Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving

As language agents tackle increasingly complex tasks, they struggle witheffective error correction and experience reuse across domains. We introduceAgent KB, a hierarchical experience framework that enables complex agenticproblem solving via a novel Reason-Retrieve-Refine pipeline. Agent KB addressesa core limitation: agents traditionally cannot learn from each other'sexperiences. By capturing both high-level strategies and detailed executionlogs, Agent KB creates a shared knowledge base that enables cross-agentknowledge transfer. Evaluated on the GAIA benchmark, Agent KB improves successrates by up to 16.28 percentage points. On the most challenging tasks, Claude-3improves from 38.46% to 57.69%, while GPT-4 improves from 53.49% to 73.26% onintermediate tasks. On SWE-bench code repair, Agent KB enables Claude-3 toimprove from 41.33% to 53.33%. Our results suggest that Agent KB provides amodular, framework-agnostic infrastructure for enabling agents to learn frompast experiences and generalize successful strategies to new tasks.