.
2026-07-30	Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory
• Scales separately pretrained parametric memory to 6.9B parameters. • Distributed indexing and sparse retrieval support training on 300B tokens. • Authors report better parameter efficiency across general and domain benchmarks.
2026-07-29	Metis: Memory Foundation Model
• Embeds persistent, evolving memory within a foundation model. • Memory attention and dedicated mid-training learn gradient-free state updates. • Experiments characterize native memory capabilities; model checkpoints are released.
2026-07-28	Raven: High-Recall Sequence Modeling with Sparse Memory Routing
• Maintains fixed memory slots with sparse input-dependent updates. • Learned routing selects which slots decay and receive new information. • Recall benchmarks remain strong at up to sixteen times training context length.
2026-07-28	MemSFT: Mitigating Alignment Tax with an External Parametric Memory
• Separates domain specialization from backbone updates using parametric memory. • A learned router combines memory and backbone distributions during generation. • Biology, geoscience, and law tests improve specialization with little general-capability loss.
2026-07-23	Naju: A Native Discrete State-Space Model with Independent Retention and Writing for Long-Sequence Memory
• Decouples retention and writing in a discrete state-space memory. • Independent forget and write gates remove the complementary-gate constraint. • Diagnostics show strong retention and overwriting at four times training length.
2026-07-23	MemTools: A Unified Research Framework for Interoperable Agent Memory
• Standardizes interoperable components for agent-memory research. • Declarative contracts separate lifecycle modules, datasets, protocols, and representations. • Integration experiments demonstrate controlled comparisons across heterogeneous memory systems.
2026-07-13	Extending LLM Context via Associative Recurrent Memory
• Extends pretrained LLM context through associative recurrent memory. • Combines continued pretraining, synthetic curricula, and selective memory-layer integration. • Experiments generalize beyond native windows and report 30% fewer FLOPs within them.
2026-07-08	Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity
• Expands linear RNN memory capacity through sparse addressing. • Replaces dense Gated DeltaNet updates with sparse memory reads and writes. • Matched-compute experiments improve retrieval and in-context learning; learned initialization adds knowledge gains.
2026-07-08	TF-Engram: A Train-Free Engram with SSD-Backed Memory for Large Language Models
• Builds train-free phrase memory across GPU, DRAM, and SSD storage. • Offline semantic construction and predictive prefetching support hidden-state injection. • Qwen3-0.6B tests improve downstream scores while reducing GPU storage demand.
2026-07-08	Co-LMLM: Continuous-Query Limited Memory Language Models
• Externalizes factual knowledge through continuously queried memory during pretraining. • Vector keys retrieve attributable textual values from an annotated knowledge base. • Multi-scale experiments improve perplexity and factual precision over compared LMLMs and LLMs.
2026-07-07	Do You Remember? Toward Memory-Centric Multimodal AI
• Studies reconstructive visual memory in multimodal language models. • Sparse local moving-average updates maintain a shared visual memory matrix. • Facial-image experiments improve reconstruction; findings are limited to the studied setting.
2026-06-24	Bridging the Post-discharge Gap: A Traceable Multi-agent Framework for Safe and Continuous Care
• This paper addresses post-discharge follow-up, where safe continuous care requires agents to reason over longitudinal patient-specific constraints rather than answer isolated medical questions. • Healink is a memory-enhanced multi-agent framework with triage routing, unified memory, and prescription-grounded response generation, designed to keep follow-up advice traceable to patient history and care instructions. • Retrospective and physician-blinded evaluations report improved completeness and perceived clinical utility, making the work relevant as a domain-specific example of memory-governed agent deployment.
2026-06-23	TRUSTMEM: Learning Trustworthy Memory Consolidation for LLM Agents with Long-Term Memory
• This paper studies memory consolidation for long-term LLM agents, where write, revise, and delete operations can omit important evidence, corrupt existing memories, or persist hallucinated content as future system state. • TrustMem introduces a Memory Transition Verifier that judges each memory update by coverage, preservation, and faithfulness, then constructs preference pairs among candidate transitions under the same memory state to train a preference-guided consolidation policy. • Experiments on MemoryAgentBench, HaluMem, and Mem-α show that TrustMem improves both memory utility and reliability, indicating that memory agents need transition-level verification before generated updates are committed to persistent memory.
2026-06-23	Governed Shared Memory for Multi-Agent LLM Systems
• This paper frames shared memory for fleets of LLM agents as a governed distributed-systems problem involving scoped access, temporal correctness, provenance, synchronization, and policy-controlled propagation. • The proposed architecture defines primitives for scoped retrieval, temporal supersession, provenance tracking, and policy-governed memory propagation, and instantiates them in the production multi-tenant service MemClaw. • ArgusFleet live evaluation reconstructs all tested depth-four provenance chains and surfaces practical enforcement issues, showing why long-context retrieval alone is insufficient for production multi-agent memory.
2026-06-23	Qwen-AgentWorld: Language World Models for General Agents
• This paper introduces Qwen-AgentWorld, a family of language world models that predict agentic environment dynamics from observations and actions, positioning world modeling as a core mechanism for reasoning, planning, and scalable agent training. • The models are trained from more than 10 million real environment interaction trajectories across seven agentic domains. The training pipeline combines continued pretraining for world-modeling capabilities, supervised fine-tuning for next-state prediction, and reinforcement learning with hybrid rubric-and-rule rewards to improve simulation fidelity. • Qwen-AgentWorld supports two complementary uses: as a decoupled simulator for scalable agentic reinforcement learning, and as a unified agent foundation model whose world-model training improves downstream performance across seven agentic benchmarks. The paper also introduces AgentWorldBench for evaluating language world models from real-world interactions.
2026-06-18	AtomMem: Building Simple and Effective Memory System for LLM Agents via Atomic Facts
• This paper introduces AtomMem, which adopts high-value atomic facts as the fundamental representational units of long-term memory. It seeks to balance the redundancy and overhead associated with preserving complete conversations against the loss of detail caused by excessive summarization, while mitigating the content drift and instability introduced by unconstrained memory updates. • The system first employs a Fact Executor to selectively extract independent, verifiable, and information-dense facts from extended interactions, subsequently organizing them into hierarchical event structures and temporal profiles. The former preserve event-level episodic relationships, while the latter track dynamic changes in user preferences, goals, and attributes over time. • During retrieval, an associative memory graph connects facts distributed across different events and time points to support cross-session, multi-hop, and temporal reasoning. Experiments on LoCoMo show that AtomMem achieves leading results across multiple reasoning categories, demonstrating that atomic representations must be integrated with event hierarchies, temporal evolution, and associative retrieval; merely partitioning memory into shorter text segments is insufficient to establish stable long-term memory.
2026-06-17	OpenRath: Session-Centered Runtime State for Agent Systems
