LearnActCoder: Role-Aware Error Memory for Adaptive Clinical Coding Agents

LearnActCoder uses a Mistake Knowledge Database to route false-negative errors to a recall-oriented coder and false-positives to a precision-oriented judge. This inference-time adaptation increases CPT F1 by 5.9 percentage points on MIMIC-III notes compared to reflection-style memory methods. The framework utilizes lookup-table grounding to stabilize clinical coding performance.