In self-improving AI systems, the traditional approach treats the agent as the object of improvement, optimizing prompts, workflows, or even the agent's own code. This agent-centric perspective can lead to expensive maintenance and difficulty in transferring gains, as improvements become tied to specific agent designs, task distributions, or adaptation runs.
We explore a complementary paradigm: knowledge-centric self-improvement, where agents remain generic and disposable while the persistent object is a curated knowledge base that agents can leverage for future tasks. We conduct controlled case studies to operationalize this idea via a simple protocol. Agents attempt a task and then contribute evidence-grounded insights to a shared knowledge base through task-level and cross-task forums, followed by knowledge distillation.
Since self-improvement occurs in the knowledge rather than the agent, improvements become more inspectable, transferable, and portable. Across abstract reasoning, coding, and terminal benchmarks, this protocol improves solve rates while reducing costs relative to agent-centric baselines. The resulting distilled knowledge also transfers to held-out tasks and across LLM families, indicating that the improvement is not merely an LLM- or run-specific behavior. These results support a new view of self-improving agentic systems: progress can primarily be driven by curated persistent knowledge.
Code is available at GitHub.
Blogger's Review: This research introduces a paradigm shift towards knowledge-centric self-improvement, significantly reducing the cost of enhancements in AI systems while increasing their transferability, marking an important evolution in AI research. The construction and utilization of knowledge bases could become crucial for the future development of AI.