Large language model (LLM) based multi-agent systems achieve impressive results across many tasks, yet their memory and collaboration structures are usually static, limiting continual adaptation. Prior work evolves memory or collaboration separately from interaction experience, which can lead to misalignment: an evolved workflow may require experiences the memory does not provide, and evolved memory may encode guidance the workflow cannot use. CoEvolve addresses this by introducing a skill‑centered framework that aligns memory and collaboration co‑evolution. Reusable local collaboration patterns are abstracted as skills that bind evolving collaboration decisions to their supporting memories. By contrasting original and counterfactual trajectories, CoEvolve extracts local evolution signals that coordinate updates to the skill library, skill memory, and executor memories each round. Experiments on the ALFWorld and WebShop benchmarks show CoEvolve achieves the best overall performance among baselines, and ablation studies confirm the benefits of aligned co‑evolution and counterfactual‑guided evolution.
Review