• This paper addresses fragmented runtime state in modern agent systems, where transcripts, tool effects, memory events, workspace placement, branch provenance, and replay evidence are often recorded separately and become difficult to inspect or reproduce. • OpenRath introduces a PyTorch-like programming model whose central abstraction is a first-class Session value passed between agents and workflows. A Session is branchable, inspectable, replayable, backend-aware, and composable, carrying conversation chunks, sandbox placement, lineage metadata, token usage, pending work, tool evidence, and the runtime record of memory interactions. • By making fork, merge, and replay explicit runtime operations, and by defining Sandbox, Tool, Agent, Memory, Workflow, and Selector components, OpenRath provides an auditable composition model for multi-agent, multi-session systems. The report focuses on controlled runtime properties, leaving broader live-provider quality and memory-quality evaluation to follow-on work.
2026-06-16	FinAcumen: Financial Multimodal Reasoning via Self-Evolving Experience Memory Harness
• This paper introduces FinAcumen, which accumulates financial multimodal reasoning trajectories across tasks, distills successful cases into reusable strategies, and converts failed trajectories into risk-warning rules. The framework addresses the tendency of stateless tool-using agents to repeatedly explore the same strategies, invoke tools incorrectly, and generate financial hallucinations. • During inference, the system employs a selective experience-activation mechanism that injects retrieved experience into the reasoning context only when its semantic relevance to the current problem exceeds a calibrated threshold. When an experience is unreliable or insufficiently relevant, a fallback pathway explicitly suppresses its influence. A deterministic financial tool environment is then used to perform numerical calculations, information retrieval, chart interpretation, and answer verification. • Across four financial multimodal reasoning benchmarks, FinAcumen consistently improves a frozen 8B vision-language model and, under certain settings, outperforms finance-specialized models while approaching leading proprietary general-purpose models. Further analysis shows that these performance gains depend on the selective use of experience rather than the unconditional expansion of memory injection.
2026-06-16	ActWorld: From Explorable to Interactive World Model via Action-Aware Memory
• This paper observes that existing interactive video world models generally employ history-compression strategies biased toward recent frames, which may discard critical event-transition frames that determine subsequent object states. This produces an “action-forgetting” problem in which a model can navigate but fails to retain the consequences of actions such as opening doors, picking up objects, or triggering physical changes. • ActWorld introduces hierarchical action-aware memory into a chunk-wise autoregressive generation framework. Historical segments are assigned different compression pathways according to their importance for object interactions and state changes, while a persistent memory bank retains event-update tokens and object-identity tokens over the long term to preserve causal continuity across extended rollouts. • The authors also construct a dataset containing approximately 100,000 human–computer interaction videos with chunk-level reasoning annotations. Experiments show that the model supports both free-viewpoint navigation and intermediate object interaction within a unified system, improving interaction fidelity without compromising viewpoint controllability. These findings indicate that long-term history management in world models must be governed by action causality rather than temporal recency alone.
2026-06-16	PreAct: Computer-Using Agents that Get Faster on Repeated Tasks
• This paper introduces PreAct, which compiles the trajectory of a computer-using agent’s first successful task completion into a compact state-machine program. States verify current screen conditions, while transitions execute actions such as clicking and text entry, transforming repeated tasks from instances of renewed reasoning into executable procedural-memory replay. • Replay is not performed unconditionally: before executing each step, the system verifies whether the screen matches the program’s expected state and returns control to the general-purpose agent upon detecting an interface deviation. A new program is written to the long-term program repository only after it has been replayed in a clean environment and an independent evaluator has confirmed complete task execution. • Across mobile, desktop, and web benchmarks, direct program replay achieves an 8.5- to 13-fold speedup without requiring step-by-step LLM calls, while write-time validation prevents defective programs from accumulating. The experiments further show that the system’s gains arise primarily from reliable program validation and reuse rather than prompt adjustments, runtime safeguards, or more sophisticated program retrievers.
2026-06-16	CoreMem: Riemannian Retrieval and Fisher-Guided Distillation for Long-Term Memory in Dialogue Agents
• This paper introduces CoreMem, an edge–cloud long-term memory architecture unified through information geometry. It is designed to address simultaneously the hubness problem in high-dimensional vector retrieval, semantic loss during context compression, and the constraints on GPU memory and computational resources imposed by consumer-grade devices. • At the retrieval stage, CoreMem replaces isotropic cosine similarity with a locally adaptive Fisher–Rao Riemannian metric and uses a low-rank Woodbury acceleration scheme to enable real-time search. At the compression stage, Fisher-guided discrete token distillation progressively selects content from the sentence level to the token level according to parameter sensitivity, while explicitly preserving grammatical and structural information. • Experiments on LoCoMo and LongMemEval-S report improvements of 4.51 percentage points in open-domain reasoning and 4.17 percentage points in temporal reasoning, while operating within a strict 8 GB GPU-memory budget. The principal significance of this work lies in providing a unified theoretical foundation for retrieval and compression, although its efficiency claims remain dependent on the adopted low-rank approximations and edge–cloud deployment configuration.
2026-06-15	MemSlides: A Hierarchical Memory Driven Agent Framework for Personalized Slide Generation with Multi-turn Local Revision
• This paper proposes a hierarchical memory architecture for personalized presentation generation that explicitly separates cross-task long-term memory from working memory within an individual editing session. It further divides long-term memory into user-profile memory and tool memory to accommodate distinct temporal scales and functional objectives. • User-profile memory retrieves stable preferences according to task intent and provides personalized conditioning for initial generation. Working memory continuously maintains newly introduced constraints and revision requirements during the current session. Tool memory stores reusable execution experience for localized editing and is combined with a scope-constrained slide-modification mechanism to reduce unnecessary regeneration of unrelated slides and elements. • Controlled experiments show that user-profile memory improves persona consistency across multiple roles and intents, while tool memory enhances the execution reliability of closed-loop revisions. Case analyses also validate the role of working memory in preserving preferences across turns. These findings indicate that effective personalized generation requires memory to be functionally stratified by persistence and responsibility rather than consolidated into a single unified memory pool.
2026-06-13	Edu-Theater: A Data-Efficient Agent Framework for Scalable Learner Behavior Simulation through Staging Roll-Call
• To address the strong dependence of conventional learner simulation on extensive individual interaction histories, this paper proposes a cohort-aware “roll-call” simulation paradigm. It first constructs a cohort-level ability prior from the logs of similar learners and then calibrates an individual’s knowledge state through a small number of targeted diagnostic interactions. • Edu-Theater employs a teacher agent to coordinate the overall simulation process and dynamically switches between cohort-level priors and individual-level evidence through retrospective roll-call probing of historical learning logs. This design transforms historical logs from complete per-user contexts into shared cohort memory and sparse individual states, alleviating cold-start challenges and reducing invocation costs in large-scale deployment. • Experiments on two real-world educational datasets show that the framework achieves higher accuracy in simulating future behavior with substantially fewer LLM calls, while its generated data also improve downstream tasks such as adaptive testing. Its primary contribution lies in enhancing the efficiency of state reconstruction for educational agents rather than establishing a general-purpose personal long-term memory system.
