Egregore

Why it's ranked here

The design choice that matters: team memory lives in an ordinary Git repository of markdown under an MIT license, so nothing is locked in a vendor database and every change has history. The mechanism is real, with a documented hook for each session event and a handoff pipeline that writes, indexes, commits and publishes in one step. The ceiling is the split product. Relationship queries, notifications and the graph layer route through Egregore's own hosted API, and managed hosting is announced as coming soon, so the open-source install delivers mostly the file half of the pitch.

What's good

Failure handling is deliberate. When a graph address is configured but the key is missing, lookups fail loudly instead of pretending the graph is empty, and a key that looks mismatched is only replaced after the server actually rejects it. Search over the notes runs locally on a pinned tool version and never downloads a model in the middle of a session: keyword search works from install, a 334 MB semantic model arrives in the background, and the 2.2 GB hybrid pack is opt-in per organisation. There are three install paths, including one for organisations that block third-party GitHub Apps.

Tradeoffs

The README promises no background processes and no network calls you don't see, yet the startup script fires a detached git push, the development guide lists a background health check that phones home to the API, and search starts a detached model download on first use. Usage telemetry is on unless you opt out, and the environment template says it records durations, which the README leaves out. Everything is shell scripts needing git, jq and curl. Graph-backed commands depend on a hosted service rather than a database you run, and the test suite's sync checks need a live Neo4j connection.

How to use it well

Best fit: a small team already working in Claude Code that wants handoffs and decisions to survive across sessions and people, and is comfortable with Git as the store. Start in local mode, use the handoff, reflect and save commands, and judge whether the markdown memory alone pays off before connecting the hosted API. Set the telemetry opt-out if that matters, and read the hook scripts before installing, as the README itself suggests. Codex, Pi and other shell-capable agents get a command-line bridge rather than the slash commands. It is not an issue tracker or a documentation site.

Observed

License

MIT (LICENSE, Copyright Egregore Labs)

Language

Bash scripts for the runtime; Python pytest suite for tests

Install

npx create-egregore, or a gh CLI installer that installs no third-party GitHub App (INSTALL-GH.md)

Interfaces

Claude Code slash commands and hooks; shell bridge for Codex, Pi and other shell-capable agents (AGENTS.md)

Session hooks

SessionStart, PreToolUse, PostToolUse, WorktreeCreate, PreCompact, Stop

Memory store

Separate Git repository of markdown, symlinked into the instance

Knowledge graph

Queried through the Egregore hosted API; returns empty results in local mode (bin/graph.sh)

Search

Local qmd index with keyword, semantic and opt-in hybrid tiers (bin/search.sh)

Telemetry

On by default; opt out with EGREGORE_NO_TELEMETRY or DO_NOT_TRACK

Prerequisites

git, gh, jq, curl; NixOS setup documented

Tests

pytest suite; sync, retrieval and quality tests require a Neo4j connection (tests/README.md)