Researchers have trained AI agents to summarize and prune their own working memory, and the results beat the standard approach of tacking on an external retrieval system.
The paper introduces MemPO, short for Self-Memory Policy Optimization. Most methods for handling long-running agent tasks work the same way: store everything in an external memory module and retrieve relevant chunks when the model needs them. The model itself stays passive. MemPO takes a different route, training the underlying policy to decide what to keep, summarize, or discard during interaction. A credit assignment mechanism rewards the model when its memory choices actually help task completion. Across benchmarks, MemPO improved F1 scores by 25.98 points over the base model and 7.1 points over the prior best, while cutting token consumption by 67.58% and 73.12% respectively.
Token count maps directly to cost and latency. A 70% reduction is not a research curiosity; it is the difference between an agent pipeline that scales and one that does not. The deeper shift is philosophical: instead of treating memory as an engineering problem you solve by adding infrastructure, MemPO treats it as something the model should learn to handle itself.
The approach will not displace external retrieval overnight. Enterprise teams have invested heavily in vector stores and RAG pipelines, and those systems have real advantages in transparency and auditability. But if these numbers survive scrutiny outside the benchmark lab, "train the model to remember better" is a harder argument to dismiss.