Teams often describe agent memory as a cache for context. It is an easy mental model and it is wrong in ways that matter.

What a cache gets wrong

A cache assumes the underlying data is the source of truth and the copy is disposable. If the copy is stale, you evict it and fetch again. Memory has no such source. The decision someone made in a call exists nowhere else.

That changes the design. Memories need provenance, a record of who said it and when, because the only way to judge a stale memory is to see where it came from.

Designing for change

We store every memory with a timestamp, a source and a confidence. When two memories disagree, the newer one wins by default, and both stay visible so a person can see the history.

Forgetting on purpose

Good memory also forgets. Expired policies, retired customers and revoked access should fade out of retrieval without being deleted from the audit trail.