[Shared Memory is Destroying Multi-Agent Systems? DecentMem Improves Accuracy by 24% While Saving Half the Tokens] According to Deep Tide TechFlow, on June 14, in multi-agent systems, shared memory was originally considered an important mechanism for improving collaboration efficiency. However, the latest research indicates that this design may instead introduce issues such as information pollution and context interference, leading to a decline in overall reasoning stability. To address this issue, DecentMem has proposed a more structured memory management approach. By filtering, compressing, and hierarchically invoking memory, it reduces the spread of irrelevant information between agents, thereby lowering noise interference. Experimental results show that this method can improve overall accuracy in multi-agent tasks by approximately 24% while reducing token consumption by about half, meaning it significantly lowers reasoning costs while enhancing performance. Industry experts believe this direction could influence the design of future agent architectures.