43 The Predictive Map
Overview: How the Brain Uses the Past to Prepare for the Future
43.1 From possible futures to learned structure
Unit VI ended with a problem that Chapter 41 made explicit: a controller cannot use an option it has never discovered or inferred. Frontal systems can allow a rule, an absent goal, or a possible consequence to govern what the body does now. But those systems do not manufacture a possible world from nothing. Before an animal can choose a route, it must have learned or inferred that the route exists. Before it can imagine arriving somewhere, it must know something about the states through which it could pass, the events that tend to occur there, and the actions that connect one situation to another.
That problem is easy to hide inside the ordinary word memory. Textbooks divide memory into episodic and semantic, declarative and nondeclarative, short-term and long-term. Those distinctions matter because the nervous system does not learn a bicycle skill, recognize a familiar object, maintain a rule for five seconds, and remember a particular breakfast through one interchangeable mechanism. The action unit has already treated skills, habits, and reinforcement learning. The frontal unit treated the temporary maintenance of goals and rules. This unit concentrates on the hippocampal and medial-temporal systems that bind places, contexts, events, and relations.
The taxonomy tells us that memory is not one thing. It does not yet tell us what a hippocampal memory contributes to control.
A useful first answer is that remembering changes what the animal can do next. A cached seed can be recovered. A dangerous route can be avoided. A shortcut can be taken although it has never been traveled from end to end. A present cue can be interpreted according to the event in which it occurs. An outcome that is not visible can be located by reconstructing the relations that lead to it.
But memory serves the future is still too broad. Every useful form of learning affects later behavior. The distinctive claim of this unit concerns the kind of structure the hippocampal–entorhinal system learns:
The hippocampal–entorhinal system learns structured relations among places, states, events, and transitions. Coupled to systems for bodily need, value, goals, and action, that structure allows an organism to locate itself, retrieve what happened in context, infer what may come next, and construct routes through situations that are not currently present.
The phrase predictive map names that structured representation. Map does not mean a picture viewed by an inner navigator. It means that relations among states are preserved well enough to support movement, inference, and generalization. Predictive does not mean that the hippocampus forecasts everything. It means that the representation is shaped by transitions: where the animal can go from here, what tends to follow what, and which states are likely to be encountered under a particular course of action.
This definition separates several interacting contributions that are often collapsed. A map can represent the states available to an animal and the transitions connecting them. Hypothalamic and interoceptive systems report what the body currently needs. Amygdalar, orbitofrontal, and striatal systems help assign salience, outcome identity, and value. Frontal systems maintain goals and alternatives. Sensorimotor systems turn one selected route into movement. These contributions overlap and change one another through recurrent connections. They are not successive processing stages or separate psychological faculties. The hippocampal formation supplies learned structure within that larger controller.
I begin with the evolutionary identity of the system, far from human autobiography. A bird hides food. A fish learns a place. A rat pauses at a junction while hippocampal activity runs down paths its body has not yet taken. Only after that older architecture is visible will we return to H.M., constructive recollection, and the human ability to assemble scenes that never happened.
43.2 An ancient spatial-relational system
The hippocampus did not arrive with human recollection, or even with mammals. The hippocampal formation belongs to an old part of the vertebrate pallium, although its adult shape differs greatly among lineages. In mammals, the hippocampal formation is laminated; in humans, it is folded into the medial temporal lobe. Birds possess a differently organized hippocampal formation along the dorsomedial pallium. The medial cortex of nonavian reptiles and parts of the teleost dorsolateral pallium have been compared with the hippocampal system through developmental position, gene expression, cell types, connectivity, and function [@abellan2014hippocampal; @tosches2018evolution].
43.2.1 Finding relatives through development
This is the problem introduced in Chapter 3: related brains can acquire very different adult shapes because evolution changes the timing, extent, and location of developmental processes. Homology means descent from a corresponding structure in a common ancestor. Similar appearance or similar behavior can suggest a comparison, but neither establishes that ancestry on its own. A bird need not possess a miniature mammalian hippocampus for its hippocampal formation to be homologous.
