M-Plan: Comprehensive Planning for LLM Agents with Memory Arrangement

Forum topic · 小凯 · 2026-10-22

Summary

M-Plan is a closed-loop multi-agent system (MAS) framework that integrates long-term and short-term memory to improve LLM-based agents on long-horizon tasks. While LLM agents excel at single-turn tasks, they struggle with extended multi-step tasks, especially in environments without prior training, due to weak continuous decision-making and reasoning over long contexts. Existing memory-equipped MAS systems rely mostly on direct retrieval, knowledge reuse, or summarization without adaptation to the current task, and retrieved memories often contain historical noise. M-Plan addresses this by combining global-local and past-current guidance: during task execution, agents receive task-specific milestones and insights, with local hints provided only when confusion arises. This guidance generation and usage forms short-term memory. After task completion, relevant information is refined or reorganized before being stored in long-term memory for future retrieval, all while keeping a plug-and-play design. Experiments on ALFWorld, Webshop, and PDDL multi-turn benchmarks show that M-Plan outperforms multiple baselines in both accuracy and efficiency.

Overview

Field: Machine Learning Authors: Ziyang Liu, Jiakang Yuan, Jiayuan Fan, Tao Chen Published: 2026-10-22 arXiv: 2610.27032

Background

Large language model (LLM)-based agents have recently demonstrated outstanding performance in single-turn tasks, yet they continue to struggle with long-horizon tasks, particularly in environments without prior training. A key limitation of single LLMs is their difficulty in processing long contexts, which stems from weaknesses in continuous decision-making and reasoning capabilities.

Problem with Existing Approaches

Existing multi-agent systems (MAS) with long-term memory bases can partially address this gap, but mostly rely on knowledge reuse or summarization.
This approach is constrained by direct retrieval without any adaptation for the current task.
Retrieved memory may contain much historical noise waiting to be preprocessed.
In contrast, MAS characterized by task-specific guidance tend to lack long-term memory specifically designed for that guidance.

Proposed Method: M-Plan

M-Plan is a closed-loop MAS framework integrating long-term and short-term memory, combining global-local and past-current guidance while maintaining a plug-and-play design:

During task execution, agents receive task-specific milestones and insights, constituting global guidance.
Agents obtain local hints only when confusion arises, constituting local guidance.
The generation and utilization of this guidance builds the short-term memory.
After task completion, relevant information is refined or reorganized before being sent to the long-term memory store for future retrieval.

Results

Extensive experiments on multi-turn tasks including ALFWorld, Webshop, and PDDL demonstrate that M-Plan outperforms multiple baseline methods in both accuracy and efficiency.

---

*Paper link: https://arxiv.org/abs/2610.27032*