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

Researchers have introduced AtomMem, a long-term memory system for large language models (LLMs) that extracts and organizes 'atomic facts' from interactions to improve information retention and recall.

AtomMem addresses the limitation of LLMs' fixed context windows, which hinder long-term information accumulation. The system uses a Fact Executor to extract high-value atomic facts from long-form interactions^[1]. These facts are then organized into hierarchical event structures and temporal profiles to capture episodic contexts. During retrieval, an associative memory graph connects fragmented memories. A separate study on agent memory from a data management perspective found that existing evaluations focus on end-to-end task success metrics, neglecting system-level concerns^[2]. The study proposed an analytical framework that decomposes agent memory into four core modules and found that no single memory architecture dominates across all scenarios. AtomMem's design aligns with this finding, offering a scalable and economically viable solution for deploying intelligent personalized agents. Experiments on the LoCoMo benchmark confirmed that AtomMem achieves state-of-the-art performance across various reasoning tasks^[1].