egregore-labs/egregore

Shared memory for people and their AI agents.

Website · Community · MIT License

---

Egregore gives people and their AI agents a shared organizational memory. Decisions, discoveries, and handoffs accumulate across sessions, so each person can build on what the others have learned. Your organization owns the memory. Claude Code, Codex, Pi, and Prime Agent can all work with it.

One person works through a design decision in Claude Code and leaves a handoff. A teammate using Codex picks it up, asks why that approach was chosen, and retrieves the decision with its sources. The work continues from the reasoning that produced it.

An earlier workflow demo:

https://github.com/user-attachments/assets/71a690fb-8a4e-4d0b-9a61-e6cdc7595430

Start an Egregore

Recommended setup: Node.js 22 or later with npm, Git, jq, Bash, and Python 3.11 or later. Local setup uses the GitHub CLI for authentication and offers to install it if needed. Install and authenticate your chosen agent before starting a session.

npx create-egregore@latest

The installer guides you through GitHub authentication, naming your Egregore, and choosing an agent. It creates your instance and shared-memory repositories, sets up the local workspace, and prepares memory search. Keyword search is available first; semantic indexing can finish in the background.

Open a new terminal and run:

egregore

The launcher lets you choose your Egregore and agent, syncs shared memory, and starts a session with your identity and pending work. You can return to the same memory through any supported agent.

Join an existing Egregore. Ask its owner to invite your GitHub username, then use the repository's owner and name:

npx create-egregore@latest join <owner>/<egregore-repo>

Setup brings in the shared memory and managed project repositories. Onboarding introduces you to the organization and its working conventions.

Update an existing installation. In the launcher, select your Egregore, press s for Settings, then choose Runtime update. Review and activate the available update. Each person updates their own installation.

Work from shared memory

Memory lives in Markdown files in a separate Git repository, linked into your workspace as memory/. Decisions, handoffs, patterns, questions, and ongoing explorations retain their history. The organization's identity and conventions give that memory a shared frame of reference.

Local keyword and semantic search help your agent find relevant work and cite the records behind an answer. Ask “What did we decide about onboarding, and why?” or “Find the handoff about the redesign.” Egregore retrieves the relevant sources as you work.

The same workflows are available across supported agents. Use natural language, or /handoff, /reflect, and other commands in Claude Code, Pi, and Prime Agent. In Codex, use $handoff, $reflect, and the corresponding skills.

Workflow	What it does
handoff	Carry decisions, reasoning, and open threads into another person's session
reflect	Preserve a decision, pattern, or finding
deep-reflect	Investigate a question across accumulated knowledge
quest	Start or contribute to an open-ended exploration
ask	Leave a question for a teammate
activity	See recent work across the team
dashboard	See your personal status and pending work
todo	Keep track of personal tasks
save	Commit and push your changes, then create or update a pull request
invite	Give a teammate access to your Egregore

For example, ask “Invite octocat to this Egregore.” Egregore handles the access workflow and returns the joining instructions. Each person works in their own session; shared memory carries the context between them.

Your installation

The open-source runtime includes memory, local search, coordination workflows, and adapters for all four agents. Once dependencies and search models are provisioned, local memory and search can operate offline. GitHub synchronization and hosted model providers need their own connections.

Session startup, memory synchronization, search indexing, and worktree support are implemented through shared scripts and agent-specific hooks or extensions. Background work can keep repository references and search indexes current. The source is available in bin/, .claude/, .codex/, .pi/, and .prime/.

Telemetry. Egregore records local diagnostic identifiers and metrics to help inspect runtime behavior. Collection is on by default and easy to turn off. These diagnostics exclude prompts, code, file contents, and conversations. Nothing is uploaded automatically. Ask Egregore to show, clear, or turn off telemetry, or set EGREGORE_NO_TELEMETRY=1 in your .env. Your chosen agent's own telemetry settings are separate. See the telemetry specification for the recorded fields and controls.

Egregore Connect. Optional hosted services add membership administration, managed integrations, and notifications. Local memory and search remain on your machine. Run egregore connect to explore the setup, or see egregore.xyz for availability.

Contribute

Egregore is built by its users. From an existing installation, ask “Help me contribute this improvement upstream,” or use /contribute ($contribute in Codex). The workflow prepares a fork, branch, and pull request to this repository.

For direct contributions, see CONTRIBUTING.md. Egregore is MIT-licensed.

Built by Egregore Labs.

Find the others.