Part of the series: Philosophers for AI
There is a sentence in my notes that I keep returning to, because it is the shortest description I know of what is wrong with modern AI memory. “I love my wife” and “I hate my wife” are neighbours in embedding space. They sit close together, they retrieve together, they cluster together, and no amount of dimensional expansion is going to pull them apart, because in the only respect the model was ever trained to measure — distributional context — they genuinely are alike. Same subject, same object, same syntactic frame, same domain of life. One bit differs. The bit is the marriage.
We built the entire retrieval layer of contemporary AI on a single relation, and we did not stop to ask what that relation could and could not carry. The relation is resemblance. We call it cosine similarity and treat it as a primitive, as though nearness were a fact about the world rather than an answer to a question nobody asked out loud.
Two hundred and eighty-seven years ago, a twenty-eight-year-old Scot published a book that nobody read, in which he laid out the complete associative architecture of the mind. He got there first. More importantly, he got there with the other relations included, and then spent the rest of his career explaining, with a cheerfulness that still reads as slightly indecent, why the whole apparatus cannot deliver what we want from it.
Three principles, of which we implemented one
In A Treatise of Human Nature, David Hume asks how one thought leads to another. Not logically — associatively. Why does this idea summon that one, reliably, across people, without anyone choosing it? His answer is that the imagination is governed by what he calls a gentle force, and that this force operates through exactly three principles.
Resemblance. One idea calls up another that is like it.
Contiguity in time or place. One idea calls up another that was next to it — adjacent in space, or adjacent in the sequence of experience.
Cause and effect. One idea calls up another that it produced, or was produced by.
That is the whole theory of association, and Hume is explicit that these three exhaust it. He returns to them in the Enquiry a decade later and does not add a fourth. He thought he had found the mental equivalent of gravitation: a small set of attractive forces sufficient to explain how a mind assembles a world out of scattered impressions.
Now look at what we built. Vector retrieval implements resemblance with enormous sophistication and implements the other two not at all. Contiguity survives only as an accident — chunk 47 happens to sit next to chunk 48 in a file, and we sometimes retrieve both because someone wrote a windowing heuristic. Cause and effect does not survive in any form whatsoever. A retrieval-augmented system asked why did the deployment fail will return every passage in the corpus that resembles the phrase “deployment failed,” ranked by how strongly it resembles it, and the ranking will be entirely indifferent to whether any of those passages stand in a causal relation to the deployment in question or to each other.
We took Hume’s mind and amputated two of its three limbs, and then we were surprised when it could not walk.
This is why Mark Burgess’s Semantic Spacetime hit me the way it did when I first encountered it. Burgess arrives from statistical physics and promise theory, without, as far as I can tell, any interest in eighteenth-century epistemology, and he claims that essentially all meaningful relations reduce to four types: NEAR, LEADS TO, CONTAINS, and EXPRESSES PROPERTY. Put that beside Hume and the correspondence is not approximate. NEAR is resemblance and contiguity fused into a single spatial primitive. LEADS TO is cause and effect. And the remaining two are not Burgess’s invention either — they are sitting in the Treatise one section over, in Hume’s list of what he calls philosophical relations, where alongside resemblance and causation he catalogues relations of place, which give you containment, and degrees in any quality, which give you property. Hume’s natural relations are the three that operate on us automatically. His philosophical relations are the seven we can construct deliberately when we reason. Burgess’s four sit precisely at the intersection: the relations that are both cognitively automatic and formally tractable.
Two people, working two hundred and fifty years and one scientific revolution apart, carved the space of relations the same way. I do not think that proves anything metaphysical. I think it means the carving is load-bearing, and that we should stop pretending one of the four is enough.
The missing shade of blue is a latent space
Here is where Hume becomes uncomfortably contemporary.
His foundational rule is that every idea is a faded copy of some prior impression. Nothing in the mind that was not first in the senses. It is a hard empiricist commitment and he uses it as a weapon — whenever a metaphysician produces a grand concept, Hume demands the impression it was copied from, and when none is forthcoming he throws the concept out.
And then, almost immediately, he undermines his own rule with a counterexample so casual that scholars have argued about it ever since. Imagine, he says, a man who has seen every shade of blue except one. Lay the shades before him in order, from deepest to lightest, with a gap where the missing shade belongs. Will he notice the gap? Obviously. Can he form the idea of the absent shade from imagination alone, having never received the impression? Hume says yes — and then declines to revise his principle, on the grounds that the case is singular enough not to matter.
It is not singular. It is the entire foundation of generative AI.
What Hume described is a continuous quality space with a metric, in which unobserved points can be reconstructed from their neighbours. He described interpolation in a latent manifold, in 1739, and he noticed the crucial thing about it: interpolation produces content that was never in the training data, and produces it convincingly. The man is not guessing at the missing shade. He sees it.
