Reproduce “A-TMA: Decoupling State-Aware Memory Failures in Long-Term Agent Memory” on OpenResearch

About this paper

Zitong Shi, Yixuan Tang, Anthony Kum Hoe Tung

Researchers at the National University of Singapore introduced A-TMA (Adaptive Temporal Memory Alignment), a state-aware overlay designed to decouple and manage evolving facts across the memory bank, retrieval, and QA levels in LLM agent memory systems. This approach significantly mitigates "ghost memory" failures, achieving substantial accuracy improvements on the LTP benchmark, with InsideOut+A-TMA gaining 0.505 in conflict accuracy and Graphiti/Zep+A-TMA gaining 0.240.

Key insights

• "Ghost memory" is a critical state coordination failure in LLM agent memory where old, current, and transitional facts intermingle, leading to misaligned retrieval and misleading answer generation.
• Decomposing memory failures into bank state role assignment, retrieval evidence construction, and QA state resolution is crucial for effectively diagnosing and mitigating the problem.
• Providing explicit state labels (e.g., "current memory," "historical memory") to the answer generation model, combined with a query-aware retrieval process, enables more accurate state-aware reasoning.

Summary by alphaXiv.