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

As LLM-based agents engage in longer, multi-session interactions, they face a fundamental trade-off: storing raw conversation history leads to noise and context window bloat, while aggressive summarization often discards critical, fine-grained details. AtomMem addresses this by asking: How can we build a memory system that maintains high information density, supports stable long-term evolution, and enables precise retrieval of relevant past experiences?

AtomMem introduces a modular architecture centered on atomic facts—self-contained, structured units of information. The system operates through four main components:

1. Fact Executor: A fine-tuned model that parses raw dialogue into structured atomic facts, performing coreference resolution and temporal anchoring to ensure each fact is independent and unambiguous.
2. Episodic Consolidation: Related atomic facts are grouped into hierarchical events, providing narrative context that simple fact-matching misses.
3. Temporal Profile Modeling: The system tracks evolving user attributes (e.g., preferences, habits) in a profile layer, using a session-based batch mechanism to update states while preserving historical versions.
4. Associative Memory Graph: During retrieval, AtomMem constructs a graph where facts are nodes connected by entity, event, and temporal edges. It uses a hybrid retrieval strategy—combining metadata filtering, semantic similarity, and random walks—to activate relevant memories across different sessions.

Evaluations on the LoCoMo and LongMemEval benchmarks demonstrate that AtomMem consistently achieves state-of-the-art performance. By replacing flat, noisy storage with a structured, graph-based representation, the system significantly improves the agent's ability to answer complex, multi-hop, and temporal questions. The results suggest that organizing memory into atomic facts is a scalable and economically viable strategy for building personalized, long-term assistants.