What to remember (and what not to)

A useful brain is small and reliable. Each memory takes up room in the recall briefing and can mislead an agent if it is wrong or stale. The brain's AGENTS.md passes these guidelines to agents; they apply to you too.

Worth remembering

What	Type	Example
A preference of yours	preference	"I prefer PR reviews in short bullet points, by severity."
A correction you made to the agent	preference or gotcha	"In this project, tests run with pnpm test, not npm test."
A decision and the why	decision	"We moved to SQS because of operating cost; our own queue took 6 hours a month."
A gotcha you discovered	gotcha	"Integration tests fail if the local database is not cleaned first."
A step-by-step that repeats	procedure	"How to publish a new version."
A fact that changes over time, with the date	fact	"Bruno has led Atlas since 2026-09-01."
Where the session stopped	handoff, with the handoff_save tool	"Three consumers left; start with the billing one."
An outside source (article, meeting notes)	in raw/, with ingest	"The article on queue costs on AWS."

Not worth remembering

• What is already in the code, in git, or in the documentation. The agent reads it at the source, which is always up to date.
• Details that only matter to the current conversation.
• Secrets. Passwords, tokens, keys, the contents of a .env. Never, not even summarized. Cabeção blocks the known formats (see Security), but it does not recognize everything.
• Other people's personal data without a need for it.

Habits that help

• Ask for a handoff at the end. "Before you stop, leave a handoff" makes the agent call handoff_save, and the next session starts where this one stopped. The handoff is valid for 30 days (expires_at) and the next one for the same project supersedes the previous one, because a stale handoff only confuses.
• Let remember handle repetition. Before writing, it looks for similar memories in the same scope and type: the same one is not written again, one that adds updates the existing one, and a fact that changed supersedes the old one, which stays in history. When the rule cannot decide, the agent gets the candidates and chooses. Still, ask for a search before saving something important.
• Run lint now and then. cabecao lint points out facts that contradict each other, broken links, memories nobody confirmed for months, and expired notes. For old memories: confirm the ones that hold (confirmed with today's date) and forget the ones that do not with cabecao forget. An old memory never confirmed already loses weight in search.
• Do not delete what changed: supersede it. When a decision changes, the new note carries supersedes: ["[[decisão-antiga]]"]. The old one stays there, expired, and answers questions about the past with --as-of.
• Use project scope. What only applies to one project goes with scope: project:<nome>, so it does not show up in the others.
• Review now and then. cabecao undo --list shows what agents wrote. Undo or fix whatever is wrong.
• Rules are few and yours. See Create rules.