AtomMem Extracts Atomic Facts for Stable Long-Term LLM Agent Memory

AtomMem decomposes multi-session interactions into atomic facts, organizes them into hierarchical event structures and temporal user profiles, and retrieves via an associative memory graph. The approach targets the instability of unconstrained memory updates common in existing systems. No benchmark numbers are provided in the abstract.