Cairn is open-source memory for coding agents. It keeps what each session learns about your repository and hands it to the next one, in whichever agent you open, before the first prompt.
$ uv tool install --python 3.12 cairn-brain
› The session-start hook hangs while a sync runs. Fix it. reading src/cairn/hooks.py, src/cairn/store.py ✗ database is locked: the hook waits behind the sync Opening the store wrote schema_version every time. Read-only opens skip the write now. ✓ tests pass · 2 files changed ● cairn_remember kind=gotcha Agent hooks never fail on a busy store stored with its source, commit 81efc41
▲ cairn · cairn — what this project remembers What the team knows (32 memories) gotcha Agent hooks never fail on a busy store (commit 81efc41, 6h ago) Handle with care src/cairn/hooks.py — 6 of 13 commits were fixes or reverts › Add a timeout to the session-start hook. Memory says hooks must never fail on a busy store. Keeping the read-only open; the timeout returns no brief.
A replay with staged prompts. The memory, the panel and the brief text are real output from Cairn's own repository on October 6, 2026. Filled dot: verified live. Ring: wired, not yet tested live.
01 What forgetting costs
Before editing src/cairn/router.py, an agent asked Cairn what depends on it. The cited answer cost 391 tokens. The 33 files behind that answer come to 144,228 tokens.
times fewer tokens. The answer still lists dependents, tests to run, past fixes and the task that owns the code.
Cairn on its own repository; savings as of October 8, 2026, other counts unchanged since October 6. Sources: cairn status, cairn recap, the session-start hook and the Overview page. Each answer is logged with its size and the size of the files it cites.
02 What it remembers
This is what Claude Code shows when a session opens in Cairn's own repository, unedited. It is built from the local store with no model calls. Every other agent receives the same memory as a written brief.
▲ cairn · cairn — what this project remembers 1Pick up where you left off · 3h ago in OpenCode Asked Verify cairn MCP server status by calling cairn_status tool… Done Successfully invoked cairn_status MCP tool; tool returned… Next Session idle pending next user request 2New since then + gotcha Injected context, hints and tool descriptions must name real… + commit fix: session capture never climbs above the repository root + 1 more commits 3What the team knows (32 memories) gotcha The MCP server creates nothing where Cairn is not set up… (commit 7a6c581, 4h ago) gotcha Agent hooks never fail on a busy store: Opening the store… (commit 81efc41, 6h ago) gotcha Windows support: cross-process file locks use msvcrt where… (commit 1481d1c, 1d ago) 4Handle with care src/cairn/hooks.py — 6 of 13 commits were fixes or reverts, last 2h ago src/cairn/cli.py — 4 of 13 commits were fixes or reverts, last 3h ago 5Open work 90 drift findings, e.g. `pyproject.toml` changed after T001… → cairn drift 875 files mapped · 32 memories · 1,216 session notes · http://127.0.0.1:4747
03 Across the lifecycle
People and agents trade work from spec to ship. Cairn is the memory both sides share, so each stage starts with what the last one learned and the next cycle starts ahead. Pick a stage to see the real output.
What becomes possible
04 Agents, data, cost
cairn global install wires each agent at user level, writing only what that agent supports. cairn uninstall removes exactly that. Switch agents mid-week and the memory comes along.
Session-start hooks for the panel and the brief, plus context on each code prompt
Verified liveA session-start hook; approve Cairn's hooks once under /hooks
Verified liveA plugin that adds the memory to the system prompt
Verified liveAn instructions file kept out of git
Verified liveAn always-on rule kept out of git
Verified liveA session-start hook, in the IDE and cursor-agent
Wired, not yet tested liveA session-start hook, also under gemini -p
Wired, not yet tested liveAlso wired: VS Code with Copilot, and Windsurf. Any other MCP client can run cairn mcp. Agents reach 13 MCP tools, including impact, why, recall and search. What each agent gets · MCP tools
Privacy
Where it all starts
Session capture, the code map, memories and the token ledger, in the repository's .cairn/ folder. None of it leaves unless you connect a team server.
Already your repository
Specs, ADRs and shared config travel with the code and get reviewed like code.
Optional, self-hosted
Shared memories and pushed sessions, with private blocks removed and secrets masked, under each member's token.
Optional, per call
Prompts for summaries and facts, only when a model is configured. Turn it off and everything above still works.
Cost
$ cairn models --ledger tier model calls input output fast claude-haiku-4-5-20251001 878 1,021,846 1,909,422 balanced claude-sonnet-5 217 432 442,679 deep claude-opus-5-5 22 5,213 18,497
Cache column trimmed. With the Claude Code provider, these tokens count against your plan's usage. Models and cost
05 Plans
Everything on this page up to here is in the open-source release. The paid plans add a shared view on top, for teams that want one.
Free · available now
Apache-2.0 · v0.2.2
InstallEarly access
Talk to us
06 For partners and investors
The case in three parts, for anyone deciding whether to build on Cairn, partner with us or back it.
Why now
Developers mix Claude Code, Codex, Cursor, Copilot and others, often in the same week. What the team knows ends up scattered by tool and by person.
Why Cairn
One memory for seven agents. It runs on the developer's machine under Apache-2.0 and keeps what it knows in the repository, the team's git and the team's own server.
The business
Developers adopt the open-source release one repository at a time. Teams pay for the hosted view across repositories. Enterprises pay for self-hosting, SSO and SCIM, audit and data controls.
Three commands, about 15 seconds of setup with no model. Then open your agent the way you always do. You need Python 3.11 to 3.13 and git. No Docker, database server or API key.
# install the CLI $ uv tool install --python 3.12 cairn-brain # wire every agent on this machine $ cairn global install # map the repository, seed memory, add git hooks $ cd your-repo && cairn init --no-deep