MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems
MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems
要約
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, or corrupted over time. In this work…