Collaborative Multi-Agent Optimization for Personalized Memory System

A research team led by Wenyu Mao has introduced CoMAM (Collaborative Reinforcement Learning Framework for Multi-Agent Memory Systems), addressing a critical limitation in personalized LLM memory systems. Current approaches optimize individual agents independently through prompt engineering or fine-tuning, but this often fails to translate to better overall system performance. The new framework treats the execution of multiple specialized agents—handling tasks like memory construction and retrieval—as a sequential Markov Decision Process (MDP), embedding inter-agent dependencies directly into the state transitions.

CoMAM introduces a novel reward mechanism that combines both local task rewards (like information coverage during memory construction) and global rewards (such as query-answer accuracy). It quantifies each agent's contribution using group-level ranking consistency between these reward types, creating adaptive weights to assign global credit. This allows the system to integrate local and global rewards, ensuring that improvements at the agent level directly enhance the overall memory system's effectiveness. Experimental results demonstrate that this collaborative optimization approach significantly outperforms existing leading memory systems.

The framework represents a shift from isolated agent optimization to holistic system design, where reinforcement learning coordinates multiple specialized components. By formalizing the collaboration problem and providing measurable contribution metrics, CoMAM enables more efficient training of complex multi-agent systems. This advancement is particularly relevant for creating AI assistants that can maintain coherent, personalized memory across extended interactions, moving beyond the limitations of fixed context windows.