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

Researchers introduce SkeMex, a self-evolving skill-based memory framework that enables medical AI agents to improve after deployment without retraining model weights. The system distills clinical interaction trajectories into reusable procedural skills, organized across multiple memory branches, and uses environment feedback to determine which experiences are genuinely useful for future decision-making.

SkeMex addresses a fundamental limitation in current medical AI systems: while these agents increasingly support interactive clinical decision-making, their memory mechanisms struggle to efficiently capture and reuse valuable experiences. Traditional approaches store raw historical data that becomes redundant, noisy, and difficult to manage at scale, constraining long-horizon clinical reasoning and generalization across cases.

The research builds on growing recognition that AI systems need to learn from deployment feedback rather than rely solely on static pre-training. In clinical contexts, this challenge intensifies because incorrect memory retention could propagate harmful patterns across subsequent patient cases. SkeMex tackles this through structured skill extraction—converting raw interaction trajectories into encoded procedural knowledge—organized into task-specific and action-level branches with built-in governance mechanisms.

The framework's closed-loop lifecycle (Read-Write-Assess-Govern) represents a significant architectural advance. By continuously evaluating which skills remain valuable based on environmental feedback and actively removing harmful entries, the system maintains memory quality without human curation. This approach could substantially improve AI reliability in clinical settings where experience accumulation directly impacts patient outcomes.

The demonstrated transferability across model backbones suggests the skill memory framework operates as a modular layer, potentially applicable to different AI architectures. For the broader medical AI industry, this work provides a practical post-deployment improvement pathway that doesn't require expensive model retraining. The planned public release of code and data may accelerate adoption of similar self-evolution patterns across healthcare AI systems.

• →SkeMex enables medical agents to improve after deployment by distilling clinical interactions into reusable structured skills without retraining models.
• →Context-dependent utility estimation guides which memories are retained, preventing accumulation of redundant or harmful historical data.
• →The framework organizes skills across general, task-specific, and action-level branches with automated governance for continuous evolution.
• →Experiments show consistent performance improvements across diverse clinical tasks in both offline and online settings.
• →Skill memory transfers across different model architectures, suggesting broad applicability to varied medical AI implementations.