Cairn

Institutional memory for coding agents. Run one command in a repository. You and your agents get a map of the code, the intent behind it, what happened to it, what the team learned and what earlier agent sessions did. All of it comes through one CLI, one MCP server and one page.

uv tool install cairn-brain        # or: pipx install cairn-brain
cd your-repo
cairn

Why

Coding agents are capable, but they forget everything between sessions. They re-read the same files, miss the caller three hops away, re-apply the change that was reverted last spring and never hear about the convention the team agreed on in review. Cairn keeps that context for them, and every commit, spec and session adds to it.

Related MCP server: reprox-mcp

Install

uv tool install cairn-brain        # or: pipx install cairn-brain

You need Python 3.11 or newer and git. You don't need Docker, a database server or an API key. The first sync downloads a small local embedding model (about 70 MB) once. Without network access, Cairn uses a built-in hashing embedder instead, so search still works, only less semantically.

Start

cd your-repo
cairn

The first run sets the repository up without asking any questions:

• It adds the spec workflow (.cairn/workflow/ and the /cairn.* agent commands).

• It wires every agent it detects to the cairn MCP server and adds a Cairn block to AGENTS.md. Claude Code also gets session capture hooks, a session-start briefing, a status line, /cairn:* commands, four sub-agents and the session skills.

• It installs git hooks (post-commit, post-merge, post-checkout, post-rewrite) that re-sync in the background, so a commit never waits on Cairn.

• It runs the first sync, starts the local server and prints the page's address.

After that, cairn shows status. Setup is idempotent, and cairn uninstall removes what it added.

What you get on day one

None of this needs a model:

• cairn impact <file|symbol>: dependents, tests likely affected, files that usually change together, past fixes and reverts, the spec task that owns the code, team conventions, recent agent work and a risk level.

• cairn why <file|symbol>: rationale comments, the commits behind the lines, the requirement and task that produced the code, and recorded decisions.

• cairn drift: done tasks whose files are missing, requirements that no task covers, and code that changed after its task was closed.

• The map: cairn trace path|explain|query, cairn hubs, cairn areas, and cairn graph for views, the report, the wiki and exports.

• Memory seeded from the repository: accepted ADRs, answered clarifications in specs, rules in contributing and style docs, and reverted or explained fixes in git history.

• Session capture: prompts, files read and changed, and commands, recorded into .cairn/sessions.db. It's on for Claude Code by default. For other agents, run cairn sessions integrate install <agent>. Without a model, the events are turned into deterministic observations.

• Measured savings: every context pack an agent receives is logged with its size and the size of the files behind it. The Overview page shows the totals.

• cairn ui: one page with the map, specs, timeline, memory and sessions.

Model features

Cairn finds a model on its own ([models] provider = "auto"). It uses the first of these that works:

1. ANTHROPIC_API_KEY (or CAIRN_API_KEY)

2. OPENAI_API_KEY, or a [models] base_url for any OpenAI-compatible endpoint, including local ones

3. The signed-in Claude Code CLI (claude on your PATH), which runs on your own plan with no key

With a model, Cairn adds:

• observations and session summaries written from raw agent activity

• timeline facts with validity windows ("true from March until the May refactor")

• memory reconciliation: a new memory updates, merges with or retires the ones it overlaps

• narrated answers (cairn impact --explain, cairn why --explain, cairn ask)

• semantic drift checks (cairn drift --deep)

• document, paper and image extraction into the map (cairn graph extract)

Each job goes to the cheapest model tier that does it well. Every call is recorded in a ledger (cairn models --ledger). Model work runs on its own when a model is available, and it spends that provider's quota (or your Claude Code plan):

• After each agent turn, observations and summaries are written. To only queue the events, set [recall] worker_spawn = false. You can process them later with cairn sessions worker.

• Every sync builds timeline facts, capped at 150,000 tokens per sync by default. Change the cap with [deep] budget_tokens, or turn this off with [deep] enabled = false.

See models and cost.

The five layers

Teams

The same server can serve a whole team. It adds sign-in, roles, API tokens, git projects that re-sync on push, and an audit log:

cairn team init --email you@example.com --name "Your Name"     # first owner; prints a one-time password
cairn serve --team                                            # behind your HTTPS reverse proxy
cairn project add https://github.com/acme/api.git             # cloned on the server; first sync from the page
cairn team invite alice@example.com --role member             # prints a one-time invite link

Developers point their agents at the server with cairn agents connect --server https://cairn.example.com --project <project id> and export CAIRN_TOKEN. See teams.

Commands

Full reference: docs/cli.md.

Documentation

• Onboarding: the in-depth walkthrough — what a developer does, what an agent does automatically, real screenshots of every page, and how to read the savings numbers

• Architecture: engines, stores, the read model, sync, the server, MCP and the platform

• CLI reference

• MCP tools

• Agents: what cairn init wires for each agent, session capture, commands and sub-agents

• Configuration: every setting, environment variable and file location

• Models and cost: providers, tiers, budgets, the ledger and how savings are measured

• Teams: team mode, roles, invites, tokens, git projects, webhooks, audit and deployment

• Decisions (ADRs)

• Contributing

Cairn builds itself spec-first. Its constitution is in .cairn/workflow/memory/constitution.md, and its spec, plan and tasks are in specs/001-cairn-core/.

Licence

Apache-2.0 (see LICENSE). Cairn includes software from other open-source projects. Their notices are in NOTICE and their licence texts are in licenses/.