Development gives us a way to look beneath the adult shape. We can examine the embryonic medial pallium, identify its boundaries relative to neighboring tissue, and compare the molecular patterns that accompany the production and differentiation of its cells. The question is not simply whether two adult regions contain the same molecule. It is whether corresponding developmental territories deploy related combinations of molecules as their cells mature.
Two methods make those comparisons visible. In situ hybridization uses a labeled nucleic-acid probe to locate a particular RNA in a tissue section. It shows where a gene is being expressed at the RNA level. Immunohistochemistry, including fluorescent labeling, uses antibodies to detect a particular protein. It can show where the protein product occurs and whether two markers occur in the same cells. RNA and protein measurements are complementary; they need not have identical distributions or timing. Neither is, by itself, a record of a cell’s complete lineage.
Antonio Abellán and colleagues compared developmental gene-expression patterns in mouse and chicken. Combinations including Lef1, Lhx2, and Lhx9, considered relative to neighboring boundaries, helped identify medial-pallial territory. More restricted patterns involving Prox1, Lmo3, and Lmo4 helped distinguish candidate subdivisions corresponding to mammalian fields such as the dentate gyrus and the CA fields, subdivisions of the hippocampus proper. The investigators proposed dentate/CA3-like and CA1/subicular-like sectors in the chicken on that combined evidence, rather than by matching the silhouette of its adult brain to a mouse brain [@abellan2014hippocampal].
Protein markers extend the comparison to nonavian reptiles. Sara Jiménez and Nerea Moreno examined developing and adult turtle and frog medial pallium using molecular labeling. In the turtle, Prox1 protein was concentrated in medial cortex at late developmental stages and in adults, while Er81 and Lmo4 helped distinguish the neighboring dorsomedial territory. In comparison with mammalian patterns, these findings support proposed dentate-like and CA-like subdivisions. The changing pattern across stages matters: a marker absent early can appear as cells differentiate [@jimenez2022medialpallium].
No single stain is a hippocampal birth certificate. A marker may occur in several cell types or outside the hippocampal system. A missing signal may reflect developmental timing, an insensitive probe, or an antibody that does not recognize the protein reliably in that species. The useful evidence is therefore combinatorial: several markers, their position, their timing, and their relation to developing cell populations. Cell-type comparisons based on many expressed genes add another independent scale of evidence; single-cell RNA sequencing in reptiles has revealed relationships between reptilian and mammalian hippocampal cell populations [@tosches2018evolution].
Together with connectivity and comparative anatomy, these methods make a strong case for shared medial-pallial ancestry. Exact correspondence between an avian or reptilian subdivision and each mammalian subfield is a more specific claim and remains open to refinement. Expression patterns identify candidates; lineage tracing addresses where cells came from, and developmental perturbations test what a gene does. Those are related questions, not interchangeable experiments.
Gross position is especially misleading in ray-finned fishes. Development of the teleost pallium involves eversion, with growth and rearrangement that place the ventricular surface dorsally. This is more complicated than turning a mammalian hemisphere inside out. Tissue related by developmental origin can end up in an adult position that does not look medial. Comparative anatomy must reconstruct that transformation before assigning homology [@folgueira2012eversion].
43.2.2 What the conserved system does
The strongest conserved behavioral result concerns flexible spatial learning. Fernando Rodríguez and colleagues trained turtles and goldfish to locate a goal defined by its relation to the surrounding environment. The animal could not solve the problem by approaching one conspicuous beacon; it had to use the arrangement of cues to identify a place from different starting positions. Damage to reptilian medial cortex or the lateral pallium of goldfish impaired this place strategy while sparing approach to a visible cue [@rodriguez2002conservation]. The lesion did not abolish vision, movement, appetite, or every form of learning. It selectively damaged the use of relational environmental layout.
Together with the anatomical evidence, that result supports a spatial contribution conserved across these lineages. It does not establish that the ancestral system was spatial and nothing else.
Goldfish with lesions of the same hippocampal pallial territory are also impaired when two events must be associated across a stimulus-free temporal gap, while learning remains possible when the events overlap in time [@gomez2022trace]. In another experiment, goldfish first learned overlapping premise pairs—A over B, B over C, C over D, and D over E. Intact animals then preferred B over D, a relation they had never been directly taught. Lesioned animals retained the trained premise discriminations but failed this novel transitive inference [@soteloparrilla2025transitive]. The dissociation implicates this territory in using the learned premises to answer a new question. It does not, by itself, decide whether the fish constructs an explicit ordered map or solves the test through another form of relational learning.
