AGENT KB: A Hierarchical Memory Framework for Cross-Domain Agentic Problem Solving

Xiangru Tang, Tianrui Qin, Tianhao Peng, Ziyang Zhou, Daniel Shao, Tingting Du, Xinming Wei, He Zhu, Ge Zhang, Jiaheng Liu, Xingyao Wang, Sirui Hong, Chenglin Wu, Wangchunshu Zhou

Keywords: LLM Agents, Workflow, Experience, Memory

Abstract:

As language agents tackle increasingly complex tasks, they struggle with effective error correction and knowledge reuse across different domains. We present Agent KB, a hierarchical memory framework that enables cross-domain agent learning through a novel \textcolor{cyan}{\texttt{Reason}}\textbf{-}\textcolor{purple}{\texttt{Retrieve}}\textbf{-}\textcolor{blue}{\texttt{Refine}} pipeline. Our dual-phase approach combines workflow-level knowledge retrieval with targeted execution pattern refinement, allowing agents to break free from limited reasoning pathways by incorporating diverse problem-solving strategies. Evaluations on GAIA benchmark demonstrate substantial performance gains, with Agent KB improving success rates by up to \textbf{16.28} percentage points overall. Most notably, on challenging tasks, Claude-3.7 with Agent KB increased performance from \texttt{38.46\%} to \texttt{57.69\%}, while GPT-4.1 showed similar improvements on intermediate tasks (\texttt{53.49\%} to \texttt{73.26\%}). For SWE-bench code repair tasks, our system significantly improved resolution rates, with Claude-3.7 achieving a \textbf{12.0} percentage point gain (\texttt{41.33\%} to \texttt{53.33\%}). Agent KB provides a modular, agent-agnostic infrastructure that facilitates continuous improvement through knowledge sharing across task boundaries and agent architectures.
Our code is publicly available at \url{https://anonymous.4open.science/r/agent_kb-35C6/}.

Submission Number: 44