A field guide to agent memory
What an agent should remember, what it should look up, and what it should be told to forget.
Nimind Team
Academy
Memory is one word doing four jobs. Untangle them and most of the design decisions become obvious, including the ones that stop a project from collapsing under its own context window.
Four layers
- Session memory — the current branch of work. Cheap to grow, dies with the session.
- Project memory — decisions, voice and the reasons behind them. Written by people, read by agents.
- Company memory — policies, roles and org. In the repo, with everything else.
- Lookup — anything that can be fetched fresh, which is most facts.
Most “the agent forgot” complaints are lookup failures, not memory failures. The agent never knew the price list; it guessed from an old document because nobody pointed it at the new one.
Write decisions, not transcripts
A transcript is a haystack. A decision is a needle. “We are not localising the help center this quarter, because support is staying English-first until March” is worth a thousand tokens of meeting. It is also the kind of line an agent can act on.
Forgetting is a feature
Every layer needs an expiry. A project note about a deadline in March is noise in June. Session memory should end with the session; project memory should be reviewed when the project changes shape; company memory should be edited, not appended to forever.
The practical test: if a new hire would not need it on their first day, it probably does not belong in the repo.
Published September 10, 2026 · Nimind Team
Back to all posts
