Can agents fix their own memory systems?

Long-horizon agents struggle when their memory systems—the mechanisms that retrieve and use past information—stay rigid after deployment. MemPro treats the entire memory pipeline as evolvable code: an agent diagnoses its own failures, edits the pipeline structure itself, and tests improved versions in a version tree. On LongMemEval, LoCoMo, HotpotQA, and NarrativeQA, this beats static memory systems and prompt-tuning baselines within a few iterations, with ongoing gains. Code released.