Rethinking How to Remember: Beyond Atomic Facts in Lifelong LLM Agent Memory

Researchers introduced TriMem, a three-level memory system for LLM agents.

It stores source-linked dialogue segments, atomic facts, and higher-level semantic profiles to balance detail and abstraction.

TextGrad is used to iteratively refine extraction and profiling prompts, helping the system improve over time without retraining.

TriMem outperformed baseline memory methods on LoCoMo and PerLTQA, suggesting stronger retrieval and reasoning for long-term agent memory.