The Brain Stores Quadruples: Why Context Is the Missing Primitive in Artificial Intelligence.
Article 1 of 4 in the series: Thinking Fast, Slow, and Situated
I. The Amnesiac Scuba Diver
In 1975, a group of scuba divers waded into the freezing waters off the Scottish coast near Oban. They were not there to explore a shipwreck. They were subjects in a cognitive psychology experiment that, half a century later, holds a missing blueprint for artificial intelligence.
Researchers Duncan Godden and Alan Baddeley asked eighteen divers to memorize lists of thirty-eight unrelated words. Half learned the words on a dry beach. The other half learned them twenty feet underwater, hearing the words through a diving communication device between the rhythmic hiss of their breathing apparatus. When tested on recall, the results were lopsided. Divers who learned words underwater recalled roughly 13.5 words when tested underwater, but only about 8.4 on land. The inverse held for the land-learners. Matching the learning environment to the recall environment produced approximately 50% better performance. Mismatching them caused roughly 40% of the learned material to vanish.
The water was not a metadata tag attached to the memory. The water was the geometry of the memory itself.
In 2021, Jaap Murre at the University of Amsterdam attempted to replicate the experiment. His team used sixteen divers in a heated indoor pool, running all four conditions in a single ninety-minute session rather than across four days in open ocean. The replication failed. But the failure is more interesting than it looks. Context-dependent memory is not a binary switch. It scales with the richness and distinctiveness of the contextual encoding. A warm pool at six feet provides weak contextual differentiation from the poolside. The freezing Scottish ocean, with its pressure, cold, and limited visibility, provides overwhelming differentiation from a sandy beach.
The principle is not that context helps memory.
The principle is that context is memory.
Endel Tulving formalized this in 1973 as the Encoding Specificity Principle: a retrieval cue is effective only to the extent that it overlaps with the information encoded at the time of learning. Steven Smith and Edward Vela confirmed the broader principle in a 2001 meta-analysis spanning decades of research. The specific numbers from the Scottish coast may be debatable. The principle they exposed is not.
But Tulving did something else, something the AI community has spent seventy years ignoring. He divided human memory into two fundamentally different systems. Semantic memory stores context-free, objective facts: Paris is the capital of France. Water boils at 100 degrees Celsius. Episodic memory stores situated, contextual, first-person experiences: I went to Paris last week and it rained. I burned my hand on the stove when I was six.
For seventy years, from the Semantic Web to LLM pre-training, the entire AI industry has been a single-minded attempt to build a purely Semantic AI. Strip the context. Extract the proposition. Store the fact. The assumption, never examined, has been that intelligence means reducing a messy, situated reality into sterile, objective propositions. Biology tells a different story. Survival does not run on encyclopedias. It runs on episodes.
II. The Tragedy of the Naked Triple
In 2001, Tim Berners-Lee, James Hendler, and Ora Lassila proposed the Semantic Web. At its foundation was a design choice that would shape two decades of knowledge representation: the RDF triple. Subject, predicate, object. The atomic unit of machine-readable knowledge.
(Ibuprofen → Treats → Inflammation)
This statement is logically sound. It is also the kind of statement that kills people. Ibuprofen treats inflammation unless the patient has renal impairment, is on blood thinners, has a history of gastrointestinal bleeding, or is in the third trimester of pregnancy. The triple has no mechanism for any of those conditions. It stores the conclusion and amputates the situation.
Now raise the stakes. An autonomous vehicle encounters an obstacle in the road. (Obstacle → Requires → Hard Braking) is a perfectly valid triple. It is logically sound. But if the context is a wet, icy highway and the obstacle is a thousand-pound moose, hard braking at 70 mph sends the vehicle into an uncontrolled skid, and the moose’s body comes through the windshield. Swerving is required. The contextual coordinate, the one the triple has no room for, is the difference between a safe stop and a fatality.
Context collapse is not a database error. It is a fatal flaw in machine reasoning.
