MemCo is a memory‑centric collaboration framework that enables large language model (LLM) agents to make effective sequential decisions in interactive environments they have never seen before. The system maintains two complementary memory spaces: a local memory that stores environment‑specific observations, actions, and feedback, and a global memory that extracts transferable workflows from the local trajectories of many agents. During online interaction, MemCo routes relevant entries from both memories based on the agent's current state and decision phase, allowing the language model to reuse experiences of other agents without blindly copying environment‑specific details. Experiments on several interactive decision‑making benchmarks show that, compared with isolate‑memory and shared‑memory baselines, MemCo improves task success rates and markedly reduces redundant exploration. The code is released at https://github.com/SYannL/nvdamas.
Review: By combining hierarchical memory organization with state‑aware retrieval, MemCo strikes a practical balance between fine‑grained environmental context and cross‑task generality, offering a promising route for zero‑shot transfer of LLM agents.