AI Agent Memory: What to Store, Retrieve and Forget

Most people do not want an agent to remember everything. They want it to remember the few things that save them from repeating themselves, and to forget what has become irrelevant. The difference matters. A memory store that mixes preferences, private conversation, business records, guesses, and audit history will eventually remember the wrong thing with great confidence.

Start with the kind of memory you need

The current task needs short-lived working state: the case being discussed, the next step, and the tools already used. A user’s explicit language preference may be worth keeping. A product price or account balance should normally be read from the source of truth when needed. An approval belongs in an audit record, where it can be found and reviewed, not in a casual profile note.

Store a fact with its source, owner, timestamp, and expiry where appropriate. “The customer asked for email updates” is a useful memory when the customer actually said it. “The customer is probably unhappy” is an interpretation. It may help for the current conversation, but it should not quietly become a permanent label.

Memory still has to respect the person asking

A lookup should know the tenant, user, role, purpose, and time range. Check access before the result reaches the model. If two memories disagree, show the conflict or prefer the newer authoritative record; do not let the model choose whichever sentence sounds most convincing.

Forgetting is part of the feature

A delete request is not complete if an old embedding, summary, cache, or copied transcript can still surface the information. Give temporary memories an expiry date and test outdated preferences, contradictory facts, cross-tenant requests, and a person asking the agent to forget something.

Good memory is quiet. It shortens the path to a useful answer while staying honest about what is a preference, what is an inference, and what must be checked in the real system.