The comparative conclusion is therefore richer than “navigation came first.”
Across vertebrates, the clearest conserved contribution of the hippocampal pallium is flexible spatial and relational memory. Navigation provides a particularly clear experimental window onto that contribution. Temporal and inferential findings in living fish make a solely spatial ancestral repertoire harder to assume, although living species do not give us a direct recording of ancestral function.
Physical navigation made the operation easy to see. A moving animal occupies one location, faces one direction, encounters boundaries, follows transitions, and arrives at outcomes. Space supplies a concrete geometry in which relations can be measured. The same general problem—binding elements according to the context and sequence in which they occur—does not have to remain confined to physical location.
Food-storing birds show how ecology can tune this old system. Across passerine species, food storers tend to have a larger hippocampal complex relative to brain and body size than nonstorers, and reversible disruption of the avian hippocampal formation impairs memory for stored locations [@krebs1989specialization; @shiflett2003inactivation]. In wild mountain chickadees, better spatial learning in juveniles predicts survival to the following winter, supplying direct evidence that natural selection can act on the relevant cognitive variation [@sonnenberg2019selection].
The relation between adult experience and gross hippocampal anatomy is less uniform. Some studies report seasonal increases in hippocampal volume or neuron number around periods of intense caching [@smulders1995seasonal]. Others find stable volume across sampled seasons or no anatomical difference after experimental manipulations that improve cache retrieval and spatial performance [@hoshooley2004stable; @pravosudov2002food]. The defensible synthesis is not that each bout of caching enlarges the hippocampus on demand. It is that dependence on spatial memory, cognitive performance, and hippocampal organization have repeatedly evolved together, while short-term anatomical plasticity varies with species, population, age, method, and timescale.
The account developed here treats navigation as a biological starting point for understanding the system, not as proof that all nonspatial functions arrived later. Evolution preserved and modified medial-pallial circuitry; how much of its relational repertoire was already present in common ancestors must be inferred. Human episodic memory belongs within that history without being reduced to a rat’s place field or a fish’s maze choice.
43.3 The bird that hides a worm
A scrub jay takes a wax-moth larva, carries it away, and pushes it into the ground. Later it returns and uncovers the cache. Food-storing birds repeat this behavior across many sites and can recover stored items after intervals long enough that direct sensory traces at the hiding place are of little use.
Nicola Clayton and Anthony Dickinson used that ecology to ask what information a jay could bind into one memory. The birds cached preferred but perishable wax worms and less preferred peanuts that remained edible. After a short delay, the jays searched for worms. After learning that worms decayed over a longer interval, they shifted their search toward peanuts. Recovery therefore depended jointly on what had been stored, where it had been stored, and when the caching event had occurred [@clayton1998episodic].
Clayton and Dickinson called the result episodic-like memory. The qualification matters. The experiment established a behavioral capacity to bind what, where, and when. It did not establish the bird’s subjective experience of recollection or the specifically human sense of mentally returning to a personal past. The behavioral achievement is nevertheless substantial. The bird did not merely learn that worms are good or that one location often contains food. It used the identity, place, and elapsed age of a particular cache to decide where searching was still worthwhile.
This is the future-directed force of an accurate record. The jay benefits from remembering where the worm was because that information changes whether digging there now is likely to succeed. The past event remains important in its own right: confusing two cache locations would be costly. But the memory earns its biological keep when it alters a later action.
A recent experiment makes the event structure unusually visible. Selmaan Chettih, Emily Mackevicius, Stephanie Hale, and Dmitriy Aronov recorded hippocampal neurons while black-capped chickadees cached and retrieved seeds from a laboratory arena. Each caching event was accompanied by a sparse, transient pattern of activity distributed across a small subset of neurons. The investigators called these patterns barcodes. Separate caches at the same location produced different barcodes; caches only a few centimeters apart also produced patterns that were largely unrelated. When the bird later interacted with or retrieved that particular cache, the corresponding pattern reappeared [@chettih2024barcoding].
