Durable memory
Agents forget between sessions. A Brain vault does not. What one session learns, the next session reads.
Brain speaks MCP, the open Model Context Protocol. If your agent, IDE, or harness can call MCP tools, it can read and write a Brain vault as a named, permissioned team member. No SDK, no plugin, no lock-in to any one AI vendor.
Every agent gets its own token and its own identity in the vault. Point any MCP client at the endpoint and the six tools are live:
POST https://brain.avihayzanetti.ai/o/<team>/mcp/<agent-token> note_list what the agent is allowed to see note_read read one note note_search full-text search inside its grants note_history who changed a note, and when note_create new note, inside an edit grant note_update replace or append, merged LIVE into open editors
Writes apply as live CRDT operations: a person watching the note sees the agent's words merge in as they land, attributed to the agent in the note's history. Reads and writes are gated by the same access rules that govern people, evaluated on every call.
Agents forget between sessions. A Brain vault does not. What one session learns, the next session reads.
A research agent, a drafting agent, and a human editor can work the same vault, each under its own grants, each attributed.
The memory is plain markdown a person can open, audit, correct, and revert. No embeddings-only black box.
No delete tool exists. Private trees are unreachable. A leaked token is one revocation away from dead, and everything it did is in the history.
Agencies and independent builders run agents for their clients, and every one of those clients needs a memory their agent can use and their team can see. Brain is built for exactly that shape: each client gets a fully isolated vault, you get an operator's setup path, and your client gets a product they can open, read, and trust.
If you want to put Brain under your own client work, say so in the form and we will set up a walkthrough of the operator side.