A-TMA resolves LLM agent ghost memory
Researchers proposed A-TMA, a state-aware overlay designed to resolve "ghost memory" failures in LLM agents by decoupling memory maintenance, retrieval, and resolution. The framework structures preserved facts into temporally labeled evidence packets, enabling agents to resolve conflicting timelines on the new LTP benchmark.
Trying to fix agent memory by simply deleting old facts is a short-sighted strategy that destroys historical context. A-TMA proves that keeping obsolete facts and dynamically labeling temporal states is the key to preventing "ghost memory" failures. Storing superseded and transitional records allows agents to handle queries about the past, providing a timeline-native approach to long-term memory.
Furthermore, by decoupling the memory architecture into maintenance, retrieval, and resolution phases, developers can systematically identify and fix memory failures rather than relying on noisy end-to-end task accuracy metrics. Retrofitting current state-of-the-art memory systems like Graphiti or Zep with a state-aware overlay like A-TMA shows significant performance improvements on temporal benchmarks.
DISCOVERED
1h ago
2026-07-06
PUBLISHED
2h ago
2026-07-06
RELEVANCE
AUTHOR
omarsar0
