MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems

Researchers introduced MemTrace, a tracing framework that turns LLM memory pipelines into executable graphs for debugging.

They also released MemTraceBench, a benchmark spanning multiple memory systems such as RAG, Long-Context, Mem0, and EverMemOS.

An automated fault-attribution method helps pinpoint root causes of memory failures, including retrieval misalignment.

These signals can guide prompt optimization and improve downstream task performance by up to 7.62%.