The same neuronal population also carried smoother information about location. Nearby places produced more similar place-related activity, whereas neighboring caching events received distinct barcodes. A population could therefore carry both spatial context and a sparse pattern distinguishing one event from another.
That arrangement solves a genuine memory problem. A place code alone cannot distinguish yesterday’s seed from today’s seed hidden at the same site. A wholly unique event code, disconnected from spatial context, would make it difficult to retrieve the memory by returning to the place. Combining a relational context with a distinctive event index allows similarity and specificity to coexist.
Barcode reactivation was recorded during later interaction with the cache, not at a distance before the bird chose its destination. The experiment supports event-specific encoding and retrieval; whether these patterns help select a future cache destination remains to be tested.
43.4 A hippocampal formation, not a memory box
The word hippocampus is often used as though it named one compact storage device. In mammals, it consists of connected fields with distinct cell types and projection patterns; in humans, these fields occupy the medial temporal lobe.
In mammals, cortical information reaches the hippocampal formation largely through the entorhinal cortex and neighboring perirhinal and parahippocampal regions. Principal entorhinal routes arise in superficial layers: layer II supplies major inputs to the dentate gyrus and CA3, while layer III supplies CA1 and the subiculum. Dentate granule cells send sparse mossy-fiber projections to CA3. CA3 pyramidal neurons possess extensive recurrent connections and send Schaffer collaterals to CA1. CA1 and subicular outputs return to deep entorhinal layers and reach other cortical and subcortical targets, allowing hippocampal processing to modify the wider network. This familiar sequence is useful, but it is not a one-way conveyor belt. Direct entorhinal inputs bypass intermediate stages, hippocampal fields communicate along their long axis, and parahippocampal–subicular pathways form parallel loops rather than a single serial chain [@amaralwitter1989organization; @witter2000communication].
The system is also coupled to structures that textbooks often place in other chapters. Retrosplenial and posterior parietal cortex contribute information about scenes, landmarks, and orientation. Anterior thalamic and mammillary systems participate in navigation and episodic memory. Septal inputs help organize hippocampal rhythms and state. Amygdalar and hypothalamic connections bring emotional and bodily relevance. Striatal and frontal interactions allow learned relations to influence choice, rules, and action. What later appears as one memory is constructed through this recurrent network.
Several computational ideas have been associated with the subfields. Sparse dentate activity may help distinguish similar inputs. Recurrent CA3 circuitry may support rapid association and completion from partial cues. CA1 receives both processed hippocampal input and more direct entorhinal input, placing it well to register relations, sequences, and discrepancies. These are productive accounts of circuit bias, not labels to memorize as if each subfield performed only one verb.
The architecture explains why memory box is the wrong metaphor. A box stores a completed object in one place. The hippocampal formation binds distributed cortical information, differentiates overlapping events, reconstructs a relation from partial evidence, and returns changed activity to the wider system. A lesion does not simply empty a container. It removes a set of operations from a recurrent network.
43.5 Drawing a map from movement and sensation
Edward Tolman introduced the cognitive map while behaviorism still dominated American psychology. A rat, he argued, did not merely accumulate a chain of turns reinforced one after another. It learned relations among places well enough to change routes, approach a goal from a new direction, and take shortcuts that had never been directly reinforced [@tolman1948cognitive]. The map was inferred from flexible behavior.
John O’Keefe and Jonathan Dostrovsky then found hippocampal neurons whose firing depended strongly on where a rat was located [@okeefe1971spatial]. A place cell is active in one or several regions of an environment—its place field. The discovery did not reveal a tiny cartographic symbol corresponding to each location. It revealed a population code compatible with Tolman’s central claim: the hippocampus carries internal structure that can support behavior beyond a memorized stimulus–response chain.
Other signals help assemble that structure. Head-direction cells fire according to the direction the animal faces. Grid cells in medial entorhinal cortex show periodic fields that repeat across an environment in an approximately hexagonal arrangement. Border cells respond near environmental boundaries. Speed, self-motion, landmarks, odors, goals, and task context also influence hippocampal–entorhinal activity [@taube1990head; @hafting2005microstructure; @solstad2008borders].