2026-06-13	T-Mem: Memory That Anticipates, Not Archives
• This paper observes that mainstream long-term conversational memory systems generally rely on lexical or embedding similarity and can therefore handle only descriptive recall in which queries exhibit explicit surface-level overlap with historical content. Such systems struggle to identify associative memories connected through causal, narrative, or latent semantic relationships. • Inspired by episodic future thinking, T-Mem proactively generates triggers such as Entity, Bridge, Scene, and Horizon during memory writing. It simultaneously establishes descriptive and associative access pathways at both the factual and full-conversation granularities, enabling historical information to be reactivated by potentially relevant future contexts rather than by its original wording. • Experiments show that T-Mem achieves leading results on both LoCoMo and the cognitively associative LoCoMo-Plus, reducing the performance gap between the two benchmarks to 5.45 percentage points. Ablation studies further demonstrate that scene-level associative triggers are particularly important for LoCoMo-Plus, indicating that future-oriented write-time indexing is the principal mechanism for overcoming the limitations of similarity-based retrieval.
2026-06-12	GitOfThoughts: Version-Controlled Reasoning and Agent Memory You Can Replay, Diff, and Merge
• This paper introduces GitOfThoughts, which maps an agent’s reasoning tree onto a Git repository: scored reasoning nodes correspond to commits, parent–child reasoning relationships correspond to commit history, and scores and verification results are recorded through notes and tags, respectively. This design supports deterministic replay, line-by-line comparison, provenance tracing, and cross-agent merging. • The study compares five memory substrates—no memory, Markdown, vector databases, graph databases, and Git—on GPQA-Diamond and MATH-500. A preregistered replication experiment finds that no memory format consistently improves accuracy on genuinely novel problems, and the positive trends observed initially fail to replicate after the sample size is expanded. • Memory gains emerge primarily above a “copyability threshold,” where retrieved cases are highly similar to the current problem, and the resulting improvement more closely resembles answer reuse than methodological transfer. The central value of GitOfThoughts therefore lies not in improving generalization accuracy, but in providing auditability, reproducibility, and collaborative governance of reasoning processes while maintaining broadly comparable accuracy.
2026-06-11	Maestro: Workload-Aware Cross-Cluster Scheduling for LLM-Based Multi-Agent Systems
• This paper proposes Maestro, a workload-aware cross-cluster scheduling system designed for LLM-based multi-agent systems in resource-constrained cloud environments. • It predicts token output length and memory usage per stage to drive a hierarchical scheduler supporting dynamic multi-model co-location and elastic memory allocation. • Experiments show that Maestro significantly reduces memory overhead while improving SLO fulfillment and overall system efficiency.
2026-06-11	Learning What to Remember: A Cognitively Grounded Multi-Factor Value Model for Agentic Memory
• This work introduces a cognitively grounded multi-factor memory value model to guide storage, retention, and forgetting decisions under limited memory budgets for LLM agents. • It integrates seven interpretable factors (e.g., emotional intensity and goal relevance) into a unified framework for encoding and retrieval policies. • Results show improved memory efficiency and retention quality compared to baselines, along with strong interpretability and lightweight deployment.
2026-06-05	Rosetta Memory: Adaptive Memory for Cross-LLM Agents
• Rosetta Memory transforms stateless LLMs into persistent, cross-model memory systems that remain effective across different model switches and task distributions. • It introduces conditional operators that optimize how memory is stored and presented, enabling compatibility across heterogeneous LLM architectures. • Experiments demonstrate strong robustness across benchmarks, maintaining performance even under unseen model substitution scenarios.
2026-06-04	Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads
• This paper redefines agent memory as a stateful long-horizon workload rather than a simple retrieval-augmented component. • It introduces a stage-aware analytical framework that decomposes memory system costs across build, retrieval, and generation phases. • Through analysis of multiple systems, it derives system-level insights on scheduling design, latency trade-offs, and scalable deployment.
2026-06-04	TokenMizer: Graph-Structured Session Memory for Long-Horizon LLM Context Management
• Addresses context-window loss in long productive LLM sessions where important decisions and files disappear. • Models session history as a typed knowledge graph and serializes it into compact resume blocks. • Improves decision and file recall at lower token cost, providing a practical context-management system.
2026-06-02	eMEM: A Hybrid Spatio-Temporal Memory System For Embodied Agents
• Highlights that embodied agents need memory searchable by meaning, space, and time simultaneously. • Builds a hybrid graph memory with multiple indexes and tool-facing retrieval functions. • Adds eMEM-Bench around cognitive psychology paradigms to evaluate embodied memory behavior.
2026-06-02	Exploring Cross-Scenario Generality of Agentic Memory Systems: Diagnostics and a Strong Baseline
• Examines whether agentic memory systems generalize beyond the single scenarios they are often designed for. • Compares multiple memory systems across diverse scenarios and introduces AutoMEM as a self-managed baseline. • Shows that active agent control over storage and retrieval is more robust than fixed passive pipelines.
2026-05-30	CoMIC: Collaborative Memory and Insights Circulation for Long-Horizon LLM Agents in Cloud-Edge Systems
• Targets long-horizon agents running across resource-constrained edge models and stronger cloud models. • Uses decentralized execution with centralized reflection to filter reusable experience and circulate insights. • Improves progress and grounding without updating model parameters, showing a practical cloud-edge memory pattern.
2026-05-30	Cost and Accuracy of Long-Term Memory in Distributed Multi-Agent Systems Based on Large Language Models
• Provides an independent evaluation of long-term memory backends in distributed multi-agent systems. • Measures accuracy, latency, CPU time, RAM, disk I/O, network use, and total cost under cloud-edge scenarios. • Finds large cost-accuracy differences across vector, graph, hybrid, RAG, and full-context approaches.
2026-05-30	Hierarchical Long-Term Semantic Memory for LinkedIn's Hiring Agent
• Describes a production long-term semantic memory system for LinkedIn's hiring agent. • Extracts signals from noisy behavioral data, stores them hierarchically, and supports low-latency retrieval under privacy constraints. • Reports improvements in correctness and personalization, offering a rare deployed memory-system case study.
2026-05-12	Beyond Similarity Search: Tenure and the Case for Structured Belief State in LLM Memory
• Argues cross-session LLM memory is a state management problem, not a search problem, and introduces a typed belief schema with five belief types, epistemic status, versioned supersession, and a `why_it_matters` field that converts extracted facts into imperative instructions rather than declarative facts. • Demonstrates that cosine similarity over dense embeddings achieves mean precision of 0.12 on a 72-case retrieval suite, while alias-weighted BM25 with hard scope isolation achieves 1.0, passing 72/72 cases; under multi-turn topic drift, vector search produces drift scores of 0.43–0.50 on noise-critical turns while BM25 maintains 0.0. • Ships as a local-first OpenAI-compatible proxy that injects curated belief context into every LLM session transparently; includes a reusable 72-case benchmark covering alias resolution, scope disambiguation, supersession chain exclusion, and session-level noise isolation.
2026-05-06	Continual Knowledge Updating in LLM Systems: Learning Through Multi-Timescale Memory Dynamics
