MedCache Improves Clinical AI Agent Memory for Patient Records

A new framework called MedCache has been developed to address the challenges of designing effective memory for AI agents in longitudinal clinical settings. These agents must process evolving patient states from diverse sources like multiple visits, time points, and medical specialties. The research introduces a benchmark of multi-visit, multi-specialty patient records to evaluate how agent memory should be structured. The study systematically investigates various memory design choices, including curation, organization, retrieval, and memory-augmented reasoning. Key findings indicate that temporal validity is more crucial than simply retaining vast amounts of history, and that specialty-factorized memory can be beneficial but requires careful handling of shared evidence. MedCache integrates these insights by constructing temporally valid patient memory, organizing evidence into specialty-specific views, routing queries to relevant memories, and adaptively invoking specialists. Experiments demonstrate that MedCache significantly improves reasoning accuracy and memory efficiency compared to existing single-agent and multi-agent baselines. It also shows strong generalization across different model backbones and external datasets, suggesting its potential for robust application in complex clinical scenarios.