These signals solve complementary problems. Vestibular, proprioceptive, and motor information allow path integration: updating position from the animal’s own movement. External landmarks correct the cumulative drift that path integration inevitably produces. Boundaries anchor the representation to environmental geometry. Direction and speed specify how the body is moving through that geometry. The map is continuously rebuilt from the relation between self-motion and sensory evidence.
Development reveals part of that dependency. When young rats first begin to explore beyond the nest, head-direction signals are already comparatively coherent. Place representations are present but continue to refine, while stable adult-like grid organization emerges later [@langston2010development; @wills2010development]. Directional and self-motion information is available early enough to provide a scaffold for later spatial organization. The timing alone does not establish a causal dependency, and it is not a replay of the order in which the functions evolved.
The adult code is also less tidy than a textbook gallery of cell types suggests. Many entorhinal neurons combine information about position, direction, speed, and boundaries, and their tuning adapts to behavior [@hardcastle2017multiplexed]. Place fields remap when context changes. The same physical corridor can be represented differently depending on the route being traveled or the episode in which it occurs [@wood2000episodes; @grieves2016routes]. Rewards and goals can distort the density and organization of fields. Grid patterns stretch, fragment, or reorient when environments change.
A useful map should behave this way. Euclidean coordinates alone do not tell an animal whether two visually similar corridors belong to different journeys, whether a barrier makes a nearby goal unreachable, or whether the same place is safe in one context and dangerous in another. The hippocampal map is not detached geometry. It is an embodied, learned representation of a world through which this animal can move.
43.6 What makes the map predictive
A map becomes predictive when it represents not only where the animal is, but the transition structure of the situation.
Four terms help separate the problem.
A state is a situation the animal can occupy. It may be a physical location, a moment in a sequence, or a task context that is not directly visible. A transition specifies which state tends to follow another after an action or event. A value specifies what currently matters: food may be valuable when hungry and irrelevant when sated. A policy specifies how the animal tends to act.
The hippocampal formation does not have to compute all four. Its central contribution can be the representation of states and their relations. Other systems can alter value or select a policy while consulting that structure. This separation explains how one learned map can guide different needs. The route to a water source does not have to be relearned each time thirst increases. Bodily state changes the value of destinations already represented in the map.
The successor representation provides one influential formal account. Instead of representing only the current state or storing a full model from which every future must be calculated anew, a successor code represents the states likely to follow the present one under a policy, with nearer and more probable successors weighted more strongly. A hallway leading repeatedly to a doorway will come to represent something of that doorway’s future occupancy. If the policy remains fixed, changing reward can revalue those predicted occupancies without first relearning them. Changed barriers or a changed policy require the successor representation itself to be updated.
Kimberly Stachenfeld, Matthew Botvinick, and Samuel Gershman showed that this framework can explain several features of place and grid activity that a purely geometric map leaves unexplained, including sensitivity to policy, barriers, and future occupancy [@stachenfeld2017predictive]. Human behavior in sequential tasks also carries signatures expected from successor-like representations, intermediate between a fully cached habit and complete model-based recomputation [@momennejad2017successor].
The successor representation is not a synonym for hippocampus and not the final theory of its code. Hippocampal populations also carry event identity, temporal context, goals, and sensory content, and an animal can plan in ways that exceed the current policy. The model matters because it makes predictive map precise enough to test. It asks whether neural similarity reflects not only physical proximity but expected transitions through the learned world.
The distinction also prevents all prediction from becoming one thing. Cerebellar forward models estimate the sensory consequences of commands over short delays. Dopamine-dependent systems learn prediction errors about value. Allostatic systems anticipate bodily demands. A hippocampal predictive map estimates relations and likely transitions among states. These computations interact, but sharing the word prediction does not make them interchangeable.
43.7 Running routes without moving
Tolman noticed that rats often paused at a junction and turned first toward one path, then another, before committing. He called the behavior vicarious trial and error. The animal appeared to sample alternatives while its body remained at the choice point.
