DecentMem boosts multi-agent AI accuracy by 23.8% with decentralized memory

Multi-agent systems (MAS) are increasingly used for complex tasks, but they typically rely on a centralized memory repository shared across agents. This approach creates communication bottlenecks, privacy risks, and homogenization of agent behavior. In a new paper, researchers introduce DecentMem, a decentralized memory framework that gives each agent its own dual-pool memory: an exploitation pool storing consolidated past trajectories, and an exploration pool containing LLM-generated candidates for unseen contexts. The two pools are dynamically reweighted using stage-wise feedback from an LLM-as-a-judge. The team also provides theoretical proofs showing DecentMem guarantees global reachability of the solution space and achieves O(log T) cumulative regret, matching the stochastic bandit lower bound up to constants.

Empirically, DecentMem was tested across three MAS frameworks (AutoGen, DyLAN, AgentNet), three Qwen3 backbones (4B/8B/14B), two Gemma4 backbones (E2B/E4B), and five benchmarks covering math, code, QA, and embodied tasks. Results show an average accuracy improvement of up to 23.8% over the strongest centralized memory baseline and up to 52.5% over the no-memory baseline, while reducing token usage by up to 49%. These results demonstrate that decentralized memory not only preserves agent diversity and privacy but also significantly boosts performance and efficiency. DecentMem represents a major step toward truly self-evolving, scalable multi-agent systems that can continuously improve from experience.