DecentMem Increases Accuracy by 24% and Reduces Token Usage in Multi-Agent Systems

According to monitoring by Beating, a team from the University of Cambridge and the University of Chicago has open-sourced the multi-agent memory framework DecentMem, which replaces global shared memory with decentralized private memory. Traditional systems commonly rely on shared memory, but when agents access the same context, they often converge toward similar decision paths, eliminating the advantages of division of labor. The core idea of DecentMem is that collaboration must depend on cognitive diversity—preserving private memory is essential to maintaining complementary pathways. DecentMem enables agents to maintain dedicated dual-pool memories: an Experience Pool (E-pool) that stores historical experiences and reflective records, and an Exploration Pool (X-pool) that continuously generates new candidate ideas. An online decision-maker dynamically adjusts the weighting of calls to each pool based on阶段性 scores from a large language model evaluator, helping agents autonomously balance exploitation and exploration. Theoretically, self-evolving search is modeled as a random walk on a graph, ensuring global reachability and enabling agents to escape local optima. In tests on AutoGen, DyLAN, and AgentNet, DecentMem outperformed the strongest centralized memory baseline by an average of 8.6%, with peak improvements reaching 23.8%, while halving token consumption. The study found that the greater the reliance on spontaneous discussion rather than fixed procedures, the more pronounced the advantages of decentralization; in the DyLAN framework, which emphasizes free negotiation, the system required approximately 60% fewer iterations to achieve equivalent performance and converged about 2.5 times faster.