MemPO: Self-Memory Policy Optimization Cuts Agent Token Use 67-73% While Adding 25.98% F1 on Long-Horizon Tasks

arXiv·medium signalXBlueskyLinkedInCopy link

MemPO replaces external memory retrieval modules with a policy-level mechanism where agents autonomously summarize and selectively retain their own memory during task execution, improving credit assignment based on what information actually proved useful. On long-horizon benchmarks, MemPO achieves 25.98% F1 improvement over the base model and 7.1% over prior best methods while cutting token consumption by 67-73% through selective retention. The approach shifts the memory management burden from retrieval systems to the model itself, representing a fundamentally different architecture for persistent agent context.