Researchers have built a framework that lets AI agents actively search their own memory instead of waiting for the right chunk to surface automatically.
A newly published paper introduces NapMem, a system that organizes a user's conversation history into what the authors call a "multi-granularity memory pyramid" — stacking raw chats, typed memory records, topic summaries, and user profiles in linked layers. Instead of handing the agent a pre-fetched snippet and hoping it is the right one, NapMem exposes these layers as callable tools. The agent is trained via reinforcement learning to pick which layer to consult based on the query and what it has already found. Benchmarks on PersonaMem-v2, LongMemEval, and LoCoMo show the approach is competitive on memory-heavy tasks without degrading the agent's general reasoning skills.
Most memory systems for AI chatbots work like a search engine: query in, relevant chunk out, done. That breaks down when the right answer depends on context spread across multiple abstraction levels — a raw conversation from months ago, a mid-level topic summary, and a high-level user profile all at once. NapMem's navigation approach targets that mismatch directly, and because the retrieval policy is learned, it can in theory be tuned for specific tasks.
The paper is academic, and "competitive on a benchmark" is a long way from deployed in production. But framing memory as an action space rather than a lookup table is a conceptual shift that teams building personalized AI assistants will likely adopt regardless of whether NapMem itself ships.