arXiv cs.AI
· Papers
MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems
arXiv:2605.28732v3 Announce Type: replace-cross Abstract: Memory is essential for enabling large language models to support long-horizon reasoning, yet existing memory systems remain unreliable and difficult to debug. Tracing memory's dynamic evolution is crucial to understand how information is synthesized, propagated