Skip to content

Brains management

A Brain is the durable, curated knowledge a workspace accumulates. The org Brain is deliberately not one big store — it's composed of smaller, access-scoped mini-brains, and a Group Brain is a named set of member mini-brains. This lets knowledge stay visible only to those cleared for it.

Where Brains are managed

There isn't a single "Brains" console page; management is where the work is:

  • Create a Brain from the left sidebar's Brains section.
  • Manage a Brain — rename, archive, delete, review its records, and set grants — from the right rail's Brains panel in a chat.

Access is granted to groups at read / write / admin, exactly like integrations; see Groups & roles for the clearance model.

the right-rail Brains panel with a Brain's records and grants

Records, promotion, and versioning

  • Promote to Brain — findings, docs, and canvases become brain records. On promotion, a record is embedded (vectorised) so it can be found by meaning, not just keywords.
  • Semantic search runs over a Brain's records via pgvector.
  • Supersession — updating a record creates a new version that supersedes the old one; history is preserved, not overwritten.
  • Verify / unverify — records can be marked verified, and a review queue lets a curator work through what's been promoted.

Group Brains and mini-brains

  • A mini-brain is an access-scoped store — the unit that groups get cleared on.
  • A Group Brain composes several member mini-brains under one name, so a chat can draw on a coherent set of knowledge while each piece keeps its own clearance.

The Brain MCP server

A read-only Brain MCP server ships separately (mcp/brain-mcp). It exposes list_brains, get_brain, search_brain, and get_record over stdio to any MCP client, talking to the running API. It's a retrieval surface — reading and searching Brains, not writing them.

Brain tools (act/write) — coming soon

A Brain can be bound to integrations (see Integrations & plugins), but a Brain invoking a tool — reading a mailbox, querying a database, calling an MCP server — is not built yet. The tool-handler model and the write-tool confirmation UX are pending design decisions. Today Brains are knowledge and retrieval only.

LEMA Chat — sovereign, open-source multi-agent chat.