🎭 MUSE

The AI Coding Governance System — Roles, Memory, Skills, in Plain Markdown

The nine Muses of Greek mythology were daughters of Mnemosyne — the Titaness of Memory. They transformed their mother's gift of total recall into mastery of the arts and sciences.

MUSE inherits this lineage. It helps preserve decisions and evidence across AI conversations, with explicit boundaries when source history is incomplete.

New in 3.9.0: a lighter way to run several long-lived sessions as one team. Sessions run on automatic compaction and are replaced only for an account switch, a confused session or a long idle wait, not at a context percentage; parallel work inside a lane goes to subagents; full test suites run in CI; /handoff save shrinks to the checkpoint, the relay snapshot and new decisions; the machine load guard reads memory pressure. Complete usage and upgrade guide.

New to MUSE? Start here

Think of MUSE as your AI's project notebook: decisions, progress and a next step to come back to.

You explain what you want; the AI plans, works, checks and records progress. Remember: Resume to pick up work · Save to keep progress · Bye to formally wrap up. For a daily stop, save and pause.

Read the beginner guide → · Shareable web guide · 简体中文

Optional tools: Jev tool selection and data boundaries covers browser decisions, review, routing and on-demand retrieval. These third-party projects are not bundled with MUSE; ordinary Resume and Save do not require them.

MUSE is a Markdown-based governance system for AI pair programming, with a Python runtime for guarded continuity. It goes beyond format specs (like AGENTS.md or .cursorrules) by providing a full system — role isolation, persistent memory, 67 skills, cross-role directives, and visual dashboards — with portable local tooling for supported clients.

┌─────────────────────────────────────────────────────┐
│ L3  Memory Infrastructure   mem0 · MemOS · memsearch│  ← MUSE doesn't compete here
├─────────────────────────────────────────────────────┤
│ L2  Governance System       🎭 MUSE                 │  ← Only player at this layer
│     Roles · Memory · Skills · Directives · Dashboard│
├─────────────────────────────────────────────────────┤
│ L1  Workflow Kits           Spec Kit · sudocode      │  ← MUSE covers this + more
├─────────────────────────────────────────────────────┤
│ L0  Format Specs            AGENTS.md · .cursorrules │  ← MUSE is compatible
└─────────────────────────────────────────────────────┘

AGENTS.md defines the format. MUSE builds the system on top of it.

Inspired by LCM (Lossless Context Management) + lossless-claw. MUSE adapts these ideas into Markdown workflows and a Python runtime for guarded checkpoint updates.

📖 中文文档 / Chinese Docs

---

✨ Why MUSE?

Problem: AI coding assistants have context window limits. Format specs like .cursorrules or AGENTS.md give your AI instructions — but they can't manage memory, roles, or cross-session knowledge. Long conversations forget early content. New conversations start from scratch.

Solution: MUSE wraps your AI with a governance protocol — role isolation, persistent dual-layer memory, 67 reusable skills, cross-role directive queue, and visual dashboards. All in plain Markdown.

Without MUSE	With MUSE
AI forgets early context in long conversations	Pre/Post Compaction protocols preserve critical info
New conversations need manual context setup	/resume auto-assembles context in 5 steps
One persona for all tasks	Role isolation: build / QA / growth each see only their context
No cross-session memory	Dual-layer: memory/ (short-term) + MEMORIES.md (long-term)
Same mistakes repeated	/distill distills lessons → /search finds them later
No visibility into project health	Web Dashboard + VS Code Extension

🎯 Benchmark (run ./scripts/benchmark.sh on your own project):

• /resume loads 16.9× more context than a bare .cursorrules cold start
• 15/15 governance features vs 0/15 for format-spec-only setups
• 344 lines of accumulated project knowledge vs 0 retained across sessions

Optional

• Full documentation
• MUSE Protocol Spec v1.0
• Chinese docs
• Changelog
• Skill Index

Works with: Claude Code · OpenClaw · OpenCode · Cursor · Windsurf · Gemini CLI · Codex CLI · Copilot · Aider · Antigravity — or any AI tool that supports system prompts.

