Researchers from Cambridge and the University of Chicago have open-sourced DecentMem, a memory framework for multi-agent systems that replaces the standard shared-memory approach with decentralized, private memory. The core insight: traditional shared memory causes agents to converge on similar decision paths, effectively killing the benefits of specialization. DecentMem insists that real collaboration requires cognitive diversity — and that means keeping memory private.
DecentMem gives each agent a dual-pool memory: an exploitation pool (E-pool) that stores historical experiences and reflections, and an exploration pool (X-pool) that continuously generates new candidate strategies. An online decision-making mechanism dynamically adjusts the weighting of these two pools based on periodic evaluations from an LLM judge, letting agents autonomously balance exploitation and exploration.
Theoretically, the framework models self-evolutionary search as a random walk on a graph, whose global reachability helps agents escape local optima. In tests on AutoGen, DyLAN, and AgentNet, DecentMem beat the strongest centralized-memory baseline by an average of 8.6%, with a 23.8% improvement in the best case — all while cutting token consumption in half. The researchers also found that the more a collaborative task relies on improvisational discussion rather than fixed protocols, the bigger the advantage of decentralization. In DyLAN, which emphasizes free-form negotiation, the system needed about 60% fewer rounds to reach the same performance, speeding convergence by roughly 2.5x.