PatientMem: A Structured Longitudinal Memory for LLM-Based Clinical Reasoning

Clinicians reviewing a patient's history in the ICU must piece together information scattered across multiple discharge notes, which can be slow and demanding under time pressure. We introduce PatientMem, a structured longitudinal memory built in two stages: an LLM extracts a structured summary from each discharge note, and a fully deterministic merge step consolidates these summaries into a single patient-level memory that tracks clinical state across admissions. On EHRNoteQA, a multi-admission clinical question answering benchmark, PatientMem improves accuracy by $14$-$19$ percentage points (pp) over using only the latest note. On the same benchmark, TOST equivalence testing at pre-specified sensitivity margins of $\delta = 1$, $3$, and $5$ percentage points shows that PatientMem is equivalent to the full concatenated record at the $\delta = 3$ and $\delta = 5$ pp margins. We further investigate confidence-based routing to the original notes when memory may be insufficient and find that model confidence alone does not reliably identify such cases. As a secondary contribution, we provide clinician-reviewed corrections to the $2.36\%$ of the evaluated cohort containing erroneous or ambiguous ground truth.