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

MemTrace introduces a framework to debug Large Language Model memory systems by turning them into executable graphs. It traces every operation to find the exact faulty step causing an error, moving beyond simple reporting. This allows researchers to pinpoint specific issues in memory architectures and create targeted fixes through prompt optimization.

Open the full episode

More from AI Radio