Abstract
LLM agents enhanced with persistent memory can recall past interactions, but existing systems suffer from two main limitations: flat, unstructured storage loses the relational context required for multi-hop and temporal reasoning, and reliance on expensive LLM-based classification makes them impractical for latency-sensitive deployment. Moreover, these systems lack mechanisms to validate new information against stored knowledge, leading to a silent accumulation of contradictions.
We present MOSAIC (Memory-Organized Structured Agent for Information Collection), a structured, conflict-aware long-term memory framework for LLM agents that significantly enhances accuracy and efficiency. MOSAIC introduces three key capabilities:
- Entity-typed graph storage: Semantic classification preserves the relational structure across events, personas, and relationships, enabling multi-hop and temporal reasoning over conversation history.
- Hash-accelerated dual-path retrieval: Replacing LLM-based classification with locality-sensitive hashing achieves near-instantaneous lookup with negligible accuracy loss.
- Active conflict detection: At save time, new information is cross-referenced against existing graph neighbors, triggering updates or deletions for contradictory entries.
Evaluated on LoCoMo (long-conversation QA), HaluMem, and a novel clinical-guideline error compounding test, MOSAIC achieves 89.35% accuracy on LoCoMo (+27.21 pp over the best baseline), best HaluMem-Medium extraction F1 (86.77%) and HaluMem-Long extraction F1 (85.84%), best QA correctness on both Medium and Long (73.10%, 70.75%), and detects 66% of injected factual conflicts - 4.7 times higher than the best baseline (14%) - while hash-accelerated retrieval keeps average search latency at 0.58 s per question.
Blogger's Review: The introduction of the MOSAIC framework represents a groundbreaking solution for long-term memory in LLM agents. Particularly in contexts where complex reasoning and rapid response demands are increasing, its structured storage and conflict detection mechanisms are crucial, and it is expected to significantly enhance the performance of intelligent assistants in future applications.