• Proposes Memini, a directed associative memory system for LLMs that continually updates knowledge through multi-timescale memory dynamics driven directly by incoming document streams. • Implements coupled fast and slow variables on graph edges based on the Benna-Fusi model, inherently enabling episodic sensitivity, gradual consolidation, and selective forgetting without explicit rules. • Leverages spreading activation over experience-shaped, dynamic edge weights for retrieval, allowing the memory substrate to autonomously reorganize and adapt to evolving evidence.
2026-05-06	Storage Is Not Memory: A Retrieval-Centered Architecture for Agent Recall
• Proposes True Memory, a six-layer agent memory architecture that shifts from extraction-based storage schemas to a multi-stage retrieval pipeline operating over verbatim conversational events. • Introduces a three-signal encoding gate (novelty, salience, prediction error) and achieves strong performance across LoCoMo, LongMemEval, and BEAM-1M benchmarks using a single SQLite file on commodity CPUs.
2026-05-05	MEMTIER: Tiered Memory Architecture and Retrieval Bottleneck Analysis for Long-Running Autonomous AI Agents
• Proposes MEMTIER, a tripartite memory architecture for long-running autonomous agents that addresses context collapse, compaction discontinuity, structural blindness, and missing attribution loops in existing flat-file systems. • Features a structured episodic JSONL store, a five-signal weighted retrieval engine, an asynchronous consolidation daemon, and an attention-attributed cognitive weight update loop with a PPO-based policy framework. • Achieves significant performance improvements on the LongMemEval-S benchmark, demonstrating that high-precision, structurally isolated memory (episodic vs. semantic) drives long-horizon success while highlighting legacy linear-combination retrieval as the current bottleneck.
2026-05-05	What Happens Inside Agent Memory? Circuit Analysis from Emergence to Diagnosis
• Analyzes functional circuits inside LLM agent memory systems across Qwen-3 scales with mem0 and A-MEM frameworks, tracing write, manage, and read circuits. • Discovers that routing-control circuits can emerge before content circuits in small models, and leverages feature-space separation for unsupervised stage-wise fault localization of silent memory failures.
2026-05-05	Deco: Extending Personal Physical Objects into Pervasive AI Companion through a Dual-Embodiment Framework
• Proposes Deco, a dual-embodiment framework extending personal physical objects into pervasive AI companions using multimodal LLMs and augmented reality with synchronized digital-physical interaction. • Introduces Reciprocal Memory as a design principle for sustaining emotional bonds, with a 7-day deployment study showing improvements in companionship, emotional connection, and sustained engagement.
2026-04-18	Skilldex: A Package Manager and Registry for Agent Skill Packages with Hierarchical Scope-Based Distribution
• Introduces Skilldex, a package manager and registry for LLM-agent skill packages, using a format-consistency score to check compliance with Anthropic skill conventions. • Adds the skillset abstraction to bundle related skills and shared resources, with three-level scope distribution, human-feedback suggestion loops, a metadata registry, and an MCP server.
2026-04-18	Bolzano: Case Studies in LLM-Assisted Mathematical Research
• Presents Bolzano, an open-source multi-agent LLM system that coordinates parallel proving and verification agents while maintaining a persistent knowledge base across turns. • Solves six problems in mathematics and theoretical computer science, with four publishable-level results and three produced almost autonomously.
2026-04-18	GenericAgent: A Token-Efficient Self-Evolving LLM Agent via Contextual Information Density Maximization (V1.0)
• Introduces GenericAgent, a self-evolving LLM agent for long-horizon tasks that maximizes decision-relevant information density in limited context. • Combines a minimal toolset, hierarchical on-demand memory, self-evolution, and context truncation/compression, converting verified trajectories into reusable SOPs and code.
2026-04-11	Mosaic: Cross-Modal Clustering for Efficient Video Understanding
• Proposes Mosaic, a cross-modal clustering system for efficient long-video understanding. • Finds implicit clustering structure in VLM KV caches and uses clusters as the unit for cache organization, maintenance, and retrieval. • Reduces fragmented transfers and management cost, achieving up to 1.38x speedups in streaming video inference.
2026-04-10	EpiAgent: An Agent-Centric System for Ancient Inscription Restoration
• Proposes EpiAgent, an agent-centric system for ancient inscription restoration. • Uses an Observe-Conceive-Execute-Reevaluate loop in which a central LLM planner coordinates multimodal analysis, historical experience, tools, and self-reflection. • Outperforms prior methods on real restoration tasks and improves both quality and generalization.
2026-04-10	Building an Internal Coding Agent at Zup: Lessons and Open Questions
• Reports the lessons from taking Zup's internal coding agent CodeGen from prototype to production. • Shows that reliability depends not just on the base model but also on tool design, safety guardrails, state management, and human oversight. • Uses string-replacement editing, layered safety constraints, and progressive supervision to raise adoption and reliability while outlining open questions.
2026-04-09	PSI: Shared State as the Missing Layer for Coherent AI-Generated Instruments in Personal AI Agents
• Proposes PSI, a shared-state architecture that connects separately generated personal AI modules into a coherent tool ecosystem. • Exposes current state through a personal context bus and supports writeback so modules can reason across each other and stay synchronized. • Demonstrates that the system can automatically integrate later-generated tools in a three-week personal AI deployment.
2026-04-09	Omakase: proactive assistance with actionable suggestions for evolving scientific research projects
• Proposes Omakase, a proactive research assistant for evolving scientific projects. • Monitors project documents, identifies timely deep-research queries, and distills long reports into project-grounded actionable suggestions. • User studies show that its suggestions are more timely and actionable than raw reports.
2026-03-15	SuperLocalMemory V3: Information-Geometric Foundations for Zero-LLM Enterprise Agent Memory
• For persistent memory in enterprise-grade agents, it proposes a mathematically grounded framework that unifies retrieval, lifecycle management, and consistency verification, overcoming the heavy reliance of traditional memory systems on engineering heuristics. • On the LoCoMo benchmark, it achieves an average improvement of 12.7% over engineering baselines, with gains of 19.9% on the most challenging samples, validating the effectiveness of information-geometric methods for complex memory retrieval. • This work advances agent memory beyond the stacking of engineering components toward a system design paradigm that is provable, governable, and compliant, with particular significance for high-privacy and high-sovereignty scenarios.
2026-03-05	Memory as Ontology: A Constitutional Memory Architecture for Persistent Digital Citizens
• Proposes a new Memory-as-Ontology paradigm for persistent digital agents, addressing the problem of identity continuity across successive generations of AI models. • Establishes a constitutional memory architecture that supports the long-term existence and evolution of digital citizens through a four-layer governance hierarchy, multi-layer semantic storage, and full lifecycle design. • Relegates the underlying computational model to a replaceable substrate, prioritizing the persistence of digital identity and governance over the optimization of short-term memory access and retrieval efficiency.
2026-03-04	Memex(RL): Scaling Long-Horizon LLM Agents via Indexed Experience Memory
• Existing approaches that handle long-context window limitations through truncation or summarization are prone to losing critical interaction evidence. • Proposes Memex, an indexed experience memory mechanism for lossless compression, and introduces the MemexRL reinforcement learning framework to optimize agents' read-write archiving strategies. • The system enables agents to autonomously decide when to archive and retrieve information, substantially improving long-horizon task success rates while significantly compressing the working context.
2026-02-25	MemoPhishAgent: Memory-Augmented Multi-Modal LLM Agent for Phishing URL Detection