Tool	Install Command	Format
Claude Code / OpenClaw	./scripts/install.sh --tool claude	.agent/skills/ + CLAUDE.md
OpenCode	./scripts/install.sh --tool opencode	.agents/skills/ + CLAUDE.md
Cursor	./scripts/install.sh --tool cursor	.cursor/rules/*.mdc
Windsurf	./scripts/install.sh --tool windsurf	.windsurf/rules/*.md
Gemini CLI	./scripts/install.sh --tool gemini	.gemini/skills/ + GEMINI.md
Codex	./scripts/install.sh --tool codex	AGENTS.md entry + skill links + continuity runtime
Copilot	./scripts/install.sh --tool copilot	.github/copilot-instructions.md
Aider	./scripts/install.sh --tool aider	CONVENTIONS.md
Antigravity	./scripts/install.sh --tool antigravity	.gemini/antigravity/skills/

Recommended companion: nah — context-aware permission guard for Claude Code. Deterministic classifier that auto-allows safe operations, asks for ambiguous ones, and blocks dangerous patterns (e.g. curl | bash). Eliminates permission fatigue without sacrificing safety. pip install nah && nah install

---

🚀 Quick Start (5 minutes)

Option A: Interactive Setup (Recommended)

# Clone MUSE
git clone https://github.com/myths-labs/muse.git

# Run interactive setup — configures language, model, and preferences
cd muse && ./setup.sh

The setup wizard will ask for your language, AI model, and docs preferences, then configure everything automatically.

Option B: Multi-Tool Install (Cursor / Windsurf / Gemini CLI / Codex CLI)

# Clone MUSE
git clone https://github.com/myths-labs/muse.git

# Install for your tool (auto-converts to the right format)
cd muse && ./scripts/install.sh --tool cursor --target /path/to/your-project

# Or auto-detect installed tools
./scripts/install.sh --target /path/to/your-project

Supported: claude, openclaw, opencode, cursor, windsurf, gemini, codex, or all.

Option C: Manual Setup

# Clone MUSE
git clone https://github.com/myths-labs/muse.git

# Copy templates to your project
cp muse/templates/CLAUDE.md your-project/CLAUDE.md
cp muse/templates/USER.md your-project/USER.md
cp muse/templates/MEMORIES.md your-project/MEMORIES.md
mkdir -p your-project/memory your-project/.muse

# Copy skills & workflows
cp -r muse/skills your-project/.agent/skills
cp -r muse/workflows your-project/.agent/workflows

# Add MUSE entries to your .gitignore
cat muse/templates/.gitignore-template >> your-project/.gitignore

Your project should look like:

your-project/
├── CLAUDE.md              # 📜 Constitution (AI iron rules)
├── ETHOS.md               # 💡 Builder philosophy (5 principles)
├── USER.md                # 👤 Your preferences
├── MEMORIES.md            # 🧠 Long-term lessons
├── .muse/                 # 🎭 Role states
│   └── build.md           # ⚙️ Dev execution
├── memory/                # Short-term memory
│   └── YYYY-MM-DD.md
├── .agent/
│   ├── skills/            # Skills library
│   └── workflows/         # resume/bye/sync/distill/ctx/sprint/retro
└── [your code]

2. Customize CLAUDE.md

This is MUSE's core — the AI's "constitution". Edit to match your project:

# Iron Rules
1. All communication in [your language]
2. Check Skills before ANY task
3. Large files: view ≤300 lines at a time
4. Context ≥ 80%: immediately exit
5. Verify before claiming done
6. End conversations with /bye

3. Start using

You: /resume strategy Lane A  ← Recover your actual role and workstream
     ... work ...
You: /save                    ← Save progress and keep working
You: Save and pause for today. ← A daily stop usually needs no Bye
You: /bye                     ← Formally review and close out a stage

strategy Lane A is an example; retain your project's role and Lane. In the same conversation, just say "continue."

