MOSAIC Boosts LLM Agent Memory Accuracy and Efficiency

Large Language Model (LLM) agents often struggle with long-term memory, as existing systems typically rely on flat storage that loses relational context and expensive LLM-based classification for retrieval. This leads to poor multi-hop and temporal reasoning, along with the silent accumulation of contradictory information. A new framework, MOSAIC (Memory-Organized Structured Agent for Information Collection), addresses these limitations by providing a more accurate and efficient long-term memory solution. MOSAIC introduces three key innovations: first, it uses entity-typed graph storage to preserve relational structure across events, personas, and relationships, enabling sophisticated reasoning over conversation history. Second, it replaces slow LLM-based classification with hash-accelerated dual-path retrieval, achieving near-instantaneous lookups with minimal accuracy loss. Third, MOSAIC incorporates active conflict detection at save time, cross-referencing new information against existing knowledge to trigger updates or deletions for contradictions. Evaluated on several benchmarks, MOSAIC achieved 89.35% accuracy on long-conversation QA, significantly outperforming baselines, and detected 66% of injected factual conflicts—4.7 times higher than the best alternative—all while maintaining an average search latency of 0.58 seconds per question.