About ten minutes. Not another long essay: the shortest honest path from “the world does not arrive as prompts” to the hypothesis this program exists to test.
Most interactions with AI begin because someone asks. But most situations don’t begin with a question. A customer stops responding. A shipment runs late. A machine starts behaving oddly. A project slowly falls behind. Nobody prompts anything. Something changes.
Modern AI has become remarkably good at reasoning once it is told what problem to solve. Being told is the easy case. In the world, the problem has to be found before it can be solved.
Before a system can decide what to do, something has to determine what is going on, for whom it matters, and what the moment calls for. We call that recognition. It comes before reasoning, and today it mostly happens implicitly, inside a prompt.
Context is what a system can know: history, documents, roles, relationships, signals. Situation is what that knowledge means right now. Having the right information is not the same as understanding the situation, and no amount of retrieval closes that gap on its own.
A payment failed: an event. The order remains unpaid: a state. The customer wants to complete the purchase: an intent. Together with what is known, these become a situation: a valuable customer blocked by what may be a temporary failure. Only the last one tells the system what to do.
Seven days of silence from a customer is a cooling prospect to a salesperson, a resolved issue to support, and an emerging collection risk to finance. The facts are shared. The situations are not. A situation is a relationship between what is happening and whoever it is happening to.
A prospect evaluating a proposal becomes a competitive evaluation, then a stalled decision, then an active negotiation. Nothing says these need fixed names. What matters is that the system notices when what is happening has changed, and adjusts.
Conventional memory remembers information. Situation memory would remember experience: what was recognized, from whose perspective, what response was chosen, what followed, and whether the interpretation turned out to be right.
Capable models already infer situations implicitly. Situational AI asks what changes when situations become explicit objects: detected, represented, compared, tracked, remembered, and revised as evidence changes, rather than disappearing when an inference ends.
Tokens are the unit of language modeling. States are the unit of control. Our hypothesis is that intelligent behavior needs a native unit that binds perception, goals, judgment, action and outcome: the situation. This is a claim about computational role. It is not yet a claim that it works.