Sprint workflow (v2.29+):

You: /sprint           ← Think → Plan → Build → Review → Test → Ship → Reflect
You: /retro             ← Weekly stats from git log + memory (commits, shipped, lessons)

---

🏛 Architecture

graph TB
    subgraph "🏛 Constitution Layer"
        A["CLAUDE.md"] --> B["Iron Rules + Safety"]
        A --> C["ETHOS.md<br/>Builder Philosophy"]
    end
    
    subgraph "🎭 Role Layer"
        R[".muse/*.md<br/>build · qa · growth"] --> S["Role Isolation"]
        R --> T["📡 Directive Queue<br/>cross-role messaging"]
    end
    
    subgraph "🧠 Memory Layer"
        D["memory/*.md<br/>short-term"] -->|"/distill"| E["MEMORIES.md<br/>long-term"]
        D --> D1["Drift Detection<br/>stale >7d auto-flag"]
        E --> E1["4 Categories<br/>preference · feedback<br/>project · reference"]
    end
    
    subgraph "🤖 Coordinator"
        CO["coordinator-mode"] --> CO1["Understand → Dispatch"]
        CO --> CO2["Sub-agent Synthesis"]
    end
    
    subgraph "⚡ Skills Layer (67)"
        G["Trigger Skills<br/>on-demand"]
        H["Lifecycle Skills<br/>strategic-compact"]
    end
    
    subgraph "🔄 Workflow Layer"
        I["/resume"] --> J["Work"]
        J -->|"/sprint"| SP["Think→Plan→Build<br/>→Review→Test→Ship"]
        J -->|"/ctx"| K{"Healthy?"}
        K -->|"🟢"| J
        K -->|"🔴"| L["/bye"]
        L --> D
        L --> L1["Digital Twin<br/>profile update"]
    end

LCM Concept Mapping

LCM Concept	MUSE Implementation	Description
SQLite persistence	memory/ + MEMORIES.md	Markdown as database
Leaf nodes	memory/YYYY-MM-DD.md	Daily conversation snapshots
Condensed nodes	MEMORIES.md	Cross-day distilled lessons
Condensation	/distill	Leaf → long-term memory
Assembler	/resume	Context assembly
lcm_grep	./scripts/search.sh (TF-IDF)	Ranked semantic search
compact:before	Pre-Compaction Protocol	Save before compress
contextThreshold	/ctx 80% red line	Auto health check

---

📖 Commands

Command	Description	Input
/start	First-time setup — configures project, roles, language	None (interactive)
/resume [scope]	Boot — restore context & start work	build, growth, etc.
/sprint	Feature sprint — Think → Plan → Build → Review → Test → Ship → Reflect	None (guided)
/retro	Weekly retrospective — stats from git log + memory	None needed
/settings	Change language, AI model, or preferences	Subcommand (optional)
/ctx	Context health check (🟢🟡🔴)	None needed
/bye	One-click wrap-up — save, sync, archive	None needed
/distill	Condense memory/ → MEMORIES.md	None needed
/search	TF-IDF search across memory, roles, skills	Query string
/sync [direction]	Cross-file sync in multi-role setup	Direction (optional)
/sync receive	Pull updates from other roles mid-conversation	None needed
/resume [project] qa	Start QA verification (independent from build)	Project name (optional)
/resume crash	Recover from context blowout	None needed
/handoff save · /handoff <lane> · /handoff status	Multi-lane account relay: save before switching accounts, continue under the new one	Lane id from .muse/orchestration.json
/away (/托) · /back (/回)	Away mode: agents decide what they may and queue the rest; on return, one list	None needed

Defensive Auto-Save (L0 Defense)

MUSE doesn't wait for context to run out. Every 10 interaction rounds, it silently updates memory/CRASH_CONTEXT.md. If a sudden blowout occurs, at most 10 rounds of progress are lost.

