The Predictive Map — How the brain uses the Past to Prepare for the Future
Overview
How the Brain Uses the Past to Prepare for the Future
There is a question we are tempted to ask about memory, and a better question hiding behind it.
The tempting question is: what does memory contain? And the textbook answer is ready to hand. Memory contains the past — episodes, facts, skills — sorted into the tidy boxes of a taxonomy that you will meet later in this unit. Declarative and non-declarative; episodic and semantic; the whole orderly cabinet. It is a satisfying answer, and it is almost entirely beside the point, because “episodic memory is memory for past episodes” is very nearly a tautology. It tells us what the word means. It does not tell us why brains bother.
The better question is the one an evolutionary biologist would insist on: what is memory for? Not what does it store, but what does storing buy the animal that stores it. And the moment you ask that question, the tidy cabinet starts to look like the wrong piece of furniture, because the answer cannot be “memory is for holding the past.” Holding the past is metabolically expensive and, on its own, useless. The past is over. No animal was ever made fitter by the bare fact of an accurate record. Animals are made fitter by what the record lets them do next.
This unit is organized around a hypothesis — and it is a hypothesis, something we are going to put on trial against the evidence rather than a conclusion you are being asked to accept on arrival: that the function of memory is prediction. That the brain remembers the past in order to prepare for the future; that the past is not the destination but the database from which the future is inferred. If that is right, then a great many things that look unrelated — food-caching birds, Tolman’s rats, place cells and grid cells, patients who have lost their hippocampus, the strange constructive errors of human recollection — should fall into a single line. Over this unit we will test whether they do, one link at a time, and we will let the evidence push back where it resists.
A word to those of you who came to this unit expecting human memory — expecting Henry Molaison, the man known for fifty years as H.M., whose story opens most textbooks’ treatment of this subject. We will get there, and the wait is deliberate. The argument of this unit is that you will understand human autobiographical memory better if you do not start with it — if you first watch the system that produces it being assembled, across half a billion years, out of something far older and more basic than the remembering of one’s own life. So we begin not with a human remembering a birthday, but with a bird hiding a worm. H.M. is coming. He will mean more when he arrives.
The bird that hides a worm
A Western scrub jay takes a wax-moth larva, carries it some distance, and buries it. Hours or days later it returns and digs it up. This is caching, and the food-hoarding birds — the jays, the nutcrackers, the chickadees — do it on a scale that should give us pause: a single Clark’s nutcracker may make many thousands of caches across a season and recover them months later, through snow, from memory.
Now ask the two questions. The tempting question — what does the bird’s memory contain? — has a famous answer, from the work of Nicola Clayton and Anthony Dickinson. The bird remembers what it stored, where it stored it, and when it stored it. We know this because of an elegant experiment: scrub jays were allowed to cache both perishable wax worms, which they prefer, and non-perishable peanuts. When the worms were still fresh, the jays dug for worms. When enough time had passed that the worms would have rotted, the jays switched and dug for peanuts instead. They were tracking not just location and content but the age of each cache. Clayton and Dickinson called this “episodic-like memory” — what, where, and when, bound together — and it was a small earthquake, because that triad had been claimed as the signature of human episodic memory, the kind that was supposed to be ours alone.
But the tempting question has led us slightly astray, and the better question puts us right. Why does the jay hold what, where, and when? Not to reminisce. The jay is not sitting on a branch enjoying the memory of an autumn afternoon. The jay holds those three facts because, taken together, they answer a question about the future: is there edible food available to me, and where? A fresh worm here; a rotted worm there, not worth the trip; a peanut over there that keeps. The what-where-when is in the service of a forecast. The memory is about a past event, but its biological job is to shape a future action. Notice that this is already true in your own intuitive formulation of what a spatial memory is. “This is where the food was; it is likely to be there again” — read that sentence carefully. The second clause is not a memory. It is a prediction. The memory is only the premise. The prediction is the conclusion, and the conclusion is the point.
