Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving
🤖 AI Summary
Language agents face bottlenecks in complex cross-domain tasks, particularly in error correction and experience reuse. This paper proposes a hierarchical experience framework that establishes the first shared knowledge base enabling cross-agent experience transfer, unifying high-level strategies and low-level execution details through semantic representation and retrieval. It introduces a Reason-Retrieve-Refine pipeline that iteratively refines agent behavior by integrating large language model reasoning with log-based retrospective analysis. The core contribution is breaking down experience silos among agents and establishing a scalable, collaborative learning mechanism. Experiments demonstrate substantial improvements: up to +16.28 percentage points on the GAIA benchmark; significant gains in multi-task performance for Claude-3 and GPT-4; and an increase in code repair success rate on SWE-bench from 41.33% to 53.33%.
Application Category
📝 Abstract
As language agents tackle increasingly complex tasks, they struggle with effective error correction and experience reuse across domains. We introduce Agent KB, a hierarchical experience framework that enables complex agentic problem solving via a novel Reason-Retrieve-Refine pipeline. Agent KB addresses a core limitation: agents traditionally cannot learn from each other's experiences. By capturing both high-level strategies and detailed execution logs, Agent KB creates a shared knowledge base that enables cross-agent knowledge transfer. Evaluated on the GAIA benchmark, Agent KB improves success rates by up to 16.28 percentage points. On the most challenging tasks, Claude-3 improves from 38.46% to 57.69%, while GPT-4 improves from 53.49% to 73.26% on intermediate tasks. On SWE-bench code repair, Agent KB enables Claude-3 to improve from 41.33% to 53.33%. Our results suggest that Agent KB provides a modular, framework-agnostic infrastructure for enabling agents to learn from past experiences and generalize successful strategies to new tasks.
Problem
Research questions and friction points this paper is trying to address.
Enables cross-domain experience reuse for agents
Improves error correction in complex agent tasks
Facilitates knowledge transfer between language agents
Innovation
Methods, ideas, or system contributions that make the work stand out.
Hierarchical experience framework for problem solving
Reason-Retrieve-Refine pipeline for error correction
Shared knowledge base for cross-agent transfer
🔎 Similar Papers
No similar papers found.
