Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

StackPlanner: A Centralized Hierarchical Multi-Agent System with Task-Experience Memory Management

Ruizhe Zhang, Xinke Jiang, Zhibang Yang, Zhixin Zhang, Jiaran Gao, Yuzhen Xiao, Hongbin Lai, Xu Chu, Junfeng Zhao, Yasha Wang

TL;DR

StackPlanner tackles memory bottlenecks in centralized LLM-based multi-agent systems by introducing explicit task memory control and a reusable experience memory. It presents a hierarchical design with a central coordinator separated from specialized sub-agents, coupled with active task memory management and structured experience memory, all learned via reinforcement learning. Key contributions include decoupled coordination, memory condensation and pruning mechanisms, and a three-part experience memory (user profiles, semantic memory, SOPs) with retrieval to improve cold-start and cross-task generalization. Experiments across multi-hop QA and agentic benchmarks show state-of-the-art performance and strong generalization, with ablations confirming the value of both memory components.

Abstract

Multi-agent systems based on large language models, particularly centralized architectures, have recently shown strong potential for complex and knowledge-intensive tasks. However, central agents often suffer from unstable long-horizon collaboration due to the lack of memory management, leading to context bloat, error accumulation, and poor cross-task generalization. To address both task-level memory inefficiency and the inability to reuse coordination experience, we propose StackPlanner, a hierarchical multi-agent framework with explicit memory control. StackPlanner addresses these challenges by decoupling high-level coordination from subtask execution with active task-level memory control, and by learning to retrieve and exploit reusable coordination experience via structured experience memory and reinforcement learning. Experiments on multiple deep-search and agent system benchmarks demonstrate the effectiveness of our approach in enabling reliable long-horizon multi-agent collaboration.

StackPlanner: A Centralized Hierarchical Multi-Agent System with Task-Experience Memory Management

TL;DR

StackPlanner tackles memory bottlenecks in centralized LLM-based multi-agent systems by introducing explicit task memory control and a reusable experience memory. It presents a hierarchical design with a central coordinator separated from specialized sub-agents, coupled with active task memory management and structured experience memory, all learned via reinforcement learning. Key contributions include decoupled coordination, memory condensation and pruning mechanisms, and a three-part experience memory (user profiles, semantic memory, SOPs) with retrieval to improve cold-start and cross-task generalization. Experiments across multi-hop QA and agentic benchmarks show state-of-the-art performance and strong generalization, with ablations confirming the value of both memory components.

Abstract

Multi-agent systems based on large language models, particularly centralized architectures, have recently shown strong potential for complex and knowledge-intensive tasks. However, central agents often suffer from unstable long-horizon collaboration due to the lack of memory management, leading to context bloat, error accumulation, and poor cross-task generalization. To address both task-level memory inefficiency and the inability to reuse coordination experience, we propose StackPlanner, a hierarchical multi-agent framework with explicit memory control. StackPlanner addresses these challenges by decoupling high-level coordination from subtask execution with active task-level memory control, and by learning to retrieve and exploit reusable coordination experience via structured experience memory and reinforcement learning. Experiments on multiple deep-search and agent system benchmarks demonstrate the effectiveness of our approach in enabling reliable long-horizon multi-agent collaboration.
Paper Structure (32 sections, 4 equations, 2 figures, 4 tables)

This paper contains 32 sections, 4 equations, 2 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Hierarchical Coordination
  4. Central Coordinator Action Space
  5. Specialized Sub-Agents
  6. Active Task Memory Management
  7. Structured Experience Memory Utilization
  8. Experience Retrieval
  9. Reinforcement Learning Formulation
  10. Experiment
  11. Experimental Setup
  12. Main Result Analysis
  13. Strong Performance Compared with Baselines.
  14. Component Analysis
  15. Model Component Ablation.
  16. ...and 17 more sections

Figures (2)

  • Figure 1: Overview of StackPlanner framework.
  • Figure 2: Case Study of StackPlanner on a deepresearch task. Task: "Please summarize the recently popular multi-agent system frameworks that are capable of performing report generation tasks."