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

StackPlanner is a hierarchical multi-agent system designed to enhance long-horizon collaboration in centralized architectures by incorporating explicit memory management. It features active task-level memory control and structured experience memory that allows agents to retrieve and utilize past coordination experiences, leveraging reinforcement learning for improved performance. This framework addresses issues of context bloat and error accumulation, making it significant for practitioners aiming to build more efficient and reliable multi-agent systems in complex tasks.