Reproduce “StackPlanner: A Centralized Hierarchical Multi-Agent System with Task-Experience Memory Management” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruizhe Zhang, Xinke Jiang, Zhibang Yang, Zhixin Zhang, Jiaran Gao, Yuzhen Xiao, Tao Feng, Yue Fang and 7 more
Researchers at Peking University developed STACKPLANNER, a hierarchical multi-agent system that enhances long-horizon task performance and cross-task generalization through active task-memory management and structured experience memory for its central coordinator. The system achieved F1 score improvements of up to 8.54% on agentic benchmarks and improved report quality (47.47 RACE Overall) on open-ended research tasks, while also reducing token usage by approximately 45% and response latency by 67%.
Key insights
- Actively managing a central coordinator's task memory (e.g., via summarization and pruning) is effective in mitigating context bloat and error propagation, leading to more reliable decision-making in long-horizon tasks.
- Storing and retrieving structured historical coordination experiences (SOPs) and factual knowledge significantly improves a central agent's planning efficiency and generalization capabilities across new and diverse tasks.
- Formulating memory management as a controllable decision variable within a reinforcement learning framework allows for the joint optimization of high-level coordination and memory state, enhancing both performance and efficiency of multi-agent systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.