Effective memory mechanisms are essential for improving the performance of LLM‑driven multi‑agent systems (MAS). Most existing approaches rely on flat, unstructured memory stores that quickly become noisy and erase distinctions among agents. To tackle this, we introduce the notion of collective‑individual memory synergy and present the CoMem architecture, which unifies private experience with shared knowledge for multi‑agent learning.\ \ CoMem consists of three core components:\
- Private Experience Sedimentation – each agent continuously preserves and updates its own useful memories, preventing valuable information from being overwritten.\
- Collective Wisdom Curation – only ideas that have been widely validated across agents are written to the shared memory, limiting noise propagation.\
- Parallel Dual‑Stream Retrieval – during inference agents query both their private memory and the group memory; clustering ensures diverse retrieval results, balancing individual expertise with collective insight.\ \ Benchmarks on ALFWorld and PDDL show that CoMem consistently outperforms flat‑memory baselines in overall success rate, task efficiency, and resistance to memory pollution. The results demonstrate that synergizing collective and individual memories enables MAS to retain rich, reliable knowledge over long‑term evolution.\ \ Review