Is Shared Memory Ruining Multi-Agent Systems? DecentMem Increases Accuracy by 24% and Saves Half the Tokens

According to DynaBeat monitoring, a team from the University of Cambridge and the University of Chicago has open-sourced the DecentMem multi-agent memory framework, replacing global shared memory with decentralized private memory. Traditional systems commonly use shared memory, but after agents read the same context, they often converge to similar decision paths, leading to the loss of division of labor advantage. The core idea of DecentMem is that collaboration must rely on cognitive differences, and maintaining private memory is essential to sustain complementary paths.

DecentMem allows agents to maintain dedicated dual-pool memory: the Experience pool (E-pool) stores historical experiences and reflection records, while the Exploration pool (X-pool) continuously generates new candidate ideas. The online decision-maker adjusts the dual-pool invocation weight dynamically based on the interim scores from the large language model referee, assisting the agents in autonomously balancing exploitation and exploration.

In theory, self-evolving search is modeled as a graph-based random walk, ensuring global reachability to enable agents to escape local optimal regions. In tests with AutoGen, DyLAN, and AgentNet, DecentMem achieved an average improvement of 8.6% over the strongest centralized memory baseline, with a relative improvement of up to 23.8% in optimal scenarios, while halving Token consumption. The research found that the more collaboration relies on spontaneous discussions rather than fixed processes, the more evident the advantages of decentralization. In the DyLAN framework, which emphasizes free negotiation, the system reduced the required number of iteration rounds by about 60% to achieve the same performance level and improved convergence speed by approximately 2.5 times.