The triple is the geometry of encyclopedias: a vacuum-sealed abstraction of reality. It represents what philosophers of language call eternal sentences, propositions true regardless of when, where, or by whom they are uttered. But almost no actionable knowledge is eternal. Knowledge that matters to a surviving organism, or to a functioning AI system, is knowledge about what is true here, now, for this agent, under these conditions.
Wikidata introduced qualifiers. The academic community has developed hyper-relational knowledge graphs using systems such as StarE, HINGE, and NaLP. These are real improvements. But they remain flat annotations bolted onto a fundamentally context-free structure. A qualifier labeled “start date: 1789” is metadata for the triple. It is not context woven into the representation. A Post-it note stuck to a photograph is not the same thing as the light conditions under which the photograph was taken.
I call this the Tragedy of the Naked Triple. A triple without context is an orphaned fact. We perform this stripping operation constantly: when we scrape text for training corpora, parse documents into RDF, or extract entities from unstructured text. Each time we perform contextual amputation, we systematically remove the situational reality in which a fact was embedded, leaving behind a husk that is logically valid but practically unreliable.
For seventy years, we have been building Semantic AI. The Quadruple is the first step toward Episodic AI, toward machines that do not merely know things, but remember being somewhere when they learned them.
III. The Contextual Processing Network
Biology never made this mistake. Evolution does not care about universal truths. It cares about situational survivability. Knowing that a rustling bush might mean a predator is useful. Knowing that the bush is dangerous only at night, in the dry season, when you are alone and downwind, is intelligence.
The vertebrate brain has been encoding context for at least 200 million years. Three regions work in concert every time you learn something new.
The Hippocampus provides the spatiotemporal grid: the “where” and “when” of an experience. Howard Eichenbaum’s 2004 work established it as the brain’s relational and contextual framework for memory organization. When the scuba divers encoded words underwater, the hippocampus was not filing words into a mental dictionary. It was mapping those words onto the water itself: depth, pressure, temperature, the visual field of murky Scottish ocean.
The Amygdala tags incoming information with emotional valence, the survival relevance of what is being experienced. This is not optional metadata. It is a weighting function that determines how deeply a memory is consolidated and how readily it is retrieved.
The Orbitofrontal Cortex integrates incoming experience against prior expectations. It computes prediction error: the difference between what you expected and what actually happened. When reality deviates from prediction, the brain allocates additional encoding resources. The surprise is part of the memory.
We know the binding mechanism. Lisman and Jensen’s 2013 work on the theta-gamma neural code showed that high-frequency gamma oscillations (30-120 Hz) are nested within slower theta oscillations (4-12 Hz) in the hippocampus, forming a cross-frequency coupling pattern that encodes multiple items in ordered sequences. Think of theta as the conductor and gamma as the individual instruments. Each gamma subcycle within a theta cycle carries a distinct piece of information, and the phase relationship encodes the binding: the contextual relationship between co-occurring elements. This theta-gamma coupling is a temporal binding protocol that weaves spatial context, emotional weight, and prediction error into a unified representation at the precise moment of encoding.
The brain does not store triples. It stores contextual manifolds. We might call them Quadruples as a provocation against the Naked Triple, but their geometry is not a list. It is a topology: C[e_i → r → e_j]. The context (C) does not sit next to the fact. It acts as the gravitational field in which the relationship is permitted to exist.
IV. Context as Gravity, Not Metadata
If the brain stores contextual manifolds, what does it mean to build an artificial system that does the same?
Here is where we must be precise, because the obvious objection is immediate. Any data engineer reading this will say: “We already do this. It is called an RDF N-Quad, or a Neo4j Property Graph. You just add a fourth field.” And they would be right to object, because if a Quadruple is just a triple with a Post-it note stuck to it, we have solved nothing. We have given the Naked Triple a jacket without changing its anatomy.
The difference is structural, and it matters.
