An agent needs three things: a short description of how the vault works, a fast way to get context on a topic, and clear rules for writing back. Brain provides each one over the CLI, the MCP server and the HTTP API.

The skill file

brain/SKILL.md is a short how-to written for agents. It says what the vault is good for, which calls to make, how to write, and what not to trust. An agent can get it four ways:

The read pattern

1

Once per session: who the owner is now

brain context returns the current state, open decisions, active goals and latest meetings. Each section states its age.
2

Before working on something: a topic briefing

brain context "<topic>" returns one pack on a person, project, concept or phrase. See Context briefings.
3

Only if needed: go deeper

brain ask "<question>" for a specific question, brain get <id> for a full page, brain query '…' for a list, brain search "<words>" when you do not know the name.
Cite the note id, for example [[people/dana-wu]], for anything taken from the vault.

What an agent must not assume

  • Old data is history. A section marked MAY BE OUTDATED or Nd ago describes the past. For current facts the owner’s own record wins.
  • Missing is not false. If nothing matches, say so and ask. Do not infer.
  • Evidence lines are matches, not verified facts. Only entries under “Sourced claims” were recorded as facts with a source.

Choosing a connection