NapMem reframes memory as structured action space
To address the limitations of passive context retrieval in conversational agents, the authors introduce NapMem, a framework that reframes user memory as a multi-granularity pyramid and treats memory access as a structured action space. Using reinforcement learning, the agent learns an active navigation policy to iteratively retrieve memory at the correct granularity for a given task.
Passive retrieval is a major bottleneck for agentic systems; treating memory as a controllable action space trained via reinforcement learning is a much-needed step toward genuine cognitive navigation in LLMs.
* The hierarchical memory pyramid provides structured abstractions from raw chats to high-level user profiles.
* Active tool-use empowers agents to iteratively query their own memory to verify evidence before answering.
* Training via reinforcement learning optimizes the navigation policy to minimize context bloat and hallucination.
DISCOVERED
1h ago
2026-07-08
PUBLISHED
3h ago
2026-07-08
RELEVANCE
AUTHOR
omarsar0