Recent research on self-improving AI agents has largely focused on improving models, reasoning policies, prompts, tools, or other agent scaffolds through experience. However, the external knowledge on which agents rely is also a critical object of self-improvement: it can become incomplete, incorrect, redundant, or poorly linked over repeated use, thereby weakening the agents’ search and downstream answers. This project will investigate data agents for self-improvement: agents that use query and interaction histories to diagnose local knowledge defects, propose evidence-grounded edits, and incrementally improve a persistent knowledge base. Building on the query-conditioned refinement paradigm of DeepRefine, the project will develop a small-scale, auditable prototype with structured insert, delete, and replace operations, source verification, versioning, and rollback. The project will evaluate whether targeted knowledge refinement improves its effectiveness for agentic tasks and its quality more efficiently and safely than leaving the knowledge base unchanged or rebuilding it from scratch.