The Continuity Problem in Artificial Intelligence
By Momar Lissa Ndiaye ("MLN"), Founder & CEO, weyoga Inc.
This essay makes a narrow claim, carefully: the dominant architecture of consumer AI contains a structural limitation that no amount of model capability can fix, because the limitation is not in the model. It is in the relationship between the model and the person using it.
Begin with definitions, stated operationally so they can be used without modification.
Continuity. A system is continuous with respect to a person if and only if information generated in their past interactions is (a) durably persisted, (b) structured along a temporal axis, and (c) available by default to condition future interactions, without the person re-supplying it. A system failing any of the three conditions is discontinuous with respect to that person.
Session-level and sequence-level facts. A session-level fact is observable within a single interaction. A sequence-level fact is a property of the ordered set of interactions — it exists only across the sequence and is undefined for any individual element.
Those two definitions are the entire essay; everything below is consequence. That you asked about leaving your job is a session-level fact. That you have asked about leaving a job every eighteen months for six years, always in the same quarter, always following the same trigger, is a sequence-level fact. Sequence-level facts are generally the important kind — they are where causes live — and they are invisible, by construction, to any observer who meets you fresh each time, regardless of that observer's intelligence. A system cannot recognize a pattern it is structurally prevented from remembering.
Today's consumer AI is overwhelmingly discontinuous, or at best shallowly continuous — a profile paragraph, a preferences file, a retrieval pass over old transcripts. The default unit of experience remains the session. And this matters because of what kind of limitation it is. The scaling thesis holds that capability gains eventually dominate all other factors — that a sufficiently capable model compensates for missing context by inference, as a brilliant diagnostician infers history from one examination. Within a session this is substantially true. But note precisely what the claim here is not. It is not "the model isn't smart enough." It is "the model cannot infer information that has never been observed." A doctor of arbitrary genius cannot determine from one visit whether your blood pressure is chronically elevated or elevated today; the sequence-level fact requires the sequence. This is an information constraint, not an intelligence constraint — and information constraints survive arbitrarily capable models. That is what makes the limitation architectural rather than temporary.
The obvious rejoinder is that memory features are shipping everywhere, and the trend is real. But the framework this series is building keeps four terms strictly apart, and this is the essay where the distinction earns its keep: memory stores. Continuity preserves. Recognition interprets. Behavior changes. Present memory implementations are recall systems — they satisfy condition (a), persistence, and stop. Continuity additionally requires temporal structure, so trajectories are visible rather than merely facts. Recognition requires more still: the system must represent "this has happened four times," not merely retain four instances, and must surface that representation at decision moments rather than upon request. A transcript archive with search is to continuity what a filing cabinet is to a biographer: necessary raw material, nowhere near the finished capability.
Why has discontinuity persisted, if the gap is this legible? Partly commercial comfort: discontinuous products are interchangeable, which suits an industry organized around model competition. But the deeper reason is an optimization mismatch that deserves emphasis: continuity does not demo. Venture-backed software is selected by what can be demonstrated in minutes; continuity accrues value on the timescale of years. Intelligence demos brilliantly — it can be exhibited to a stranger in one exchange. A system's knowledge of a particular person cannot be exhibited to a stranger at all; it is, definitionally, valuable only to the person it accumulated around. The industry's selection function is not neutral between these capabilities. It systematically breeds one and starves the other.
Two honest limitations of the argument — one of which is not a limitation but a boundary. First: a longitudinal record of a person's decision patterns is among the most sensitive artifacts imaginable. But consent here is not an implementation detail to be handled by a settings page; it is part of the architecture itself, load-bearing in the definition. Recognition without consent isn't recognition. It's surveillance. The distinction is not tone — the two produce different systems, different data structures, different incentives, and only one of them deserves to exist. Second: continuity can entrench as well as reveal — a system that remembers your patterns might calcify them, pattern-matching you into a self you are trying to leave. Continuity done well must represent trajectories, not just tendencies: what is changing, not only what repeats. Both problems are real. Neither is an argument for amnesia. They are the engineering and ethical agenda of the continuity problem, not its refutation.
The narrow conclusion: there exists a class of high-value facts about every person — sequence-level facts, where their patterns live — that current AI architecture is structurally unable to observe, and that no improvement in model capability will make observable, because the limitation is informational, not cognitive.
This gap will not be closed because models get smarter; smarter models cannot close it. It will be closed because sequence-level information has become economically valuable — the one asset in AI, as the previous essay argued, that compounds instead of commoditizing. Architecture follows value. Someone will build for the sequence deliberately.
The rest of this series is about what building for it deliberately looks like.
Part of The Recognition Layer →
Momar Lissa Ndiaye ("MLN") is the Founder & CEO of weyoga Inc., a Delaware company. — weyoga.ai · mln@weyoga.ai