Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory

A new framework called SkeMex enables medical artificial-intelligence agents to improve their clinical reasoning over time by building a structured skill memory, without altering the underlying model weights, according to research submitted on 8 Jun 2026 [1][2]. The work addresses a limitation in current medical agent systems, which are increasingly expected to handle interactive clinical decision-making rather than only static question answering. Existing memory mechanisms often store raw historical traces that are redundant, noisy, and difficult to govern, and they rarely distinguish which memories are useful for future reasoning [2]. SkeMex distills informative interaction trajectories into structured skills that encode reusable procedural knowledge and organizes them into a multi-branch repository spanning general, task-specific, and action-level experience [2]. A closed-loop "Read--Write--Assess--Govern" lifecycle supports continual evolution by writing new skills, updating utilities, promoting useful memories, and removing harmful entries [2]. The framework estimates context-dependent utility from environment feedback to guide value-aware retrieval and repository governance [2]. In experiments across diverse clinical tasks, SkeMex consistently outperformed representative memory-based agents in both offline and online settings, and it generalized across model backbones while supporting transferable skill memory [2]. The authors stated that all data and code will be released publicly [2]. The preprint was posted on arXiv, an open-access repository for electronic preprints that has been operating since August 1991 and now receives about 24,000 submissions per month [6]. arXiv hosts papers across fields including computer science, physics, and quantitative biology, and it is not peer-reviewed [6]. The SkeMex paper appears under the Computer Science and Artificial Intelligence category and is accessible through the arXiv abstract page, which also features community-developed tools via the arXivLabs framework [5][4]. arXivLabs, launched in 2020, allows third-party collaborators to build experimental features such as bibliographic explorers and code-finding tools that appear as tabs on article pages, under guidelines that require adherence to openness, community, excellence, and user data privacy [5].