A July 2026 paper introduces NapMem, a framework that treats long-term user memory as a structured action space to be navigated rather than context passively retrieved. The system organizes conversation history into a multi-level memory pyramid linking raw conversations, typed records, topic tracks, and user profiles through provenance relations, and trains an agent with reinforcement learning to choose which memory level and granularity to query for a given task. The authors report the approach stays competitive on memory-intensive benchmarks while preserving the underlying model’s general reasoning ability.
