AtomMem: Building Simple and Effective Memory System for LLM Agents via Atomic Facts

AtomMem is a newly proposed long-term memory system for large language models (LLMs) that addresses limitations in fixed context windows and inefficient memory management. It features a Fact Executor for extracting high-value atomic facts, which are organized into hierarchical structures and temporal profiles, enabling coherent episodic context retrieval through an associative memory graph. Experimental results on the LoCoMo benchmark demonstrate that AtomMem achieves state-of-the-art performance in reasoning tasks, providing a scalable solution for building intelligent, personalized AI agents.