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

2026-06-19 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

AtomMem is a novel long-term memory system for LLM agents, designed to overcome the limitations of fixed context windows and unstable memory representations. It introduces a Fact Executor that extracts high-value atomic facts from interactions, organizing them into hierarchical event structures and temporal profiles. During retrieval, AtomMem activates an associative memory graph to connect fragmented memories. Evaluated on the LoCoMo benchmark, AtomMem achieves state-of-the-art performance across various reasoning tasks, including Multi-Hop and Temporal tasks, with a 5.5% and 31.1% J-score improvement over LightMem, respectively. It also demonstrates competitive cost efficiency, reducing total token consumption by approximately 61.4% compared to Mem0. The system uses a Qwen3-14B model fine-tuned with SFT LoRA as its fact extractor and all-minilm-L6-v2 for embeddings.

Key takeaway

For AI Engineers building long-term conversational agents, AtomMem offers a robust solution to memory limitations. You should consider adopting its atomic fact-based, hierarchical memory architecture to improve reasoning accuracy and reduce operational costs. Implementing a fine-tuned fact extractor and graph-based associative recall can significantly enhance your agent's ability to maintain consistent user understanding across extended interactions, making personalized agents more scalable and economically viable.

Key insights

AtomMem enhances LLM agent memory by structuring interactions into atomic facts, hierarchical events, and temporal profiles for efficient, associative retrieval.

Principles

• Atomic facts are superior memory units.
• Hierarchical memory improves context.
• Graph-based recall enhances associations.

Method

AtomMem extracts atomic facts via an SFT-tuned Fact Executor, organizes them into events and temporal profiles, then uses a memory graph with entity, event, and temporal edges for hierarchical, associative retrieval.

In practice

• Fine-tune LLMs for fact extraction.
• Use hybrid similarity for fact verification.
• Implement graph-based associative recall.

Topics

• LLM Agents
• Long-Term Memory
• Atomic Facts
• Graph-Based Retrieval
• Conversational AI
• Computational Efficiency

Code references

• MINE-USTC/AtomMem

Best for: NLP Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• AtomMem: Building Simple and Effective Memory System for LLM Agents via Atomic Facts — AtomMem enhances LLM agent memory by extracting atomic facts and organizing them hierarchically for stable, efficient long-term recall.
• T-Mem: Memory That Anticipates, Not Archives — T-Mem introduces a long-term memory architecture enabling both descriptive and associative recall for conversational agents.
• Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents — LLM agent memory should be actively reconstructed and dynamically adapted to reasoning context, not statically retrieved.
• NextMem: Towards Latent Factual Memory for LLM-based Agents — NextMem uses an autoregressive autoencoder and two-stage training to create efficient, reconstructible latent factual memory for LLM agents.
• How xMemory cuts token costs and context bloat in AI agents — xMemory uses a hierarchical memory structure and uncertainty-gated retrieval to optimize long-term AI agent coherence and reduce token costs.
• From raw interaction to reusable knowledge: Rethinking memory for AI agents — Structuring AI agent memory into reusable knowledge units improves performance and efficiency across diverse tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.