There is a recent and intriguing glimpse of what this might look like in the bird’s brain, though it comes with a caution we should state plainly: it is a single study, very new, and a careful reader should hold it lightly until it is replicated and extended. In 2024, Selmaan Chettih, Emily Mackevicius, and Dmitriy Aronov recorded from the hippocampus of black-capped chickadees while the birds cached and retrieved seeds in a laboratory arena studded with more than a hundred hiding sites. They reported that each individual caching event was marked by a sparse, distinctive pattern of firing across a small set of hippocampal neurons — a “barcode,” in their word — apparently unique to that one cache. Two caches a few centimeters apart got different barcodes; two caches at the very same site on different occasions got different barcodes too. And when the bird later returned to a site, the barcode for that specific cache appeared to flicker back on, as if the act of arriving had pulled up the file.
What makes the finding suggestive — and it is no more than suggestive yet — is what the barcode seems to sit on top of. The chickadee hippocampus also contains an ordinary spatial map: place cells, coding where the bird is in the smooth, continuous way place cells do, nearby locations producing similar activity. The barcodes did not replace that map; they appeared to ride on it. The spatial scaffold says where; the barcode, layered onto it, would say which event, which seed, which moment. If this picture holds up, it is a small, vivid instance of the larger pattern this unit will argue for — a spatial map, evolved for finding one’s way, with the memory of particular events indexed onto it — the map first, the episode built upon it. But one study does not establish a thesis, and we cite it here as a pointer, not a proof.
The map beneath the memory
Where did that map come from? Not from birds, and certainly not from us. It is far older — and the evidence that it is older is not a philosophical argument but a stack of experiments, which are worth seeing in some detail, because they are the empirical foundation on which this whole unit stands.
The structure doing this work is the hippocampus, and a central project of this unit is to follow it backward in time. Using molecular markers — the signatures of gene expression that let us recognize the same piece of tissue across lineages that diverged hundreds of millions of years ago — we can identify a hippocampus-like region not only in mammals but in birds, in reptiles, in amphibians, and down into the fish. We will spend a later chapter on exactly how that identification is made, because it is one of the places where the “brain-first, evolution-first” commitment of this book does its hardest and most rewarding work. For now the important fact is that the structure is genuinely ancient and genuinely conserved. The hippocampus is not a mammalian invention, still less a human one.
And everywhere we find this structure in the older lineages, we find it doing something that looks much less like remembering one’s life and much more like finding one’s way and finding one’s food. The evidence is behavioral and surgical, and it is strikingly consistent across the vertebrate tree.
Start with a fish. Place a goldfish in a task where it must locate a spot defined by its relationship to the surrounding layout — a spot it can only find by holding a kind of map of the space, what is called an allocentric or world-centered representation. A normal goldfish learns this readily. Now ablate the lateral region of its telencephalic pallium, the part that comparative anatomists and developmental biologists identify as the fish’s hippocampal homolog, and the ability collapses. Crucially, the deficit is selective: the same fish can still swim straight to a single conspicuous cue, a beacon it can simply approach. What it has lost is not the use of its eyes or its motivation to find food — it is the map. The relational, layout-based representation is gone; the simple cued approach remains. This is the work of Cosme Salas, Fernando Rodríguez, and their colleagues in Seville, and it is about as clean a demonstration as one could want that this ancient structure is, in a fish, an organ of spatial mapping. (The same group has shown that the fish hippocampal homolog is also needed to bridge time — to associate two events separated by a gap — which will matter later, when we ask whether the ancestral function was purely spatial or something more general. We will not duck that complication; we will meet it head-on in due course.)
The story repeats as we move up the tree. In reptiles — lizards and turtles — lesions to the medial cortex, the reptilian hippocampal homolog, impair spatial navigation and the learning of places while sparing simpler forms of learning. In birds, the dependence is dramatic, and it brings us back to caching. Food-storing birds live or die by their spatial memory, and their brains show it. David Sherry and others demonstrated that damaging the avian hippocampus devastates a bird’s ability to relocate its caches — the bird will still cache, and still search, but it can no longer remember where. More striking still is a comparative pattern that connects ecology directly to anatomy: across bird species, the ones that depend most heavily on stored food tend to have larger hippocampi relative to their brains, and within a species the structure enlarges over the season of heaviest caching and in the individuals who cache most. The hippocampus grows to meet the spatial demand placed upon it. Your own lecture noted the genetic version of this result in Clark’s nutcrackers — that the genes most associated with the best-performing cachers are genes that influence hippocampal size. Selection has been tuning the size of this structure to the difficulty of a navigation-and-retrieval problem.
