A July 2026 paper introduces A-TMA, a state-aware overlay for long-term agent memory systems that targets ghost memory, a failure mode where outdated, current, and transitional facts coexist and confuse retrieval. The method operates across bank maintenance, retrieval, and answer-resolution stages, retaining superseded records, assembling evidence packets for specific temporal states, and explicitly labeling facts as current, historical, or transitional. The authors built a new benchmark called LTP for temporal-conflict evaluation and show that layering A-TMA onto the Graphiti memory framework substantially improves temporal reasoning while exposing failures that standard QA-accuracy metrics miss.