Layer	Trigger	Reliability
L0	Every 10 rounds (silent)	⭐⭐⭐
L1	🔴 context detection	⭐⭐
L2	/resume crash scans convo/	⭐

Auto-Distill Detection

Every /bye automatically checks memory/ accumulation. Reminds you to /distill when:

• ≥ 7 days of un-distilled logs
• ≥ 5 new log files since last distill
• ≥ 15 total files and never distilled

---

🧩 Skill System

Loading Behavior

Type	When Loaded	Examples
Always-on	Every turn automatically	CLAUDE.md iron rules, Safety
Trigger	When task matches	git-commit, systematic-debugging
Lifecycle	On specific events	strategic-compact (compression), verification (completion)

Tier Classification

Tier	Description	Ships with MUSE?	Examples
🏛 Core	Required for MUSE to function	✅ Built-in	context-health-check, strategic-compact, verification-before-completion, using-superpowers
🔧 Toolkit	General dev tools, recommended	✅ Included	git-commit, systematic-debugging, security-review, tdd-workflow, frontend-design, ui-ux-pro-max, +19 more
🎯 Domain	User-created, domain-specific	❌ Private	Your own custom skills

Skill Lifecycle

memory/ lessons repeat → /distill finds pattern → write to MEMORIES.md
→ appears ≥3 times → upgrade to CLAUDE.md constitution
→ methodology is generic enough → create new Skill
→ useful across projects → contribute to MUSE Toolkit

---

📁 Directory Convention

Standard Project Structure

project/
├── CLAUDE.md              # 📜 Constitution
├── README.md              # Public README
├── LICENSE                # License
├── USER.md                # Preferences (private)
├── MEMORIES.md            # Long-term lessons (private)
├── assets/                # 🎨 Project assets (public)
│   ├── logo.png           # Project logo
│   ├── banner.png         # README/social banner
│   ├── screenshots/       # App screenshots
│   ├── diagrams/          # Architecture diagrams
│   └── social/            # Social media assets (OG images, previews)
├── .muse/                 # 🎭 Role states (private)
│   ├── build.md / qa.md / growth.md / ...
├── memory/                # Short-term memory (private)
│   └── YYYY-MM-DD.md
├── convo/                 # Conversation archives (private)
│   └── YYMMDD/
├── docs/                  # Documentation
│   ├── [public docs]      # → git push ✅
│   └── internal/          # Strategy/fundraising (private)
├── src/ | packages/       # Source code
└── .agent/                # Skills + Workflows (private)

Naming Conventions

Category	Pattern	Example
Memory logs	YYYY-MM-DD.md	2026-03-12.md
Conversations	YYMMDD-NN-desc.md	260312-02-muse_setup.md
Crash archives	+_CRASH suffix	260312-05-debug_CRASH.md
.muse roles	[role].md lowercase	build.md, qa.md

Assets Convention

Subdirectory	Purpose	Naming Pattern
assets/ (root)	Logo, banner, favicon	logo.png, banner.png, favicon.ico
assets/screenshots/	App/feature screenshots	feature-name.png or YYMMDD-feature.png
assets/diagrams/	Architecture, flow charts	component-name-diagram.png
assets/social/	OG images, social cards	og-default.png, x-card.png

Tip: Keep assets/ in git (public). Large video files (>10MB) should use Git LFS or external hosting.

---

🔧 Customization

Minimal Setup (Personal Project)

Just 3 things:

• CLAUDE.md — Constitution (required)
• memory/ — Short-term memory (required)
• MEMORIES.md — Long-term memory (recommended)

Standard Setup (Indie Developer)

Add the role system:

• .muse/build.md — Dev state
• .muse/qa.md — Quality verification
• USER.md — Personal preferences

Full Setup (Team / Multi-Project)

Add GM + all roles + sync:

• .muse/strategy.md — Strategy (global, one per workspace)
• .muse/gm.md — Project GM (project-level CEO)
• .muse/build.md + qa.md + growth.md + ops.md + research.md + fundraise.md
• /sync workflow — Cross-role sync

---

🔌 MCP Server (Claude Plugin)

MUSE includes a built-in MCP (Model Context Protocol) server — a zero-dependency Bash implementation that lets Claude Code, Cursor, and other tools access your project's role files, memory, and directives through a standard protocol.