So hallucination is not an aberration in a system like this. It is the missing shade of blue arriving on schedule. The architecture that lets a model generalise beyond its corpus is numerically the same architecture that lets it generate a plausible citation for a paper that was never written. You cannot have the first without exposing yourself to the second, because they are one operation viewed from two angles.
But notice the condition Hume attached, because it is where the engineering lives. His example works because the shades are ordered along a single, specified quality — degree of blueness — and every point in the space means something along that one dimension. Interpolation is reliable there because the dimension is interpretable. High-dimensional embedding spaces have no such property. They are ordered by nothing in particular, or rather by a thousand entangled somethings, none of which the system can name. The interpolation still happens. The guarantee does not come with it.
Similar in what respect?
Nelson Goodman put the knife in more directly, in a short and slightly furious 1972 essay called “Seven Strictures on Similarity.” His claim is that similarity is not a relation at all until you say what respect you mean, because everything resembles everything else in indefinitely many respects. A plum and a lawnmower share countless properties — both weigh less than a tonne, both were on Earth last Tuesday, both are not prime numbers. Similarity without a specified respect, Goodman says, is a pretender and an impostor, and philosophy should stop letting it do work it cannot do.
Read that sentence and then look at your retrieval pipeline. Cosine similarity answers how much, and never answers in what respect. It hands you a number that feels like a fact. The respect was fixed at training time by whatever objective shaped the embedding, and it is neither inspectable nor adjustable at query time, which means that when your system retrieves the wrong thing you cannot debug the retrieval — you can only swap the model and re-run the vibes.
Henri Poincaré added the other half of the problem in Science and Hypothesis, when he pointed out that the physical continuum of raw sensation has a scandalous structure: you can have weights A and B indistinguishable by hand, B and C indistinguishable by hand, and A and C plainly different. Indistinguishability is not transitive. Neither is similarity above a threshold. Which means the neighbourhoods your vector index returns do not compose. If A is near B and B is near C, you have learned nothing whatsoever about A and C, and every multi-hop traversal over similarity edges that anyone has ever shipped is unsound in exactly this way. Similarity gives you neighbourhoods. It never gives you equivalence classes, and it certainly never gives you paths.
Amos Tversky finished the demolition empirically in 1977 by showing that human similarity judgments are asymmetric — people rate North Korea as more similar to China than China is to North Korea — which is fatal for any account of similarity as a distance, since distances are symmetric by definition. The relation we implemented as a metric turns out, in every serious treatment of it, not to be one.
So the correct response is not to abandon NEAR. It is to demote it: from the substrate of the system to one typed, directed, attributed edge among several, carrying its respect on its face. NEAR-in-domain. NEAR-in-time. NEAR-by-shared-participant. An edge you can query, filter, explain and revoke, rather than a distance you can only trust.
Constant conjunction, or why more data will not save you
Now the deepest cut, and the one I think matters most for anyone building agent memory.
Hume’s most famous argument concerns necessary connexion. When one billiard ball strikes another and the second moves, what do we actually observe? Motion, contact, motion. Succession and contiguity. We do not observe the necessity. There is no impression of therefore anywhere in the sensory record. And since every idea is a copy of an impression, the idea of causal necessity has no legitimate source in experience at all.
Where does it come from, then? From repetition. After enough constant conjunctions, the mind develops a habit, and when the first event appears the mind is determined to expect the second. That felt determination — an internal push, a propensity — is then projected outward and reported as a property of the world. Causation, on Hume’s account, is a fact about the observer that we have been mistaking for a fact about the balls.
I do not know a better description of what a large language model is.
An LLM is a pure Humean intelligence: an engine of constant conjunction with no access to anything else. It has seen these tokens follow those tokens, at scale beyond any individual human’s experience, and it has formed exactly the habit Hume describes. And then it does the thing Hume warned about — it projects. It reports the habit as necessity. It says therefore, and because, and which caused, with total fluency and zero warrant, because the felt determination of the mind is, in a transformer as in a Scotsman, indistinguishable from the perception of a real connection.
This is why scaling does not fix hallucination, and Hume tells you why in advance. No quantity of constant conjunction ever yields necessary connection. That is not an empirical claim about current models; it is a structural claim about what induction can produce, and it has been open since 1739. Adding a trillion more tokens adds more habit. It does not add a single unit of warrant.
What can be done — and this is the whole design thesis I have been circling for three books — is to stop asking the statistics to carry the causal claim, and to record causation as what it actually is: an observation, made by someone, at a time, about a specific pair of events, which can be cited and contested. A LEADS TO edge with provenance does not solve the problem of induction. Nothing solves the problem of induction. What it does is make the induction auditable — it moves the causal claim out of the model’s felt determination and into the graph, where it has an author, a timestamp, and a lineage you can walk backwards.