• Proposes the MPA agent, which leverages episodic memory of past reasoning trajectories to guide decision-making against both recurring and novel phishing threats. • Achieves a 13.6% improvement in recall over state-of-the-art baselines on public datasets, with gains of up to 20% in real-world scenarios. • Analysis shows that episodic memory contributes approximately 27% of the recall improvement without introducing additional computational overhead.
2026-02-18	MMA: Multimodal Memory Agent
• Proposes the Multimodal Memory Agent (MMA), which addresses overconfidence errors in RAG by assigning dynamic reliability scores to retrieved items, incorporating source credibility, temporal decay, and conflict awareness. • Introduces the MMA-Bench benchmark to control speaker reliability and evaluate belief dynamics under text–vision contradictions. • Reveals the “visual reassurance effect,” where agents tend to inherit latent visual biases from the underlying models.
2026-02-07	M2A: Multimodal Memory Agent with Dual-Layer Hybrid Memory for Long-Term Personalized Interactions
• It proposes the M2A framework, which utilizes collaborative agents (ChatAgent and MemoryManager) to transform static multimodal personalization into a co-evolving memory mechanism that updates during interactions. • It introduces a dual-layer hybrid memory architecture linking raw message logs with high-level semantic observations via evidence IDs, employing tri-path retrieval to ensure precise and efficient context recall. • It develops a scalable multimodal data synthesis pipeline for long conversations, with experiments demonstrating that M2A significantly outperforms existing RAG and text-only memory baselines in personalized reasoning.
2026-01-28	Memory Retrieval in Transformers: Insights from The Encoding Specificity Principle
• Investigated memory mechanisms in Transformer attention layers drawing on the "Encoding Specificity Principle" from psychology. • Proposed that Q encodes retrieval context, K indexes memory traces, and V stores content, identifying that contextual cues are encoded as keywords. • Identified specific attention neurons whose activations facilitate keyword retrieval, providing a theoretical basis and extraction method for downstream tasks like unlearning.
2026-01-28	S3-Attention: Attention-Aligned Endogenous Retrieval for Memory-Bounded Long-Context Inference
• Proposed S3-Attention (Sparse & Semantic Streaming Attention), achieving O(1) GPU memory for long-context inference. • Uses Sparse Autoencoders (SAE) to decode attention states into sparse features, building a CPU-based inverted index and discarding the KV cache. • Retained near-full-context performance (e.g., 99.4% on Llama-3-8B) on LongBench via endogenous retrieval based on feature co-activation.
2026-01-19	LLM-as-RNN: A Recurrent Language Model for Memory Updates and Sequence Prediction
• Proposed LLM-as-RNN, an inference-only framework converting frozen LLMs into recurrent predictors to address immutable context history. • Represents the hidden state as a natural language memory (structured system prompt) updated via feedback-driven text rewriting at each step. • Significantly outperformed Zero-shot and MemPrompt baselines in healthcare, meteorology, and finance time-series tasks under fixed token budgets.
2026-01-14	Continuum Memory Architectures for Long-Horizon LLM Agents
• Introduced a continuum-style memory architecture to support long-horizon agent behavior beyond a single context window. • Described how different memory stores (e.g., working/episodic/semantic) interact over time for planning and recall. • Provided empirical evidence that architectural choices improve long-horizon task stability and retrieval quality.
2025-12-17	Memory Bear AI: A Breakthrough from Memory to Cognition
• Memory Bear constructs a human-like memory architecture grounded in cognitive science (ACT-R, Ebbinghaus), distinguishing between explicit and implicit memory to achieve a leap from "memory" to "cognition." • It features a three-layer architecture with an Intelligent Semantic Pruning Algorithm and a Self-Reflection Engine, significantly reducing token consumption (~90%) while actively suppressing hallucinations. • Experiments demonstrate it outperforms Mem0 and MemGPT in accuracy and latency, with validated effectiveness in healthcare (chronic disease management), enterprise operations, and personalized education.
2025-12-11	O-Mem: Omni Memory System for Personalized, Long Horizon, Self-Evolving Agents
• O-Mem is a novel memory framework based on active user profiling that dynamically extracts and updates user characteristics and event records from proactive interactions. • Unlike systems relying on semantic grouping, O-Mem supports hierarchical retrieval of persona attributes and topic-related context to enable adaptive and coherent personalized responses. • The system achieves state-of-the-art performance on LoCoMo and PERSONAMEM benchmarks while significantly improving token efficiency and interaction response time compared to previous frameworks like LangMem and MemoryOS.
2025-11-21	Episodic Memory in Agentic Frameworks: Suggesting Next Steps in Workflow Creation
• Proposed an episodic memory store for workflow traces, enabling retrieval of similar historical workflows. • Used retrieved workflow episodes to suggest plausible “next steps” during workflow creation, reducing hallucination pressure on the LLM. • Evaluated next-step recommendation quality in agentic workflow settings.
2025-11-11	From Experience to Strategy: Empowering LLM Agents with Trainable Graph Memory
• Introduced a trainable graph-structured memory to store experiences and support strategy formation over time. • Learned how to weight/route experiences in the memory graph to inform future decisions and planning. • Reported improvements on agent tasks requiring multi-step reuse of prior experience.
2025-10-21	LIGHTMEM: LIGHTWEIGHT AND EFFICIENT MEMORY-AUGMENTED GENERATION
• LightMem is a lightweight memory architecture inspired by the Atkinson-Shiffrin human memory model, designed to balance performance and efficiency in LLMs. • It features a three-stage process: cognition-inspired sensory memory for filtering redundancy, topic-aware short-term memory for structured access, and long-term memory with sleep-time updates to decouple maintenance from inference. • Experimental results on LongMemEval and LoCoMo show LightMem outperforms strong baselines in accuracy while reducing token usage by up to 100× and API calls significantly.
2025-10-10	Seeing, Listening, Remembering, and Reasoning: A Multimodal Agent with Long-Term Memory
• Introduces M3-Agent, a novel multimodal agent framework that mimics human memory by processing continuous visual and auditory inputs to build entity-centric episodic and semantic long-term memories. • Proposes M3-Bench, a comprehensive long-video question answering benchmark comprising 1,020 videos from robot and web perspectives, designed to evaluate capabilities like person understanding and cross-modal reasoning. • Experimental results demonstrate that M3-Agent, trained via reinforcement learning, significantly outperforms strong baselines such as Gemini-1.5-Pro and GPT-4o in memory retention and reasoning tasks.
2025-10-08	A-MEM: Agentic Memory for LLM Agents
• A-Mem introduces a Zettelkasten-inspired dynamic memory organization that equips LLM agents with genuine long-term memory. • Beyond simple storage, A-Mem enables self-linking and self-evolution, allowing agents to achieve significant advantages in complex reasoning tasks. • Experimental results demonstrate that A-Mem outperforms existing methods in performance, efficiency, and scalability, laying a strong foundation for building more intelligent and autonomous LLM agents.
2025-10-01	Improving Code Localization with Repository Memory
• Modeled long-term “repository memory” from commit history, linked issues, and evolving code regions. • Added tools to retrieve historical changes and summarize active code areas to support localization decisions. • Demonstrated improved localization performance for software engineering agents on SWE-bench-style benchmarks.
2025-08-12	Livia: An Emotion-Aware AR Companion Powered by Modular AI Agents and Progressive Memory Compression
• Livia is an emotion-aware AR companion designed to combat loneliness through a modular multi-agent architecture and immersive augmented reality interactions. • It introduces two novel memory compression algorithms—Temporal Binary Compression (TBC) and Dynamic Importance Memory Filter (DIMF)—to efficiently manage long-term memory while retaining emotionally significant context. • The system integrates multimodal emotion recognition (text and voice) and an adaptive personality model, demonstrating high accuracy and fostering deeper emotional bonds with users.
