Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving (2507.06229v5)

Abstract: AI agent frameworks operate in isolation, forcing agents to rediscover solutions and repeat mistakes across different systems. Despite valuable problem-solving experiences accumulated by frameworks like smolagents, OpenHands, and OWL, this knowledge remains trapped within individual systems, preventing the emergence of collective intelligence. Current memory systems focus on individual agents or framework-specific demonstrations, failing to enable cross-architecture knowledge transfer. We introduce AGENT KB, a universal memory infrastructure enabling seamless experience sharing across heterogeneous agent frameworks without retraining. AGENT KB aggregates trajectories into a structured knowledge base and serves lightweight APIs. At inference time, hybrid retrieval operates through two stages: planning seeds agents with cross-domain workflows, while feedback applies targeted diagnostic fixes. A disagreement gate ensures retrieved knowledge enhances rather than disrupts reasoning, addressing knowledge interference in cross-framework transfer. We validate AGENT KB across major frameworks on GAIA, Humanity's Last Exam, GPQA, and SWE-bench. Results show substantial improvements across diverse model families: compared to baseline pass@1, smolagents with AGENT KB achieve up to 18.7pp gains at pass@3 (55.2% -> 73.9%), while OpenHands improves 4.0pp on SWE-bench pass@1 (24.3% -> 28.3%). Similar improvements are observed across all base model families. Ablations confirm that hybrid retrieval and feedback stages are essential, with automatically generated experiences matching manual curation. This establishes the foundation for collective agent intelligence through shared memory infrastructures.

• The paper demonstrates that Agent KB aggregates cross-domain agent trajectories to enhance autonomous problem solving.
• It employs a standardized experience schema and a two-stage retrieval system to efficiently guide execution plans.
• Experimental results show benchmark improvements and reduced errors across diverse frameworks, underscoring practical applicability.

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

The paper "Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving" explores how autonomous agents can benefit from shared experiences accumulated across diverse frameworks. The authors propose Agent KB, a framework-independent knowledge base that aggregates agent trajectories into structured experiences, facilitating efficient cross-domain problem-solving.

Introduction to Agent KB

Traditional AI agents operate in isolation, limiting their ability to leverage the experiences of other agents. Agent KB addresses this limitation by constructing a universal memory infrastructure that enables agents to share experiences seamlessly across various frameworks without retraining or architectural modifications.

Figure 1: End-to-end workflow of Agent KB, highlighting construction, evolution, and task-solving stages.

Agent KB abstracts heterogeneous agent trajectories into structured experiences indexed in a central repository. It expands across domains by systematically adding, resolving conflicts, and evicting outdated entries, thereby maintaining the quality of information.

Methodology

Experience Representation

Agent KB structures agent experiences using a standardized schema:

$E = \langle \pi, \gamma, S, \mathcal{C} \rangle$

where $\pi$ is the task embedding, $\gamma$ encodes goal constraints, $S$ stores action-reasoning pairs, and $\mathcal{C}$ contains metadata ensuring cross-framework compatibility. This formulation allows experiences to be effectively shared and reused.

Two-Stage Retrieval System

Agent KB employs a hybrid retrieval method that consists of a planning stage and a feedback refinement stage:

1. Planning Stage: It retrieves past experiences to guide execution plans, producing initial solutions.
2. Feedback Stage: Utilizes execution feedback to refine the solution iteratively, applying adjustments only if they align with previous successful actions.

Figure 2: Comparison of agent workflows for protein distance calculation, illustrating enhanced accuracy through Agent KB.

The novel disagreement gate ensures that knowledge integration is coherent and non-disruptive to the agent's native reasoning flow.

Experimental Evaluation

Agent KB's efficacy was validated across major frameworks like GAIA, GPQA, HLE, and SWE-bench. It showed substantial improvements in reasoning and software engineering tasks with gains up to 18.7 percentage points in benchmarks such as pass@3 for smolagents and SWE-bench Lite.

Figure 3: Score improvements (%) across benchmarks for multiple LLMs enhanced with Agent KB.

Error Analysis

The integration of Agent KB significantly reduced retrieval and reasoning errors across different model families, showcasing robust performance and adaptability in solving complex tasks.

Figure 4: Error frequency assessment comparing smolagents with and without Agent KB.

Implications and Future Directions

Agent KB establishes a solid foundation for collective agent intelligence, paving the way for interoperable memory systems in AI. Future developments could explore richer modalities, longer-horizon reasoning, and further scaling of the knowledge base.

Conclusion

Agent KB represents an innovative approach to memory sharing among AI agents, emphasizing experience aggregation and seamless knowledge transfer. As a plug-and-play solution, it promotes universal applicability across heterogeneous agent ecosystems, providing a practical route toward achieving collective intelligence in autonomous systems.