An N-Quad or a property graph annotation is declarative. A machine reads it and applies a rule. “Valid from 1789 to 1797.” The context sits beside the fact as a passive modifier. The machine consults the tag, applies a filter, and moves on. The tag does not change the fact’s relationship to other facts. It does not reshape the space the fact inhabits. It is, to use the language of the scuba divers, a label that says “learned underwater.” It is not the water.
A contextual tensor is gravitational. It physically warps the space in which entities exist. In a high-dimensional embedding space, the concepts “Ibuprofen” and “Inflammation” do not simply sit near each other with a tag that says “unless renal impairment.” Instead, the contextual manifold of “healthy kidneys” bends the latent space, bringing them together. The contextual manifold of “Stage 4 CKD” curves them apart. The context is not a column in a database. It is the topology of the space the memory inhabits.
Think back to the moose on the highway. In a flat Semantic AI, “Obstacle” and “Hard Braking” are tightly clustered in vector space. But inject the contextual tensor of “Icy Road,” and the topology warps. The gravity of “Icy Road” bends “Hard Braking” so far from “Obstacle” that they become inaccessible to one another. “Swerving” gets pulled into the foreground instead. The context did not filter a rule. It fundamentally altered the physics of the machine’s reasoning.
Three architectural ideas make this concrete.
Temporal Embeddings: The Artificial Hippocampus. In a standard knowledge graph, an entity occupies a fixed point in a flat, Euclidean vector space. In a contextual manifold, the space itself is non-Euclidean. The temporal context operator does not simply add a translation vector to slide a point across a flat plane. It alters the distance metric itself, curving the latent space so that entities that were close under one temporal context become distant under another. A pharmaceutical compound in 2005, before a black-box warning, occupies a different position in the space of medical knowledge than the same compound in 2025. It is not a static point that has been nudged sideways. It is a mass moving through a space warped by the contextual gravity of time. This is the beginning of an artificial hippocampus: a system that maps knowledge onto a spatiotemporal grid rather than filing it into a flat index.
Contextual Tensor Decomposition: The Artificial Orbitofrontal Cortex. Traditional knowledge graph embeddings operate on two-dimensional matrices (entity-by-relation). Context graphs operate on higher-order tensors that factor in contextual dimensions. Where a triple defines a point on a plane, a quadruple defines a point in a volume, and the additional dimensions encode the conditions under which that point is valid. This is the architectural analog of prediction error: the system maintains a structured representation of the conditions that must hold for a fact to remain true, and when reality deviates from those conditions, the fact’s position in the space shifts accordingly.
Context-Aware Graph Attention: The Artificial Amygdala. Standard graph attention computes weights based on the structural neighborhood: which nodes are connected to which. Context-aware attention extends this by attending simultaneously to structural neighbors and contextual attributes. The system does not ask “what is connected to this entity?” It asks, “What is connected to this entity that is relevant given the current situational context?” This is what the amygdala does in biological memory: it weights incoming information by survival relevance, foregrounding what matters and suppressing what does not. Context-aware attention is the computational equivalent: a relevance weighting function that reshapes the graph’s topology around the situation at hand.
This is the difference between Semantic AI and Episodic AI. A semantic system looks up facts in a flat index. An episodic system reconstructs a situation and asks what was true inside that situation. The scuba divers did not retrieve words from a dictionary. They re-entered the water, and the water gave the words back.
V. Catastrophic Context Collapse
Large Language Models are linguistically fluent and structurally impoverished. They process linguistic context well through self-attention over token sequences. But they cannot distinguish what they know from what they hallucinate, what is current from what is outdated, what is universal from what is situational.
Because LLMs train on all contexts simultaneously, they exist in a state of Contextual Superposition. Every contradictory truth from every situation is averaged together into a single set of weights. Ibuprofen treats inflammation (true for healthy patients) and ibuprofen damages kidneys (true for CKD patients) collapse into a blurred statistical mean. The model inhabits a nowhere state: omniscient in aggregate, unreliable in any specific situation.
I call this Catastrophic Context Collapse.
It is the statistical averaging of contradictory truths from different situations into a single, undifferentiated response.
This is not a bug that will be fixed by scaling. Adding more parameters deepens the superposition; it does not resolve it.
