PrivMeSA uses local memory to manage privacy in medical AI consultations

A paper submitted to arXiv on 29 Sep 2026 presents PrivMeSA, a multi-agent system for medical AI that combines local and remote language models. The authors describe a privacy risk in sending patient information to more capable remote models: removing direct identifiers may not be enough. Details shared across multiple consultation turns or visits can accumulate and help identify a patient.

In PrivMeSA, a local agent manages each encounter and consults remote specialists, who may ask for additional information. The local agent controls what to disclose as the consultation unfolds. The system uses reinforcement learning to balance task accuracy against direct disclosure and the risk of re-identification through registry information. Privacy is assessed across the complete outbound transcript for an encounter, rather than only an individual message.

The system also builds a local memory from completed consultations. It distills those consultations into generalized clinical guidance and retrieves relevant lessons before information is sent in later cases. The authors say this lets subsequent cases reuse remote expertise without another remote exchange. They also report that the memory can grow without additional outcome labels or parameter updates.

The evaluation used an emergency-department benchmark built from MIMIC-IV-ED records. Compared with delegation, PrivMeSA improved mean task accuracy by up to 15.8 percentage points. In the same setting, the share of cases involving disclosure of personal details fell from 98.0% to 0.2%. The share of cases where a patient could be narrowed to ten or fewer registry patients fell from 74% to 0%.

For AI builders, the paper describes an architecture in which a local agent governs disclosure to remote models and stores generalized lessons for reuse. Its evaluation considers accumulated information across an encounter and reports task accuracy alongside two privacy measures. Those design choices and benchmark results offer a concrete example of how a system can combine remote consultation with local control and memory.