another-brain MCP server
Commit Nov 14, 2026Provides shared long-term memory for AI coding agents using a local SQLite database with hybrid semantic and keyword search.
Also known as Flowerf19/another-brain
- Stars
- 3
- Tools
- —
- Downloads, 30 days
- —
- Latest version
- —
- Last commit
- Nov 14, 2026
- License
- MIT
What is another-brain MCP?
another-brain provides shared long-term memory for AI coding agents using a local SQLite database with hybrid semantic and keyword search.
Summary generated from the project's README
About another-brain MCP server
Another Brain provides a shared long-term memory system for AI coding agents. It allows different tools like Claude Code, Cursor, and Codex to access a unified set of recollections stored in a local SQLite database. By using a single executable, it avoids the need for complex server administration or containers, ensuring that all data remains on your machine. The system uses local CPU processing to compute embeddings for semantic search.
Developers can set up the environment using a single command that handles model downloads and harness configuration. The memory system functions like a diary that manages its own lifespan; entries are assigned an importance level that determines how many days they persist before expiring. Agents can actively manage this knowledge by using tools to reinforce useful information or forget incorrect details. This hybrid approach combines semantic similarity with keyword matching to ensure both conceptual ideas and specific error strings are easily retrievable.
What you can do with it
- 1Share knowledge between different agentsEnsure that a decision made by Claude Code on one day can be recalled by Codex later.
Written from the project's README. Every use case is checked against a quote from it.
Automated checks
Ran Nov 15, 2026Full report →Not verified yet. A server needs a score of 80, no serious finding, and enough data for us to check. How checks work · What to fix for more points
- Source repository linked and reachableflowerf19/another-brain is public and reachablegithub.com/flowerf19/another-brain ↗Pass
- Package names this repository as its sourceNot checked
- Build provenance (Sigstore attestation)Not checked
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit or release 2026-11-14github.com/flowerf19/another-brain ↗Pass
- OpenSSF ScorecardNot checked
- Package name not close to a popular packageNot checked
- No critical advisory in the package or its dependenciesNot checked
- No hidden instructions to the model0 tool descriptions and the README read; no override phrases, credential paths or hidden charactersPass
- Destructive tools are labelledNot checked
- Remote endpoint answers the MCP handshakeNot checked
README
View on GitHubAnother 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.
another-brain MCP questions
Does my data leave my machine?
No, everything is stored locally. Memories live in a SQLite file in your user directory and embeddings are computed locally on your CPU.
Which AI agents can use this memory?
It currently supports Claude Code, Codex, Cursor, Gemini-CLI, and Pi.
How does the memory management work?
The system uses a self-expiring diary where entries have a lifespan based on importance and expire unless reinforced by an agent.