And RAG does not fix it either. The standard defense is that Retrieval-Augmented Generation solves the context problem by fetching relevant documents and injecting them into the prompt. But think about what RAG actually does. It behaves like Leonard Shelby, the protagonist in the film Memento.
Afflicted with anterograde amnesia, Leonard survives on two external systems. His tattoos are permanent, painfully acquired, and nearly impossible to update. His Polaroids are ephemeral, taped-on contexts scattered around his motel room. Every morning, he reconstructs a fragile, temporary understanding of his situation from these artifacts. As soon as the scene ends, he forgets everything.
The modern AI stack mirrors this tragedy exactly. Model pre-training is Leonard’s tattoos: a slow, brutally expensive process of permanently burning facts into the weights. RAG is his Polaroid system. We frantically retrieve documents and tape them into the context window so the model can function for the next inference pass. The moment the prompt clears, the Polaroids burn, and the amnesia returns.
The human brain does neither. It physically alters its architecture to store context. The hippocampus rewires. The theta-gamma coupling binds. The water becomes the geometry. RAG describes the water to the LLM in a text block.
That is not memory. That is a caption.
Hubert Dreyfus saw this coming in 1972. In What Computers Can’t Do, he argued that human expertise relies on embodied, situated understanding that cannot be reduced to context-free rule-following. Michael Polanyi got at the same thing from a different angle in The Tacit Dimension (1966): we know more than we can tell, and the surplus is precisely the contextual, episodic knowledge that resists formalization as propositions. The triple is a proposition. The quadruple is the beginning of an architecture that can represent what propositions alone cannot.
VI. The Silent Consensus
Across the vanguard of AI infrastructure, a quiet consensus is forming: the era of flat data is ending.
Diffbot’s KG-LM Accuracy Benchmark showed that, for schema-bound questions requiring relational traversal, vector-based RAG scored 0. Not low. Zero. The architecture could not answer questions that required following the connections between entities. Graph-augmented retrieval did not merely improve on this; it operated in an entirely different category.
Zep AI’s Graphiti framework introduced a bi-temporal model that tracks both when events occurred and when they were ingested. Every graph edge carries explicit validity intervals. When new knowledge conflicts with existing knowledge, the system does not overwrite. It preserves the historical state and marks the transition. This is the beginning of temporal episodic structure in machine memory: not just what is true, but what was true, and when the transition happened.
The MemOS project proposed treating memory itself as a schedulable system resource, distinguishing parametric memory (in the weights), activation memory (in the runtime state), and plaintext memory (in external stores), with managed transitions between them. It is, in effect, an attempt to build an operating system for artificial memory, as Unix was for computation.
These systems are groping their way, independently and from different engineering starting points, toward the biological truth Tulving identified fifty years ago. They are reintroducing the timeline, the interval, and the expiration date to machine knowledge. They are, each in their own partial way, trying to give AI back the water.
None of them has built the full Quadruple yet. But the convergence is unmistakable.
VII. The Fourth Dimension
We have taught machines how to speak. We have taught machines how to categorize. But we have not yet taught machines how to be somewhere. How to know that what is true here is not true there. How to know that what was true yesterday may not be true today.
Storing human knowledge as a triple is like storing a symphony by writing down the notes but deleting the tempo, the acoustics of the hall, and the silence between the beats. You have preserved the information. You have destroyed the meaning.
AGI will not be achieved by stacking more layers on a language model. It will not be achieved by building a trillion-edge knowledge graph of context-free triples. It will be achieved when artificial systems stop building Semantic AI and start building Episodic AI. When they abandon the Naked Triple and embrace the Quadruple. When they learn to store not just what is true, but where, when, for whom, and under what conditions it is true.
It is time to add the fourth dimension. It is time to teach machines that water is not a tag in memory. The water is the geometry of the memory itself.
Next in the series: Part 2, “The Theta-Gamma Engine,” on the neuroscience of contextual binding and what it takes to build an artificial hippocampus.