Put these together and the conclusion is hard to escape, and it is the deep evolutionary fact on which the unit rests: the hippocampus is, in origin, an organ of navigation. Long before any animal used it to remember the story of its own life — long before there were animals with lives in that sense at all — this structure was being used to build and consult maps of the world in the service of finding what the body needed. The memory of events, when it finally appears, will turn out to be a passenger on a much older spatial vehicle.
And navigation, properly understood, is already a forward-looking computation — which is the first hint that our hypothesis about prediction might reach all the way down. To answer “where am I?” in a way that does any good, an animal must be able to answer “where will I be if I move this way?” The machinery for that — for tracking which way you are facing and how far you have travelled — comes online in a young mammal astonishingly early. Head-direction cells, which fire according to the direction an animal faces, are already adult-like in rat pups before they open their eyes, before they have explored anything, driven by the vestibular sense of the body’s own movement. The cells that map external space mature later, built upon that earlier, bodily sense of heading and motion. The brain wires up the body-in-motion first, because tracking your own movement is the precondition for predicting where that movement will take you. Even at its evolutionary and developmental root, the map is built for going somewhere. Navigation is prediction’s first form.
The intellectual lineage here runs through one of the great rehabilitations in the history of psychology. In the 1940s, when behaviorism held that an animal was a bundle of stimulus-response reflexes, Edward Tolman insisted that a rat running a maze was building something internal — a cognitive map — that let it take shortcuts it had never been trained on, infer routes it had never walked. The idea was treated with suspicion for decades. Then, in 1971, John O’Keefe — a psychologist, it is worth saying, not a card-carrying neuroscientist — put electrodes in a rat’s hippocampus and found neurons that fired only when the animal was in a particular place. He had found Tolman’s map, written in single cells. O’Keefe and Lynn Nadel called their 1978 book The Hippocampus as a Cognitive Map, and the modern science of this structure begins there. The place cells were joined, over the following decades, by grid cells, which tile space in a hexagonal lattice and supply something like a coordinate system; by head-direction cells; by boundary cells that mark the edges of the world. The discovery of grid cells brought a Nobel Prize to Edvard and May-Britt Moser, sharing it with O’Keefe.
So the map is real, it is ancient, and it is cellular. But a map is a static thing, and prediction is not. How does an animal get from having a map to using it to see into the future? There is a beautiful intermediate step, and it too comes from Tolman.
Running the route before you run it
Tolman noticed something that his rats did at the junctions of a maze. Arriving at a choice point, a rat would often pause and look one way, then the other, then back again — hesitating, surveying, as if weighing the options. He called this vicarious trial and error: the animal seeming to try out each path “in its head” before committing its body to one. To a behaviorist this was an embarrassment; reflexes do not deliberate. To Tolman it was the cognitive map being consulted.
For seventy years this was a behavioral curiosity with no mechanism. Then, in 2007, Adam Johnson and David Redish decoded the activity of place cells in a rat paused at exactly such a choice point — and watched the map come alive in precisely the way Tolman’s interpretation demanded. The representation of the animal’s location did not sit still at the junction where the rat’s body actually was. It swept forward, down first one arm of the maze and then the other — place cells representing positions ahead of the animal, on paths not yet taken, activating in sequence as if the rat were mentally running each route in turn. The cognitive map was being read out into the future, one possible path at a time, while the body stood still and deliberated.
This is the hinge of the whole unit, so let us be clear about what it shows. The hippocampus is not merely storing where the animal has been. It is simulating where the animal might go. The map has become an engine for generating possibilities — for running the route before running it. And the developmental fact noted earlier now pays off: this forward sweep is built on path integration, on the self-motion machinery that matures first, because to simulate “if I go down this arm I will arrive there” you need exactly the apparatus that tracks heading and distance. The forward sweeps appear most strongly early in learning, when the animal is uncertain and the choice matters, and they fade as the route becomes habitual and no simulation is needed — which is just what you would expect of a deliberative system that earns its keep only when the future is in doubt.
