Continuous Clinical Memory: an architecture for context aware ICU decision support

Most clinical AI is stateless. A question arrives, a context window is assembled, an answer is produced, and everything is forgotten. That design is tolerable for consumer chat. It is the wrong architecture for intensive care, where the unit of meaning is not a message but a stay: days of vitals, investigations, decisions, rounds and handovers that only make sense together.

The blank prompt problem

Ask a stateless assistant why a patient is still on vasopressors and it can only reason from whatever fits in the prompt. It does not know that pressors started 22 hours ago at stabilization, that the last handover set a weaning threshold, or that the cardiovascular component is what keeps the SOFA score at 9. The clinician carries all of that in working memory. The software should carry it too.

The model: one evolving memory per patient

UltraCare treats the ICU stay as a single, continuously evolving clinical memory. Seven stores feed it, each written by the workspace itself as care happens.

• Charted vitals and labs, every value, trended.
• Severity scores, recomputed deterministically on every charted value.
• The structured record, from admission through management plan.
• The patient knowledge graph: comorbidities, medications, lines, allergies, linked.
• The encounter timeline, every event on every shift.
• I-PASS handovers, so what the last shift knew is never lost.
• Safety sweeps, the FAST HUGS BID status each round.

Admission creates the foundation. Every assessment deepens it. Every round updates it. Every handover preserves it. By discharge, the system holds a complete clinical memory of the stay.

What this enables, and what it must not

Grounding changes what an assistant can safely offer. Every line of a draft can be traced to a store. A summary cites the chart, not a guess. A contradiction between today and yesterday is visible because both are in memory.

It also sharpens the boundary. The assistant drafts, cites and flags. It does not prescribe, does not diagnose, and does not act without clinician sign off. Deterministic scores and transparent rules form the core. A validated machine learning predictor joins only in Phase 2, after pilot data and ICMR ethics approval.

Why this is hard to retrofit

A bolt on chatbot cannot backfill structure that was never captured. The workspace writes the memory as a side effect of care, which is precisely why it cannot be skipped: the wedge is the workflow, and the memory compounds with every shift.

The unit of clinical meaning is the stay, not the message.

See it running

The intelligence page lets you operate the instruments on a synthetic patient. The evidence page documents every method and citation. Decision support, not decision making.