Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory
Existing LLM agent memory relies on extracting atomic facts, which causes loss of detail and limits deep reasoning. TriMem proposes a new system with three granularities: raw dialogue segments for storage fidelity, extracted facts for retrieval, and synthesized profiles for holistic semantic understanding. This architecture uses prompt optimization to evolve its extraction and profiling methods over time.