2025-08-05	NEMORI: SELF-ORGANIZING AGENT MEMORY INSPIRED BY COGNITIVE SCIENCE
• Nemori is a self-organizing memory architecture inspired by cognitive science, designed to address the limitations of Large Language Models in long-term interactions by enabling persistent, adaptive memory. • It introduces the Two-Step Alignment Principle for autonomous episode segmentation and the Predict-Calibrate Principle for proactive knowledge distillation, moving beyond passive storage to active learning. • Experimental results on LoCoMo and LongMemEval benchmarks show that Nemori significantly outperforms state-of-the-art systems while using 88% fewer tokens than full-context baselines.
2025-07-23	H-MEM: Hierarchical Memory for High-Efficiency Long-Term Reasoning in LLM Agents
• Proposes H-MEM, a hierarchical memory architecture that organizes memory into four semantic levels with positional index encoding, enabling efficient layer-by-layer retrieval without exhaustive similarity computation. • Introduces a dynamic memory update mechanism that adjusts memory weights based on user feedback to reflect changing user interests and psychological states. • Experimental results on the LoCoMo dataset demonstrate that H-MEM consistently outperforms baselines in long-term dialogue tasks while significantly reducing computational cost and retrieval latency.
2025-07-10	MIRIX: Multi-Agent Memory System for LLM-Based Agents
• MIRIX is a modular, multi-agent memory system that addresses the limitations of flat memory architectures by integrating six specialized memory components—including Episodic, Semantic, and Procedural memory—managed by dedicated agents. • The framework introduces an "Active Retrieval" mechanism and a Meta Memory Manager to dynamically coordinate memory updates and retrieval, and validates these capabilities on a newly introduced multimodal benchmark, ScreenshotVQA, consisting of high-resolution user activity logs. • Experimental results show that MIRIX outperforms RAG baselines by 35% in accuracy on ScreenshotVQA while reducing storage by 99.9%, and achieves state-of-the-art performance on the LOCOMO long-form conversation benchmark.
2025-06-30	Ella: Embodied Social Agents with Lifelong Memory
• Introduces Ella, an embodied social agent equipped with a structured long-term multimodal memory system comprising name-centric semantic memory and spatiotemporal episodic memory. • By integrating this lifelong memory system with foundation models, Ella can retrieve relevant information for decision-making, plan daily activities, and build social relationships in a 3D open world. • Experimental results in a dynamic environment demonstrate Ella's ability to influence, lead, and cooperate with other agents, highlighting the potential of combining structured memory with foundation models.
2025-05-30	Memory OS of AI Agent
• MemoryOS aims to provide comprehensive and efficient memory management for AI agents. • Inspired by memory management principles in computer operating systems and the hierarchical structure of human memory, MemoryOS adopts a unique segment–page hierarchical storage architecture and comprises four core functional modules: memory storage, memory updating, memory retrieval, and response generation. • Experimental results show that MemoryOS significantly improves contextual coherence and personalized memory retention in long conversations across mainstream benchmarks; for example, on the LoCoMo benchmark, average F1 and BLEU-1 scores increase by 49.11% and 46.18%, respectively.
2025-05-28	MemOS: A Memory OS for AI System
• MemOS (Memory Operating System) is a memory operating system designed for AI systems that treats memory as a manageable system resource, unifying the representation, scheduling, and evolution of explicit, activation-based, and parameter-level memory to enable cost-efficient storage and retrieval. • MemOS adopts a three-layer architecture consisting of an interface layer, an operation layer, and an infrastructure layer. The interface layer interacts with users or upstream systems and provides standardized memory APIs; the operation layer organizes and schedules memory resources; and the infrastructure layer handles storage, security, migration, and data flow of memory. • MemOS provides operating-system-level support for cross-task adaptation, cross-modality evolution, and cross-platform migration. Its introduction marks a key transition for large models from “perception and generation only” to “memory-enabled and evolutionary” intelligence.
2025-04-28	Mem0 Building production-ready AI agents with Scalable Long-Term memory
• Mem0 is a memory architecture that dynamically extracts and integrates key information from conversations, enabling AI systems to remember important content and sustain cross-session dialogue. • The authors further propose Mem0g, which extends Mem0 by incorporating graph-structured memory (i.e., knowledge graphs), allowing AI systems to handle complex relational reasoning more effectively. • NLI tasks enhance constituency grammar induction ability, whereas SMS tasks reduce it in the upper layers.
2025-02-22	Echo: A Large Language Model with Temporal Episodic Memory
• Introduced an LLM design that explicitly incorporates temporal episodic memory to store time-indexed events. • Targeted improved recall of event sequences and time-sensitive dependencies for downstream reasoning. • Evaluated on episodic-memory-focused tests to demonstrate gains over baselines without explicit episodic memory.
2025-01-20	ZEP: A TEMPORAL KNOWLEDGE GRAPH ARCHITECTURE FOR AGENT MEMORY
• Introduces Zep, a memory layer service for AI agents powered by Graphiti, a dynamic and temporally-aware knowledge graph engine. • Zep synthesizes unstructured conversational data and structured business data while maintaining historical relationships, enabling agents to handle complex, evolving contexts. • Experimental results demonstrate that Zep outperforms MemGPT on the Deep Memory Retrieval (DMR) benchmark and achieves significant improvements in accuracy and latency on the more challenging LongMemEval benchmark.
2025-01-09	Embodied VideoAgent: Persistent Memory from Egocentric Videos and Embodied Sensors Enables Dynamic Scene Understanding
• Proposes Embodied VideoAgent, a multimodal agent that constructs persistent scene memory by fusing egocentric video with embodied sensory inputs like depth and pose to address dynamic scene understanding. • Features a VLM-driven memory update mechanism that dynamically tracks object state changes and relations during actions, ensuring the memory remains accurate over long-form interactions. • The agent achieves state-of-the-art performance on benchmarks such as Ego4D-VQ3D and OpenEQA, and demonstrates practical utility in generating synthetic embodied user-assistant interaction data.
2024-12-12	Memory Layers at Scale
• Studied trainable key–value “memory layers” as a sparse lookup module to add capacity without proportionally increasing FLOPs. • Presented a fully parallelizable implementation and explored scaling behavior with very large memory parameter counts. • Reported downstream improvements (notably on factual tasks) compared to dense and MoE baselines under matched budgets.
2024-06-16	Towards Lifelong Dialogue Agents via Timeline-based Memory Management
• Proposes THEANINE, a framework for lifelong dialogue agents that utilizes a relation-aware memory graph to store memories without deletion, preserving temporal and cause-effect connections. • Introduces a timeline-augmented response generation approach that retrieves and refines entire memory timelines, ensuring rich contextual cues are maintained for long-term interaction. • Presents TeaFarm, a counterfactual-driven evaluation pipeline designed to stress-test dialogue agents' ability to correctly reference past conversations, where THEANINE demonstrates superior performance over existing baselines.
2024-05-04	Memoro: Using Large Language Models to Realize a Concise Interface for Real-Time Memory Augmentation
• Memoro is a wearable audio-based memory assistant designed to minimize disruption during social interactions by utilizing Large Language Models (LLMs) for concise memory retrieval. • The system introduces a "Queryless Mode" that proactively infers user memory needs based on real-time conversational context, alongside a traditional "Query Mode" for explicit natural language requests. • User studies indicate that Memoro increases recall confidence and reduces device interaction time while effectively preserving the quality of ongoing conversations.

