Large language model (LLM) agents excel at single‑turn tasks but still struggle with long‑horizon tasks, especially in environments without prior training. A single LLM cannot handle long contexts well because its continuous decision‑making and reasoning abilities are limited. Existing multi‑agent systems (MAS) with long‑term memory mitigate this issue, yet they mostly rely on knowledge reuse or summarization, retrieving information without adapting it to the current task and often pulling in noisy historical data that requires preprocessing. Conversely, MAS that are guided by task‑specific instructions usually lack a dedicated long‑term memory for that guidance.
To address these gaps we introduce M‑Plan, a closed‑loop MAS framework that integrates long‑term and short‑term memory. By combining global‑local and past‑current guidance, M‑Plan improves agent performance on multi‑turn tasks while preserving a plug‑and‑play design. During execution, agents receive task‑specific milestones and insights; local hints are supplied only when confusion arises, and together they form global‑local guidance that also constructs short‑term memory. After a task finishes, relevant information is optimized or reorganized before being stored in long‑term memory for future retrieval. This design differs from prior works that focus on either long‑term or short‑term memory alone, offering a comprehensive memory‑cooperation framework.
Extensive experiments on multi‑turn benchmarks such as ALFWorld, Webshop, and PDDL across diverse models demonstrate that M‑Plan consistently outperforms multiple baselines in both accuracy and efficiency.
Review: M‑Plan’s closed‑loop memory interaction enables effective coordination of long‑ and short‑term memories, delivering more robust reasoning and decision‑making in complex multi‑turn scenarios and providing a practical pathway for deploying LLM agents in real‑world tasks.