Thalamus is a four-layer pipeline that turns every document a patient generates into a continuously computed patient state, and compiles the right slice of that state for each agent, each time.
A five-minute agent call becomes 22 atomic facts. Most don't touch anything. One does, and it can move more than one part of the patient's state at once.
Each unit of computed state is a Health Profile Pointer, or HPP: L3's counterpart to L2's ACDPs. One ACDP can affect several HPPs, and several ACDPs can support one HPP. The highlighted ACDP above touches two existing HPPs at once.
The ACDP doesn't touch any existing HPP. Roughly 80% of new ACDPs land here.
a. State updates: e.g. smoker goes from no to yes.
b. Evidence changes, state doesn't: poor sleep quality was already flagged; two weeks of new wearable data raises confidence in it without changing the HPP itself.
c. No change: the dependency is real, but the HPP's value stays the same.
If the ACDP is new and clinically relevant and doesn't map to any existing HPP, it creates one.
Every source document a patient generates, kept in full. Nothing is discarded here; this layer is the system of record underneath everything Thalamus computes.
Every document is broken down into discrete clinical facts, known as Atomic Clinical Data Points, or ACDPs. Each one carries roughly ten metadata tags today, a set built to grow past a hundred: category, organ system, asserter, temporal information, source and confidence among them. A single patient can hold thousands.
Every ACDP links back to its original document, timestamp and provenance.
L3 performs inference over the atomic layer to compute the patient's current clinical state. As new information arrives, Thalamus determines which downstream states are affected and updates only those states, while preserving the evidence behind each one.
Three capabilities sit on top of the computed patient state, each consuming the clinically coherent state differently.
Different agents need different context: a pregnancy agent doesn't need a full dermatology history. The compiler decides which patient state, which ACDPs and which context each specific agent receives, and prepares the input it's given without requiring the agent to reconstruct the patient's clinical state from raw data.
Ingests clinical guidelines from bodies including ADA, USPSTF, AHA, ASCO and ATA, and determines which apply, which recommendations are relevant, and which screenings should surface given the patient's current clinical state.
Proactively surfaces care gaps, preventive opportunities, missing follow-ups, clinical risks and unresolved issues from the patient's computed clinical state, before the downstream agent starts reasoning.
Every new document, whether a lab result, a follow-up note or a five-minute agent call, flows through the same pipeline the moment it arrives. Thalamus identifies which parts of the patient's clinical state are affected and updates those states, so the patient state continuously evolves as new evidence arrives.
Bring a real patient record set and we'll walk through how Thalamus turns it into a continuously computed patient state and what your agents can reason from.