🧰 Resources

📊 Benchmarks and Tasks

Task Type	Benchmarks & Datasets
Personalized Task Evaluation	IMPLEXCONV, PERSONAMEM, PERSONAMEM-v2, PersonaBench, PersonaFeedback, LaMP, MemDaily, MPR, KnowMe-Bench, ATM-Bench
Comprehensive Evaluation	MemoryAgentBench, LifelongAgentBench, StreamBench
Memory Mechanism Evaluation	MemBench, Minerva, MemoryBench
Long-Term Memory Evaluation	LOCCO, LONGMEMEVAL, LOCOMO, MADial-Bench, StoryBench, DialSim, Mem-Gallery, RealMem, CloneMem
Long-Dialogue Reasoning	PREFEVAL, MiniLongBench
Long-Context Understanding	LongBench V2, LongBench, BABILong, HotpotQA
Long-Context Evaluation	SCBENCH, L-CiteEval, GLE, HOMER, RULER, MM-Needle
Long-Form Text Generation	LongGenBench
Episodic Memory Evaluation	PerLTQA
Memory Hallucination Evaluation	HaluMem
Web Interaction & Navigation	WebChoreArena, MT-Mind2Web, WebShop, WebArena

💻 Systems and Open Sources

Systems below are ordered by publication date:

System	Time	Stars	GitHub & Website
Zep	2023-05-19		https://github.com/getzep/zep https://www.getzep.com/
Agentmemory	2023-07-07		https://github.com/elizaOS/agentmemory No official website
Cognee	2023-10-09		https://github.com/topoteretes/cognee https://www.cognee.ai/
Letta	2023-10-26		https://github.com/letta-ai/letta https://www.letta.com/
Supermemory	2024-02-22		https://github.com/supermemoryai/supermemory https://supermemory.ai/
Memary	2024-04-26		https://github.com/kingjulio8238/Memary No official website
Second-Me	2024-06-26		https://github.com/mindverse/Second-Me https://home.second.me/
Mem0	2024-07-11		https://github.com/mem0ai/mem0 https://mem0.ai/
Memobase	2024-10-05		https://github.com/memodb-io/memobase https://www.memobase.io/
Agent Brain	2024-12-01		https://github.com/kaderosio/agent-brain No official website
Puppyone	2024-12-06		https://github.com/puppyone-ai/puppyone https://www.puppyone.ai/
LangMem	2025-01-22		https://github.com/langchain-ai/langmem https://langchain-ai.github.io/langmem/
Vestige	2025-01-25		https://github.com/samvallad33/vestige https://samvallad33.github.io/vestige/
A-Mem	2025-02-17		https://github.com/agiresearch/A-mem No official website
Mirix	2025-04-16		https://github.com/Mirix-AI/MIRIX https://mirix.io/
MemEngine	2025-05-04		https://github.com/nuster1128/MemEngine No official website
MemOS	2025-05-28		https://github.com/MemTensor/MemOS https://memos.openmem.net/
MemoryOS	2025-05-30		https://github.com/BAI-LAB/MemoryOS https://baijia.online/memoryos/
ReMe	2025-06-05		https://github.com/agentscope-ai/ReMe https://reme.agentscope.io/
Nemori	2025-06-30		https://github.com/nemori-ai/nemori No official website
Memori	2025-07-24		https://github.com/MemoriLabs/Memori https://memorilabs.ai/
MemU	2025-08-09		https://github.com/NevaMind-AI/memU https://memu.pro/
MemMachine	2025-08-16		https://github.com/MemMachine/MemMachine https://memmachine.ai/
MineContext	2025-09-30		https://github.com/volcengine/MineContext No official website
TiMem	2025-10-25		https://github.com/TiMEM-AI/timem https://timem.cloud
EverMemOS	2025-10-29		https://github.com/EverMind-AI/EverMemOS https://evermind.ai/
MemoryBear	2025-12-17		https://github.com/SuanmoSuanyangTechnology/MemoryBear https://www.memorybear.ai/
OMEGA	2025-12-17		https://github.com/omega-memory/omega-memory https://omegamax.co/
Autohand Code CLI	2025-12-20		https://github.com/autohandai/code-cli https://www.autohand.ai/code/
Hindsight	2025-12-22		https://github.com/vectorize-io/hindsight https://hindsight.vectorize.io/
MAGMA	2026-01-06		https://github.com/FredJiang0324/MAMGA No official website
Mengram	2026-02-11		https://github.com/alibaizhanov/mengram https://mengram.io/
widemem-ai	2026-02-23		https://github.com/remete618/widemem-ai https://widemem.ai
Riverse	2026-02-25		https://github.com/wangjiake/JKRiver https://wangjiake.github.io/riverse-docs/
SuperLocalMemory	2026-03-01		https://github.com/qualixar/superlocalmemory https://superlocalmemory.com/
Cog	2026-03-15		https://github.com/marciopuga/cog No official website
NeverOnce	2026-03-18		https://github.com/WeberG619/neveronce https://pypi.org/project/neveronce/
MHN AI Agent Memory	2026-03-21		https://github.com/shahzebqazi/mhn-ai-agent-memory No official website
LycheeMem	2026-03-23		https://github.com/LycheeMem/LycheeMem No official website
GoodMemory	2026-03-24		https://github.com/hjqcan/GoodMemory https://www.npmjs.com/package/goodmemory
MemClaw	2026-03-26		https://github.com/Felo-Inc/memclaw https://memclaw.me
MemPalace	2026-04-05		https://github.com/MemPalace/mempalace http://mempalaceofficial.com/
SwarmVault	2026-04-06		https://github.com/swarmclawai/swarmvault https://swarmvault.ai
PackRat	2026-04-09		https://github.com/kevdogg102396-afk/packrat https://www.npmjs.com/package/packrat-compress
SkillClaw	2026-04-10		https://github.com/AMAP-ML/SkillClaw https://arxiv.org/abs/2604.08377
Synap	2026-04-10		https://github.com/maximem-ai/maximem_synap_sdk https://maximem.ai
Mnemoverse	2026-04-10		https://github.com/mnemoverse/mcp-memory-server https://mnemoverse.com
Formative Memory	2026-04-11		https://github.com/jarimustonen/formative-memory No official website
taOSmd	2026-04-13		https://github.com/jaylfc/taosmd Docs & benchmarks
Engram	2026-04-13		https://github.com/FBISiri/engram No official website
ToolPipe	2026-04-17		https://github.com/COSAI-Labs/toolpipe-mcp-server https://toolpipe.dev
Origin	2026-04-19		https://github.com/7xuanlu/origin https://useorigin.app
Omnigraph	2026-04-22		https://github.com/ModernRelay/omnigraph No official website
Caura	2026-04-27		https://github.com/caura-ai/caura https://caura.ai/
Mnemory	2026-05-03		https://github.com/fpytloun/mnemory No official website
kaeru	2026-05-08		https://github.com/LamantinAI/kaeru No official website
Dakera	2026-05-12		https://github.com/dakera-ai/dakera-mcp https://dakera.ai/
Perseus	2026-05-18		https://github.com/tcconnally/perseus https://perseus.observer/
Talamus	2026-05-21		https://github.com/ampres-ai/talamus https://ampres-ai.github.io/talamus/
Agentic Task System	2026-05-29		https://github.com/renezander030/agentic-task-system https://www.npmjs.com/package/@reneza/ats-cli
OWASP Agent Memory Guard	2026-05-30		https://github.com/OWASP/www-project-agent-memory-guard https://pypi.org/project/agent-memory-guard/
Engrava	2026-06-02		https://github.com/sovantica/engrava https://engrava.ai
Mimir	2026-06-06		https://github.com/tcconnally/mimir https://perseus.observer/mimir
mem0ry4ai	2026-06-11		https://github.com/cremenescu/mem0ry4ai https://cremenescu.ro/en/mem0ry4ai/
AccInt	2026-06-15		https://github.com/maxbaluev/accreted-intelligence https://accint.xyz/
Lians	2026-06-25		https://github.com/Lians-ai/Lians https://www.lians.ai/
SQLite Graph Memory	2026-07-03		https://github.com/Palo-Alto-AI-Research-Lab/sqlite-graph-memory No official website
wiki	2026-07-01		https://github.com/plasma-ai/wiki https://docs.plasma.ai/wiki
Lians	2026-07-06		https://github.com/Lians-ai/Lians https://www.lians.ai/
Tura	2026-07-06		https://github.com/Tura-AI/tura https://turaai.net/
Tree Ring Memory	2026-07-07		https://github.com/TerminallyLazy/Tree-Ring-Memory https://terminallylazy.github.io/Tree-Ring-Memory/
Data Olympus	2026-07-08		https://github.com/knaisoma/data-olympus No official website
LWC	2026-07-29		https://github.com/JanYork/llm-wiki-cli https://janyork.github.io/llm-wiki-cli/
ViBo	2026-08-13		https://github.com/vnbochkarev-netizen/ViBo-memory https://wwwvibo.com
memory-shield	2026-08-26		https://github.com/vnbochkarev-netizen/memory-shield https://wwwvibo.com/memory-shield
Open Index	2026-08-01		https://github.com/DrDroidLab/open-index No official website
Superself	2026-07-23		https://github.com/fxylabs/superself https://superselfs.com/
Compartment	2026-07-20		https://github.com/MaxFreedomPollard/Compartment No official website
OpenViking	2026-08-15		https://github.com/volcengine/OpenViking https://openviking.ai/
Engraphy	2026-08-22		https://github.com/devon-clarkk/engraphy https://engraphy.tech
chamnan	2026-08-19		https://github.com/ArcticFox2029/chamnan
Verified Memory Vault	2026-08-24		https://github.com/secondbrainstarter/verified-memory-vault https://secondbrainstarter.github.io/verified-memory-vault/
kgai	2026-08-03		https://github.com/kgaidev/kgai https://kgai.dev/?ref=awesome-ai-memory
causal-memory	2026-07-26		https://github.com/JingxuanC/causal-memory No official website
ReasonGraph	2026-02-22		https://github.com/bgokden/reasongraph https://memory.primaxiom.ai

