What Is a Persistent Virtual Patient?

A persistent virtual patient is a world that remains after the conversation ends. It keeps state, a clock, a disclosure boundary, a history of what happened, and branches that can be replayed. A chat transcript is a record of words. A persistent virtual patient is a record of a situation.

Most language-model sessions treat the patient as a prompt. You paste a vignette, ask a question, and receive an answer. Close the window and the situation is gone, except for whatever text you remembered to save. The next session starts from a new paste. Nothing in that workflow can say what the patient world held at minute twelve, which facts were still hidden, which action was committed, or what would have followed if a different action had been taken from the same moment. Those are properties of a world, and a world has to be stored.

Patientic is being built around that distinction. The system design treats the patient as durable state rather than as a disposable prompt. A person, or an external agent, can return to the same patient and inspect what is true now, what was true earlier, and which futures were explored. The aim is not to imitate a clinic. The aim is to give AI work a patient-shaped environment that does not forget itself between turns.

What has to persist

Four things have to outlive a single reply.

State is the structured situation. It includes the fictional or otherwise permitted identity, the facts currently in play, the open questions, and the commitments already made. State is inspectable. It is not a paragraph the model is free to rewrite because a new sentence sounded plausible.

Time is the position of that state on a clock the simulation owns. A patient world needs to know whether an observation happened before a decision, and whether a later branch is looking at the same moment or a later one. Without an explicit clock, "before" and "after" collapse into the order of messages in a window, which is an accident of the interface rather than a property of the case.

History is the append-only account of how the state came to be. It is the difference between a snapshot and a story you can audit. If an actor gathered a fact, that gathering is an event. If a person committed a decision, that commitment is an event. History is what makes replay possible.

Branches are alternate histories that share a past and then diverge. A branch is not a second patient pasted into a new chat. It is a fork from a known state, so two futures can be compared without pretending they had different beginnings. The shared prefix is the point. If you cannot name the state where the futures split, you do not have a branch. You have two unrelated stories.

Disclosure is part of the world

A persistent patient is not an open book. Some facts are visible to the actor at a given moment. Some are present in the world and not yet disclosed. The boundary matters because any serious look at an agent includes what it did with incomplete information. If every fact is stuffed into the first prompt, the agent never has to ask, wait, or notice that something is missing. The task becomes reading comprehension, and the patient becomes a paragraph.

The system design keeps that boundary explicit. An action can reveal a fact, and the reveal is recorded. An action can also fail to reveal it, because the actor did not ask, did not have permission, or asked at the wrong time. The hidden fact remains in the world. It is not deleted to make the prompt shorter, and it is not silently added to the actor's context because the model might find it useful.

This is also a boundary on what the system is for. The scenarios Patientic is being built to run are synthetic, educational, or otherwise permitted. The disclosure rule is a simulation control. It is not a design for how a real clinical record should be opened, and it is not a claim about a real person's care. The default is that identifiable real-patient data stays outside the world. That limit is spelled out in Synthetic Patient Simulation: What It Is—and What It Is Not.

A session is not a patient

It helps to separate three objects that chat interfaces collapse into one.

A session is a stretch of interaction. It has a start and, eventually, an end. It can be short. Two sessions can touch the same patient without being the same event.

A context window is the text a model can see on one call. It is a view. It is limited. It becomes lossy when someone summarizes it to make room. It is not the authority for what happened. If the window and the stored world disagree, the world is the record and the window is a rendering of part of it.

A patient world is the durable object. Sessions read it and propose actions against it. Context windows are assembled from it under a policy that decides what may be shown. Neither the session nor the window is allowed to overwrite the world just by speaking. A fluent paragraph is a proposal until something in the system commits it.

That separation is why "just use a longer context" does not answer the problem. A longer window can hold more text. It cannot, by itself, decide which facts are disclosed, which actions are permitted, or how two futures that share a past should be stored. Those are environment questions. They are the subject of AI Agents Need Environments, Not Just Context Windows.

Time, forks, and replay

Persistence without time is a filing cabinet. Persistence with time is a simulation. In this design, the world moves forward when a decision is committed, not whenever a model produces a paragraph. A proposed action is not yet a fact about the patient. A committed action is an event that later states can depend on. The commit is the moment a possibility becomes part of the history.

From a committed point, a fork starts another branch. The fork keeps the past it needs and then accepts a different next action. Both branches remain. Nothing is rewound by deletion. Looking backward is a way to inspect an earlier state or to leave the current branch for a new one. It is not an eraser, and it does not make the abandoned branch untrue as a record of what was tried.

Replay runs a recorded history again. The same patient, the same branch, the same events. Replay is how a person checks what an actor actually did, and how a later run can be set beside an earlier one. Comparison across branches is only meaningful if both branches name the state they diverged from. A chat log titled "version 2" does not do that. It might be a rewrite, a continuation, or a different patient entirely, and the file name will not tell you which.

patient world
  state + clock + hidden facts
        |
        v
  committed action ---> history (append-only)
        |
        +-- branch A --> later state A --> replay reads the same events
        |
        +-- branch B --> later state B --> compared from the shared fork

The diagram is the whole idea in one picture. The world is not the reply. The reply is something an actor does inside the world. History accumulates. Branches share a fork and then differ. Replay reads events that were stored, rather than asking a model to remember them.

What you can ask of it

A persistent virtual patient gives a builder or a reviewer a place to watch work happen over time. Did the actor gather information before acting? Did it respect a permission? Did a second branch, started from the same state, proceed differently? Those questions cannot be answered from a final paragraph, because the paragraph has already selected a conclusion and discarded the path.

Patientic is being built so that people and external agents operate this kind of world, rather than a private chat that happens to mention a patient. The public direction is synthetic simulation and evaluation of how agents act inside that simulation. Branching shows futures inside a specified setup. It does not predict a real person's outcome, and it does not tell a clinician what to do. If you want the product surface these essays are written toward, start from Patientic.

Related reading

• What Branching Clinical Simulation Actually Means
• Why Every AI Patient Run Needs a Simulation Receipt