NapMem Enables Active Memory Navigation for Conversational Agents.

Personalized conversational agents require robust long-term user memory to maintain context and provide relevant responses. However, many existing memory systems treat memory as a passive resource, simply retrieving pre-selected evidence for the model to consume. This research introduces NapMem, a framework designed to enable agents to actively navigate and utilize memory. NapMem structures user history into a "memory pyramid" with multiple granularities, linking raw conversations, typed memory records, topic tracks, and user profiles through provenance relations. This hierarchical structure is then exposed to the agent via memory tools. Instead of passively receiving information, the agent is trained to actively select and inspect different memory granularities based on the current query and intermediate evidence, allowing for more informed and personalized responses. Experiments on several memory-intensive benchmarks (PersonaMem-v2, LongMemEval, LoCoMo) show that a NapMem agent, trained with memory-tool reinforcement learning, achieves competitive performance. Crucially, evaluations on non-memory tasks suggest that this learned policy largely preserves the agent's general reasoning and tool-use abilities. Further analyses delve into storage, inference costs, tool-use behavior, and the impact of navigation and granularity, indicating that coupling structured storage with a learned policy for memory use significantly benefits long-term user memory in conversational agents.