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

MemTrace turns LLM memory pipelines into executable evolution graphs for fine-grained error tracing.

The team also built MemTraceBench from representative systems like Long-Context, RAG, Mem0, and EverMemOS.

An automatic attribution method identifies root causes such as retrieval misalignment and information loss.

The framework can guide prompt optimization, making LLM memory systems easier to debug and more reliable on long-horizon tasks.