Dual-Stream Memory Architecture for Clinical Discrepancy Detection in LLM Agents

A new Dual-Stream Memory Architecture separates patient narrative from structured clinical records (FHIR) to detect discrepancies in health coaching agents. The system addresses the challenge of reconciling patient self-report (current but biased) with EHR data (validated but stale). A Reconciliation Engine evaluates each extracted memory against the FHIR profile, classifying discrepancies by type, severity, and involved FHIR resources. This approach aims to prevent safety failures common in general-purpose agent memory systems that overwrite older facts with the latest user statement. The architecture is designed for LLM agents managing longitudinal healthcare journeys, moving beyond single-session tools.