New A-TMA System Improves LLM Agent Long-Term Memory Accuracy

Large Language Model (LLM) agents rely on long-term memory to function as persistent assistants, but managing evolving user facts presents a significant challenge. A common issue, termed "ghost memory," arises when outdated, current, and transitional information coexists and becomes muddled during retrieval, leading to incorrect responses. This new research proposes A-TMA, a state-aware framework that enhances existing memory systems. A-TMA meticulously tracks superseded and transitional records, constructing evidence packets tailored to the query's requested temporal state. It then exposes clear labels for current, historical, and transitional facts to the answer generation model, significantly reducing memory coordination failures. The authors also advocate for a decoupled evaluation approach, assessing memory bank maintenance, retrieval, and answer resolution separately, rather than relying solely on final QA accuracy. They developed a new benchmark, LTP, specifically for ghost memory conflicts, demonstrating that A-TMA substantially improves conflict accuracy and temporal F1 scores on various agent tasks.