Open-source memory for SI agents
Memory that keeps up.
Your agents remember what happened, show where every answer came from, and say so when they don't know.
Read the research · open source, BSD-3-Clause
01
Continuity
Pick up where you left off. A new session, a new model, a reset context: the work is still there, with its source and its date. Nothing to re-explain.
Your history isn't locked inside anyone's weights.
02
Memory
Your notes, decisions and repositories, indexed on your own machine. Embeddings run on-device by default, so your documents and queries are never transmitted.
Recall comes back in milliseconds, from anything you've chosen to add, whether it went in today or a year ago.
No account. No key. Measured: about 15 ms per query, on-device, over a 2,600-chunk brain.
03
Readback
Every answer is read back from the record, with the record it rests on. When nothing in the record supports an answer, it says so instead of guessing.
Every answer comes from the record, with its source and date attached, so you can check it. Behavioral gates refuse the moves our research links to drift and confabulation, at the point where they would do harm.
Illustration
What did we decide about the deploy flag?
Keep it off in production until the migration lands.
source decisions/deploy-flag.md · dated, with the line it came from
What is the staging database password?
Nothing in the record answers this. LBrain does not guess.
result declined, not paraphrased into an answer
Start in four commands
No account. No key. LBrain fetches its embedding model once on first run, then works offline.
pip install "lbrain[local]"lbrain init --source ~/noteslbrain import && lbrain embed --stalelbrain query "what did we decide about the deploy flag"
Works with any MCP-compatible agent: lbrain mcp.
The research behind it
Preprints with DOIs. In one, failure rates swing 34.4 points across records while varying negligibly across architectures: the failure lives in the record, so the fix can, too.