Remembering to do something later
There are two kinds of memory failure. One is forgetting what you knew. The other is knowing perfectly well what you were supposed to do, and not doing it, because the moment slid past unmarked. The second kind has a name in cognitive psychology: prospective memory, the capacity to form an intention now for execution later.
It splits cleanly into two modes. Event-based prospective memory relies on a cue in the world: you post the letter because you pass the postbox. Time-based prospective memory has no such cue. Nothing external tells you it is eight o'clock; you must generate the prompt from inside, unaided, against whatever else you happen to be doing. The two modes are not equally reliable. Time-based intentions fail far more often, and they fail in a specific, well-characterised way. The content is rarely lost. Ask someone who has missed a scheduled task whether they remember what they were meant to do, and they almost always do, immediately and completely. What failed was not storage. What failed was noticing — the intention did not surface at the moment it was needed, because nothing forced it to.
This matters because it relocates the problem. Memory research spent decades treating forgetting as a retrieval failure of content: weak encoding, decay, interference. Prospective memory failure is usually none of these. The instruction sits intact, fully retrievable on demand. What is missing is the demand itself — the interruption that should have arrived and did not, because the task was structured to require the rememberer to interrupt themselves.
Origin
The distinction was worked out through the 1970s and into the 1990s, once applied researchers noticed that laboratory recall tasks — word lists, paired associates — explained almost none of the forgetting people actually reported in daily life. Meacham and Leiman's postcard studies, in which participants agreed to mail a card back on a given date, showed how badly self-generated timing failed even for a trivial, low-stakes act. John Harris's 1984 survey of memory-aid use confirmed the pattern from the other direction: people do not trust themselves with time-based tasks, and they compensate almost entirely with external props — notes, alarms, objects placed somewhere conspicuous. Gilles Einstein and Mark McDaniel gave the field its controlled experimental paradigm in 1990, embedding a to-be-remembered task inside an unrelated ongoing activity, and produced the finding that has held up since: intentions supported by an external cue are recalled at markedly higher rates than intentions that must be self-initiated. Medicine and aviation adopted the finding almost immediately, because both fields run on scheduled acts embedded in longer, unrelated activity — exactly the structure the laboratory work had shown to be fragile.
The turn
Read against this, the three generations in the lineage — Large Language Model, Large World Model, Large Universe Model — separate along a line that has nothing to do with fluency or perception and everything to do with prospective memory.
A Large Language Model has none. Its intake is a corpus closed at a cutoff. Between one query and the next, nothing runs. There is no clock ticking inside it, no stream it is still watching, no possibility of an intention surviving an interval because there is no interval — only the frozen past and the present prompt. It has retrospective memory, in abundance, and no mechanism by which a "do this later" could ever be discharged, because later never arrives inside the system. This is not a criticism of fluency. It is a statement about what closed intake forecloses.
A Large World Model improves this by exactly one grade. While a scene is being sensed, a cue inside that scene can fire a stored intention — event-based prospective memory, functioning. But the scene ends. The intention has no stream to persist in once sensing stops, so anything due after the episode closes is unreachable in principle, not just in practice. A Large World Model can watch a room and react to something happening in it. It cannot be told to check back in six months, because in six months there is no "it" still watching.
A Large Universe Model is defined by the property that fixes this: streams that do not close, and beliefs held with provenance so that an intention can be re-evaluated when the evidence supporting it changes. This converts time-based remembering into event-based remembering. The distinguishing structural fact is not "watching more." It is that the moment which used to require someone to remember the date now arrives as a cue inside a stream already under observation. The self-initiated interruption that human time-based memory does badly is replaced by an externally supplied one — the thing every prosthesis for prospective memory, from the postcard to the pill organiser, has always tried to manufacture.
Where this shows up
Radiology gives a clean instance. A trauma CT incidentally reveals a small pulmonary nodule; the report recommends repeat imaging in six to twelve months. The report is retrospective memory functioning perfectly — the recommendation exists, written, retrievable. What happens next depends on a time-based prospective memory task distributed across a patient, a GP, and an administrative system, none of which is continuously watching for the due date. Audits repeatedly find completion well under half. Aviation shows the same failure from the executing end: Northwest 255 left Detroit in 1987 without flaps or slats configured, because a checklist item that should have fired as an event-based cue was severed by an interrupted taxi sequence. The crew were experienced and the knowledge was intact. The intention did not survive the interruption. Medication adherence sits near 50% in developed-country reviews, and a large share of the shortfall is not refusal but simple failure to notice eight o'clock — a time-based task with no environmental cue, exactly the condition Einstein and McDaniel's paradigm was built to isolate.
The economic pattern generalises. Marine survey windows, pharmacovigilance commitments, vegetation-management cycles on power lines, loan covenant tests — each assigns a human a scheduled or conditional re-examination, and each such check is a time-based prospective memory task riding on a person who is, on the population statistics, not good at this particular thing. Continuous, provenanced intake does not make the person better at remembering. It removes the need for them to.
Objections, taken seriously
Automated checking is decades old. Cron jobs, tickler files, SCADA alarms already discharge scheduled tasks. There is nothing new here.
This is correct about execution and wrong about formation. A scheduled job discharges an intention that was fully specified in advance, against a stream that was fully specified in advance. It cannot notice that a condition nobody enumerated now warrants a check. Continuous intake with provenance changes what is eligible to be a cue at all: the trigger can be a shift in the evidence behind a belief, not only a threshold someone guessed correctly years earlier. The difference is real but modest, and it is the only one being claimed here.
Automated watching degrades into alarm fatigue. Intensive care units already show that most alarms are non-actionable and get ignored. Continuous intake multiplies notifications faster than it removes forgetting.
This is the strongest objection, and it mostly lands. Watching without calibration manufactures noise. The reply is narrow: what the objection demands — that a flag arrive with its evidence, its confidence, and its revision history attached, so attention can be rationed rather than exhausted — is exactly the provenance requirement that defines the position. A system that only fires alerts has not reached it. The claim is about revisable belief, not about volume of interruption.
Continuous intake supplies no intentions of its own. Deciding what matters is exogenous. The real bottleneck is goals and authority, not observation.
Conceded entirely. The axis under discussion is intake — what a system is permitted to observe, and for how long. Terminality on that axis says only that there is no further category of evidence beyond "every stream, continuously." It says nothing about who sets the intention worth checking, who is liable when a check is missed, or who is liable when a false one fires. Those problems get harder as intake widens, not easier, and the claim made here is deliberately silent on them.
The misreading to disown
The weak version of this argument says an always-on system means nothing is ever forgotten again. That is not what prospective memory failure is. The content is almost never the casualty; retrieval at the right moment is. Widening intake without improving cue-matching does not fix this — it can make it worse, since more streams means more competing signal. What continuous, provenanced intake actually buys is something narrower: forgetting becomes auditable. You can point to which check lapsed and on what evidence it should have fired.
That is what the concept establishes, and no more. It says intake has a top rung, that the third position occupies it, and that occupying it converts a class of tasks humans are measurably bad at into a class a system with an open stream is structurally suited to. It does not say the system knows what to watch for, or who answers when it watches wrong.