🎥 Multi-media resource

Type	Website Link	Video Introduction
Basic Theory of Memory	https://www.youtube.com/watch?v=k3FUWWEwgfc	Short-Term Memory with LangGraph
https://www.youtube.com/watch?v=WsGVXiWzTpI	OpenAI: Agent Memory Patterns
https://www.youtube.com/watch?v=fsENEq4F55Q	Long-Term Memory with LangGraph
https://www.youtube.com/watch?v=L-au0tvDJbI	LLMs Do Not Have Human-Like Working Memories
https://www.youtube.com/watch?v=RkWor1BZOn0	long-term memory and personalization for LLM applications
https://www.youtube.com/watch?v=CFih0_6tn2w	A Paradigm Shift to Memory as a First Class Citizen for LLMs
Memory-Related Tools	https://www.bilibili.com/video/BV1hom8YAEhX	LLMs as Operating Systems: Agent Memory
https://www.bilibili.com/video/BV1CU421o7DL	Langchain Agent with memory
https://www.bilibili.com/video/BV1arJazVEaX	Open Memory MCP
https://www.bilibili.com/video/BV11HxXzuExk	Agentic Memory for LLM Agents
Memory-Related Papers	https://www.bilibili.com/video/BV1XT8ez6E46	AI agent Survey Memory
https://www.bilibili.com/video/BV1f12wBpEXX	MemGen: Weaving Generative Latent Memory for Self-Evolving Agents
https://www.bilibili.com/video/BV1deyFBKEFh	MLP Memory: A Retriever-Pretrained Memory for Large Language Models
https://www.bilibili.com/video/BV18FnVzpE6S	How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior
https://www.bilibili.com/video/BV1mpbrzSEH9	Agent Workflow Memory
https://www.bilibili.com/video/BV1qEtozyEoh	Introduction to the Memory Mechanism of Large Language Model Agents
https://www.bilibili.com/video/BV1FGrhYhEZK	Memory Layers at Scale
https://www.bilibili.com/video/BV1aQ1xBkE45	Agentic Memory for LLM Agents
https://www.bilibili.com/video/BV1Yz421f7uH	Evaluating Very Long-Term Conversational Memory of LLM Agents
https://www.bilibili.com/video/BV19RWdzxEsR	Lightweight plug-in memory system

🧠 Adam Framework

• Description: 5-layer persistent memory and coherence architecture for local AI assistants built on OpenClaw. Solves AI amnesia (cross-session) and within-session coherence degradation.
• Layers: Vault injection, mid-session memory search, neural graph (7219+ neurons), nightly Gemini reconciliation, coherence monitor with scratchpad dropout detection.
• Validated: 353 sessions, 6619 message turns, 8 months production use on a live business by a non-developer.
• Platform: Windows/macOS/Linux, OpenClaw, local-first, model-agnostic.
• Links: GitHub | Live Demo | Interactive Proof

🤝 Make a Contribution

Issue Template:

Title: [paper's title]
Head: [head name1] (, [head name2] ...)
Published: [arXiv / ACL / ICLR / NIPS / ...]
Summary:
  - Innovation:
  - Tasks:
  - Significant Result:

💬 Community & Support

Join our community to ask questions, share your projects, and connect with other developers.

• GitHub Issues: Report bugs or request features in our GitHub Issues.
• GitHub Pull Requests: Contribute code improvements via Pull Requests.
• GitHub Discussions: Participate in our GitHub Discussions to ask questions or share ideas.
• WeChat: Scan the QR code below to join our discussion group, get the latest research information related to Memory, or promote your related research results.

🌟 Star Trends