We can now write the progression that organizes this unit. Each step is the previous one turned toward the future:
Navigation — Where am I? ↓ Spatial memory — Where was the food? ↓ The predictive map — Where is food likely to be now? ↓ Vicarious trial and error — Let me run the route forward before I commit. ↓ Mental simulation — What will happen if I go there? ↓ Human episodic memory — What happened before, and what does it imply about what comes next?
Read that sequence from top to bottom and notice what has happened to autobiographical memory — the remembering-your-birthday kind that we started by calling the tempting answer. It is not at the summit of the evolutionary story. It is one rung from the end, and it is not even the destination. The destination is prediction. Autobiographical memory turns out to be one mechanism, among others, that an unusually clever lineage evolved in the service of seeing ahead. We did not evolve the hippocampus to keep a diary. We inherited a map for finding food, and built the diary on top of it.
Why the hypothalamus needs a map
There is a part of the brain whose job is to know what the body needs. The hypothalamus monitors the internal state — the fall in blood glucose that we feel as hunger, the rise in osmolarity that we feel as thirst, the drift in core temperature, the demands of safety and reproduction — and it generates the drives that push an animal to do something about them. It is, in a sense, the organ of need.
But need on its own is blind. The hypothalamus can register that the body is short of water; it cannot tell the body where the water is. For that, the animal must turn to the structure that holds the map of the world and the memory of where, in that world, needs have been met before. Set the two side by side and the relationship is almost a conversation:
The hypothalamus asks: What do I need? The hippocampus answers: Where have I found that before — and where am I likely to find it now?
This is a far more biologically honest picture than the one in which memory is a filing system for the past. The map exists because the body has needs that must be satisfied in places, and satisfied not just now but tomorrow and next season. Here the argument of this unit joins hands with the argument of Unit 2. There we developed the idea of allostasis — that good regulation is not merely reactive, not merely the restoring of balance after it has been lost, but anticipatory: predicting need before it becomes urgent and arranging, in advance, to meet it. And there we made the case that exploration is not idle behavior but an allostatic investment — that when the homeostatic drives are quiet, an animal is released to wander and gather information about where, in the world, its future needs might be met. That argument was made in Unit 2 in the language of drives and the hypothalamus. We can now see what it requires on the other side of the conversation. An animal can only invest in exploration if it has somewhere to put what exploration returns. The hippocampal map is that repository. It is where the fruits of exploration are stored against future need — and so it is the structure that makes allostasis, as we described it in Unit 2, physically possible. The exploring animal of the homeostasis unit and the mapping animal of this one are the same animal, seen from two sides. The hypothalamus supplies the reason to explore; the hippocampus supplies the place to keep the map that exploration builds.
Put the relationship in a sentence and it inverts the textbook picture: an animal that waits until it is desperately thirsty to begin learning where water is will die; an animal that, in its unpressured hours, has already laid down a map of where water reliably appears can predict and pre-empt. The predictive map is how a body with future needs prepares, in the present, to meet them. The hippocampus is allostasis made spatial.
The tell-tale flaw: why memory is not a video recorder
If you wanted one piece of evidence that the function of memory is prediction rather than archival storage, you could hardly do better than a property of memory that has long embarrassed the archival view: human memory is constructive. It is not a recording. It is reassembled, each time, from parts — and in the reassembly it distorts, blends, and invents. We misremember who was there. We import details that never happened. We are confident about things that are false. If the purpose of episodic memory were to keep an accurate record, this would be a catastrophic design flaw, a recorder that quietly rewrites its own tapes.
But suppose the purpose was never accurate recording. Suppose the purpose was simulation — building, on demand, a coherent model of an event in order to answer “what is likely to happen?” Then the very same property stops being a flaw and becomes a feature. To imagine a future you have never experienced, you cannot retrieve it — it has not happened. You must construct it, by recombining fragments of things that did happen into a new, plausible scene. A memory system optimized for that job would naturally be flexible, recombinatorial, and a little loose with the literal facts, because its product is not the past but a usable picture of what might come. The “errors” of memory are the fingerprints of a machine built for prospection.