Quick Setup

# From your MUSE-enabled project
./scripts/mcp-server.sh --help

Available Tools

Tool	Description
muse_get_status	Read all L0 lines → project overview (~400 tokens)
muse_list_roles	List all role files with summaries
muse_get_role	Deep-read a specific role file
muse_send_directive	Send 📡 cross-role directive
muse_write_memory	Append to today's memory log
muse_search_memory	Search across memory files

Configuration

Add to your tool's MCP config (e.g., ~/.config/claude/claude_desktop_config.json):

{
  "mcpServers": {
    "muse": {
      "command": "/path/to/muse/scripts/mcp-server.sh",
      "args": ["--project-root", "/path/to/your/project"]
    }
  }
}

Requires: jq (brew install jq on macOS, apt install jq on Linux)

---

🌐 Optional Third-Party MCP Integrations

MUSE links well-vetted third-party MCP servers as optional add-ons. Each runs independently, has its own token / config, and is fully opt-in.

MCP	Purpose	Cost	Setup
🎨 Lazyweb	UI design reference search · real-app screenshots + semantic descriptions (pricing pages, onboarding, dashboards, etc.)	Free (per-user token)	/setup-lazyweb — 3 steps, ~2 min

How it works: Open the vendor's install page, copy the generated install instructions block (token embedded), paste it to your AI agent (Claude Code / Cursor / Codex / OpenClaw / Windsurf / etc.), and say "install this". The agent picks the right install path for its client. Restart the client to load tools.

Why we link these: design taste calibration is non-trivial without real references. Lazyweb gives MUSE-driven projects access to a curated UI library — useful when an agent needs to advise on pricing pages, onboarding flows, dashboard layouts, or compare a mockup to industry patterns.

Want to suggest another MCP for the list? Open a PR · we vet for free-tier availability, no-billing safety, and per-user token isolation.

---

🤔 FAQ

Q: Does MUSE require installation? Markdown skills can be read directly. For guarded continuity, use the native installer with Python 3.7+ and Git; see docs/CONTINUITY.md.

Q: Which AI tools does it support? Ten tools with native install support: Claude Code, OpenClaw, OpenCode, Cursor, Windsurf, Gemini CLI, Codex CLI, Copilot, Aider, and Antigravity. Run ./scripts/install.sh --tool <name> to install in the correct format for each tool. Additionally, ./scripts/convert-skills.sh --tool <name> exports all 67 skills to any supported format. Import from agency-agents (35K+ ⭐) with --import agency-agents.

Q: How is this different from lossless-claw? lossless-claw is a code plugin (SQLite + DAG + sub-agents) that requires the OpenClaw runtime. MUSE uses Markdown skills and a Python continuity runtime. Its recorded-source boundaries and supported clients are described in the continuity guide.

Q: What if memory/ files pile up? Archive files older than 30 days to memory/archive/. Use /distill to extract key lessons into MEMORIES.md first, then safely archive the originals.

Q: Is there a visual dashboard? Yes! Two options:

• Online: Visit muse.mythslabs.ai/dashboard and load your project folder — 100% client-side, data stays in your browser.
• Local: Run ./scripts/dashboard.sh to generate a self-contained HTML dashboard at .muse/dashboard.html.

Q: How do I discover and share skills? Use ./scripts/skill-discovery.sh with commands like recommend "frontend", categories, stats, export <name>, or registry. See CHANGELOG for details.

Q: Is there a VS Code extension? Yes — vscode-extension/ provides an activity bar with Roles, Skills, and Memory tree views, plus in-editor dashboard and skill search. Run cd vscode-extension && npm install && npm run compile, then press F5.

💬 Follow Us

• 🌐 GitHub: Myths Labs
• 🐦 X (Twitter): @MythsLabs
• 💼 LinkedIn: Myths Labs
• 👤 Creator: @SunshiningDay — indie dev, solo-building MUSE

---

🙏 Credits

• gstack by Garry Tan — Sprint pipeline, builder ethos, and reframing methodology (YC CEO's open-source software factory)
• LCM Paper by Ehrlich & Blackman — Theoretical foundation for lossless context management
• lossless-claw by Martian Engineering — OpenClaw implementation of LCM
• nah by Manuel Schipper — Context-aware permission guard (complements MUSE's Safety Protocol)
• Greek Mythology — Mnemosyne and her nine Muses, eternal symbols of memory and creation

---

📜 License

MIT © Myths Labs

---

Built with 🎭 by Myths Labs — Solo-developed by JC

MUSE v3.9.0