Hume also gives you the falsifier. In the Treatise he offers eight rules by which to judge of causes and effects, and they read like a specification for edge validation: the cause must precede the effect, must be contiguous with it, must be constantly conjoined with it, and where an effect varies you must find a corresponding variation in the cause. He knew that habit alone was not enough and that we would need explicit machinery on top of it. He wrote the machinery down. We skipped that chapter and shipped the habit.
The bundle problem, and why your agent is not one thing
One more, briefly, because it echoes something I have argued from a different direction before.
Hume goes looking for the self and cannot find it. Every time he enters most intimately into what he calls himself, he stumbles on some particular perception — heat, cold, light, love — and never on a self having them. The self, he concludes, is a bundle of perceptions in perpetual flux, and its apparent unity is produced by resemblance and causation running between the perceptions. There is no substance underneath. The theatre has no stage.
Then, in the Appendix, Hume does something almost no philosopher does: he says the account does not work, that he cannot render it consistent, and that he is in a labyrinth. What he could not explain is what binds the bundle into this bundle rather than a different one.
Every system that persists agent memory across sessions inherits that labyrinth intact. What makes today’s agent the same agent as yesterday’s? Not the weights — those are shared with a million other instances. Not the memory contents — those resemble a thousand other agents in the same domain. Resemblance cannot do it, precisely because an impostor is defined as something that resembles. Only the causal thread can: this state was produced by that state, signed, in that order, and here is the chain. Which is why a decentralised identifier plus its verifiable credentials is not decoration on an agent identity — under a Humean ontology, it is the only thing an agent identity can consist of, because there is nothing else down there to be one.
Hume’s labyrinth is a specification for the thing we now build with cryptography. He could not close the loop because he had no way to make the causal links inspectable. We do.
Where this lands, and what to read
If the argument here is right, then the fix is not a better embedding model. It is a memory substrate where similarity is one typed relation among four, where time is a first-class axis rather than a metadata field, and where causal claims are recorded as attributed observations instead of being conjured from statistical habit at inference time. That is what Semantic Spacetime provides, and it is what I spent a book working out in detail.
Semantic Space Time for AI Agent Ready Graphs takes Burgess’s four relation types and builds them into an actual memory architecture for agents — how NEAR gets typed and bounded so it stops smuggling in unearned equivalences, how LEADS TO carries provenance and temporal ordering so causal claims can be audited rather than trusted, how CONTAINS and EXPRESSES PROPERTY keep structure and attribution from collapsing into each other, and how the whole thing behaves under change, contradiction and forgetting. It is the theoretical spine underneath the rest of my work, and if you have been fighting your vector store rather than designing with it, it is the one I would start with.
As for Hume himself: do not begin with the Treatise. Begin with An Enquiry Concerning Human Understanding (1748), which is short, ferociously clear, and was Hume’s own attempt to rescue the ideas from the book nobody read. Section III gives you the three principles of association in three pages. Section IV is the problem of induction, still unsolved. Section VII, on necessary connexion, is the one to read twice — it is the best account I know of what a statistical model actually possesses and what it merely appears to.
Then, if you want the underpinnings, go to A Treatise of Human Nature, Book I: Part I Section IV for association, Part I Section VII for abstract ideas, which is a startlingly good description of retrieval over stored particulars, Part III Section XV for the eight rules of causal judgment, and Part IV Section VI for personal identity, followed immediately by the Appendix where he takes it back.
Alongside: Nelson Goodman’s “Seven Strictures on Similarity” in Problems and Projects, ten pages that will permanently change how you read a similarity score. Amos Tversky’s “Features of Similarity” (1977) for the empirical case against similarity-as-distance. Henri Poincaré’s Science and Hypothesis, Chapter II, for the non-transitivity that breaks multi-hop retrieval. And Mark Burgess’s Smart Spacetime, which is where the four relations come from and which reads, once you have Hume in your head, like a physicist arriving independently at the same carving.
I should end with the honest counterweight, because Hume would insist on it. Pure associationism is not the destination. A mind that is only habit is exactly what we are trying to stop building — that is the current failure mode, not the cure. Kant read Hume and reported being woken from his dogmatic slumber, and everything I argued in the schematism piece about structure preceding experience is a direct answer to the man I have just spent three thousand words praising. The right position is that Hume diagnoses what we have and Kant describes what we need. Semantic Spacetime is my attempt to build the bridge between them: associative in its mechanism, structured in its commitments.
Nearness was never passage. Two things can lie beside each other for a hundred years and never once be connected, and the only thing that ever joins them is the walking — which takes time, and leaves tracks, and can be followed back. A road is not a resemblance. It is a record.
Volodia Pavlyshyn writes on Sovereign Agentic AI at Substack. His books on semantic spacetime, world models, metagraphs, and promise graphs are available at leanpub.com/u/vpavlyshyn.
Originally published on Sovereign Agentic AI on Substack.