Another Brain

Long-term memory your AI coding agents actually share. One brain, many agents: what Claude Code learns on Monday, Codex can recall on Friday.

It runs as a single installed executable — no server to start, no container, no database to administer, and nothing leaves your machine. Memories live in one SQLite file in your user directory; embeddings are computed locally on CPU.

Install

Another Brain is a standard Python package — the only prerequisite is Python 3.12+ (no daemon, no root, no container runtime). Install the published package into a venv with pip:

python -m venv .venv
.venv/bin/python -m pip install another-brain   # Windows: .venv\Scripts\python
.venv/bin/another-brain setup                   # Windows: .venv\Scripts\another-brain

Or install a checkout of this repo from its root:

python -m venv .venv
.venv/bin/python -m pip install .

The console script another-brain lands in the venv's bin directory (Scripts on Windows); activating the venv puts it on PATH.

Restart the harness and your agent has memory. setup is a one-shot bundle of model pull + connect: it downloads the hash-verified embedding model (~206 MB) once, then writes the MCP server entry into each detected harness's own config and installs a skill that teaches the agent when to use it — no JSON to edit by hand, on any OS. Both steps are idempotent, so re-running setup is safe; the individual commands stay available below.

uv is an optional convenience: uv tool install another-brain does the same install and puts another-brain on PATH in one command. The uv.lock file is a maintainer/development artifact — pip installs never read it, and uv is never needed at runtime.

Run another-brain connect with no arguments to see which harnesses are known and which are installed here. Supported today: claude-code, codex, cursor, gemini-cli, pi.

For pi the connector writes the standard {"command": "another-brain"} entry into ~/.config/mcp/mcp.json, like every other harness. With the tools renamed to bare verbs, each harness adds its own prefix on top of the wire names: pi's adapter exposes another_brain_health, Claude Code shows mcp__another-brain__health, and the MCP wire names themselves never change.

What your agent can do

Eight tools appear in the agent's toolbox (your harness may prefix them):

| Tool | What it does | |---|---| | remember | store one thing worth recalling later — a decision, a bug and its fix, a preference | | search | find memories by meaning and keywords at once | | recent | list the newest entries, or walk one day or one topic | | get | fetch one memory in full | | reinforce | a memory proved useful — keep it longer | | forget | a memory proved wrong — drop it | | health | is the brain reachable and which one is bound | | audit | what changed, when, and by which agent — never the memory text |

Memory here is a diary that forgets on purpose. Each entry gets a lifespan from its importance — 1 to 5 maps to 7, 30, 90, 180, or 365 days — and expires unless an agent reinforces it after actually using it. Nothing accumulates forever, and a memory that turns out to be wrong can be forgotten. Forgetting is soft for 30 days, so a mistake is recoverable.

Search combines two independent signals: semantic similarity (so "how do we handle expired tokens" finds a note about refresh logic) and full-text keyword match (so an exact error string or file path is findable verbatim).

0.13.0 — 2026-08-13

Changed Tool names shortened to bare verbs — remember, search, recent, get, reinforce, forget, health, audit (was brain_*); your harness still prefixes them, so Claude Code shows mcp__another-brain__remember and pi shows another_brain_remember. Hard rename, no aliases. pip is a first-class install path — python -m pip install another-brain; uv stays an optional dev convenience.

Commands

| Command | Purpose | |---|---| | another-brain | the MCP server itself (your harness runs this; you normally don't) | | another-brain setup | one-shot onboarding: pull the model + connect detected harnesses | | another-brain connect [harness…] | register the server + install the skill | | another-brain model pull / model status | download or check the embedding model | | another-brain recent [--limit N] | print the newest entries from the terminal | | another-brain doctor | full health report; exits nonzero if something is wrong | | another-brain admin restore\|hard-delete ID | undo a forget inside its grace window, or purge | | another-brain import-jsonl PATH | import a JSONL v1 export | | another-brain serve --http | optional loopback HTTP on 127.0.0.1:1905 instead of stdio |

recent, admin, connect, and doctor all work without the model installed.

Your data

Memories live in brain.sqlite3 in your per-user data directory, and the model in your per-user cache directory — another-brain doctor prints both exact paths. Nothing is uploaded; after model pull the tool never needs the network again.

| Variable | Effect | |---|---| | BRAIN_DATA_DIR | where brain.sqlite3 lives | | BRAIN_MODEL_CACHE_DIR | where the model lives | | BRAIN_ID | which brain this process is bound to (default default) | | TIMELINE_TIMEZONE | IANA zone deciding the diary day (default UTC) |

Each agent process loads its own copy of the embedding model, about 322 MiB of RAM once it has embedded something — worth knowing if you run several harnesses at once.

Platform support

Gated in CI on Linux x86_64, macOS 14+ Apple Silicon, and Windows x86_64, with Python 3.12–3.14. Linux ARM64 and Windows ARM64 work but have no CI hardware. macOS Intel, macOS 13 and older, and Alpine/musl are not supported — the install fails clearly rather than silently building from source. another-brain doctor reports the tier for your machine. Full matrix in CHANGELOG.md.

More

• CHANGELOG.md — release notes, support matrix, measured performance
• docs/deployment.md — harness setup in detail, configuration
• docs/mcp-tools.md — the tool contracts
• docs/memory-trust-model.md — how much to trust a recalled memory

MIT licensed.