This is precisely the convergence reached, from the human side, by Daniel Schacter and Donna Rose Addis, whose constructive episodic simulation hypothesis holds that remembering the past and imagining the future are two uses of one constructive system — and that memory’s constructive distortions are the price, and the signature, of its future-directed function. The neuroscience cooperates: imagining future events and remembering past ones recruit much of the same core network of brain regions, with the hippocampus at its center. Most strikingly, Demis Hassabis and Eleanor Maguire reported that patients with hippocampal amnesia, asked simply to imagine a novel scene — a beach, a museum — could not do it in the normal way. Their imagined scenes came apart into disconnected fragments, lacking the spatial coherence that binds a real scene together. The damage that had stolen their past had also stolen their future. A structure we had filed under “memory for what happened” turned out to be necessary for picturing what hasn’t.
Honesty requires a flag here, because this is live science and you should see its edges. Not every study agrees: Larry Squire and colleagues found some hippocampal patients who could imagine future events in reasonable detail, and argued that the dramatic failures in other patients owed to damage spreading beyond the hippocampus itself. The dispute is unresolved, and it is exactly the kind of disagreement a developing field should have. It also marks the boundary of how strong a claim we are entitled to make — which brings us to the necessary caution.
How far the claim reaches
It is worth distinguishing the bold version of this unit’s thesis from the version we can actually defend, because the gap between them is where careful thinking lives.
The bold version — the one that makes the best lecture and the worst overstatement — is this: episodic memory did not evolve to remember the past; it evolved to anticipate the future. It is a genuinely provocative reframing, and it organizes everything. But notice what kind of claim it is. It is a claim about evolutionary history — about what the system was originally selected for — and claims about ancient selection pressures are inferences, not observations. The evidence we actually have is of a different and more modest kind: that memory and future-thinking share neural machinery; that the machinery is demonstrably used for prospection right now; that the structure is ancient and spatial in origin; that its constructive character fits a prospective function. All of that is solid. None of it strictly proves that prospection, rather than something else — relational binding for navigation, say, with future-simulation emerging later as a bonus once the map grew rich enough — was the original selective pressure. A scrub jay’s cache memory is consistent with our story; it is also consistent with “spatial-relational memory that happens to permit retrieval.” The bird does not settle the question of origins, and we should not pretend it does.
So hold the bold claim as the organizing provocation and the modest claim as what you would defend in a seminar. That is not timidity; it is the actual shape of the knowledge. What we can say without flinching is the functional point, and it is enough: whatever its deepest evolutionary origin, the hippocampal map is profoundly, demonstrably oriented toward the future. It simulates paths not yet taken. It constructs scenes not yet lived. It answers the body’s needs not only by recalling where they were met but by predicting where they will be. The past is its raw material. The future is its product.
There is one more guardrail. “Prediction” is a seductive word precisely because almost anything can be redescribed as prediction, and a theory that explains everything explains nothing. We are not claiming the vague grandeur that “the brain is a prediction machine.” We are making a specific claim with a specific organ: that the hippocampal-entorhinal system builds a model whose job is to forecast the location and availability of the things a needful body requires, and to simulate the consequences of going to get them. That claim is specific enough to be wrong. If hippocampal damage spared the forecasting of need-relevant locations while abolishing some other, unrelated prediction — or the reverse — the claim would be in trouble. It is the specificity, anchored to scene construction and spatial forecasting and tethered to the hypothalamus’s catalog of needs, that keeps the idea honest and testable rather than letting it float free into slogan.
What follows
With that frame in place, the rest of the unit can proceed — and the familiar landmarks of the textbook will look different in this light. When we meet the patient H.M., whose surgery removed his hippocampus and with it his ability to form new memories, we will ask not only why he could not remember the past but why, as later work on such patients revealed, he could not properly imagine the future either. When we lay out the taxonomy of memory systems — episodic, semantic, procedural — we will treat it as a useful map of what gets stored, while never forgetting that the storage is in the service of something else. When we examine consolidation, and the long, strange process by which the hippocampus trains the rest of the cortex and then bows out, we will see a system arranging its database for fast and flexible future use. And when we return, at the unit’s end, to the question of what makes human memory distinctive, we will be able to give an answer that places us in continuity with the jay and the rat rather than on a pedestal above them: we are the animal whose predictive map grew rich enough to run very far forward — to simulate not just the next turn in the maze but next year, not just where the food is but who we will be when we get there.
The brain remembers the past. But it was never for the past. It was always for what comes next.