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

Researchers introduced TriMem, a three-tier long-term memory system for LLM agents.

It stores source-linked raw dialogue, extracts atomic facts for efficient lookup, and builds synthesized profiles for higher-level reasoning.

The system also uses TextGrad to iteratively refine prompts, improving how memory is used over time.

TriMem outperformed prior baselines on LoCoMo and PerLTQA, suggesting stronger support for long-horizon interactions without retraining.