Adam Johnson and David Redish later recorded ensembles in hippocampal area CA3 during such pauses. Decoded activity did not remain fixed at the rat’s current location. It swept ahead down one possible arm of the maze and then another [@johnsonredish2007paths]. The result supplied a physiological candidate for Tolman’s intuition: hippocampal activity can represent sequences through locations the body is not currently occupying. Recording such sweeps does not by itself show that they caused the ensuing choice.
Prospective sequence activity is not one uniform phenomenon. During active movement, place-cell firing can be compressed within cycles of the hippocampal theta rhythm, producing sequences that extend from recently occupied positions toward positions ahead. During quiet pauses and non-REM sleep, the hippocampus also generates brief sharp-wave ripples, often accompanied by compressed population sequences. Choice-point sweeps and ripple-associated sequences should not be treated as the same physiological event merely because the animal is pausing. Some sequences recapitulate a recent route in forward or reverse order. Others depict an alternative trajectory or a path toward a remembered goal [@pfeifferfoster2013future].
The words replay and simulation should therefore not be used as synonyms. A reverse sequence of a just-traveled route may support retrospective evaluation. A forward sweep during a choice can represent an available path. A sleep sequence may contribute to reorganization across hippocampal and cortical networks. The content and function depend on state, timing, learning, and the circuits with which the hippocampus is coordinated.
Causal evidence shows that the events matter. Selectively disrupting awake sharp-wave ripples impairs learning and performance in a spatial alternation task while leaving ordinary place-field firing intact [@jadhav2012ripples]. During learning, coordinated hippocampal–prefrontal replay distinguishes correct past and future paths from alternatives, linking internally generated sequences to a larger decision system [@shin2019replay]. These findings support an account in which hippocampal sequences contribute candidate trajectories within a larger decision network. Disrupting ripples establishes the importance of ripple-associated activity; it does not isolate the causal contribution of every decoded trajectory.
These sequences are sometimes described as the rat imagining its future. That phrase captures their prospective character, but it should not smuggle in a claim about conscious experience. The physiological result is enough: hippocampal populations can generate ordered representations of states through which the animal is not presently moving, and interventions on ripple-associated activity can impair memory-guided behavior.
The old linear story would place navigation at the bottom and human mental time travel at the top. The architecture suggests a different account. Several operations recur in different combinations:
- estimating the current state;
- binding events to context;
- learning transitions;
- generating sequences;
- inferring unobserved relations;
- recombining elements into a possible situation.
Navigation, cache recovery, relational inference, and episodic construction all draw on this family of operations. Human memory is not the summit of an animal ladder. It is one unusually elaborate use of an ancient relational system.
43.9 Exploration as map building
Chapter 15 and Chapter 41 approached the same behavior from two sides. Bodily state helps determine when open-ended information gathering is affordable and what kind of information is currently urgent. Frontal and subcortical systems arbitrate between exploiting a known option and sampling an uncertain one. This unit asks what exploration purchases.
The answer is not merely familiarity. Exploration reveals the states and transitions available to later control: where water collects, which route remains passable after rain, where cover interrupts a predator’s line of sight, which plant changes across seasons, which social partner can be approached, which cue predicts danger, and where a detour rejoins a known path. It also reveals uncertainty—places that have not been sampled, resources whose reliability is declining, and transitions that no longer behave as expected.
The information has future regulatory value. A route learned while the animal is safe can be used when escape becomes urgent. A water source discovered while hydrated can shorten search during dehydration. Alternative shelter matters most after the familiar shelter fails. The immediate costs of exploration—energy, time, exposure, and risk—are exchanged for a larger and more reliable set of later actions.
That is the sense in which this book proposes that exploration is an allostatic investment in future controllability. It does not mean that a curious animal explicitly calculates its future physiology. It means that selection can favor mechanisms that pay a present cost to improve the information on which later regulation depends.
Current need then reweights what the learned world affords. The map of a territory need not be rebuilt when the animal becomes thirsty. Thirst increases the value of states containing water and the routes that lead to them. Hunger, cold, threat, mating condition, and social context alter which destinations and transitions matter. The same relational structure can guide different policies as the body changes.
As a first approximation, the relationship can be compressed into a dialogue:
The body asks: What matters now?
The map supplies: Where, when, and through which route might it be found?
The full circuit includes hypothalamus, brainstem, amygdala, striatum, frontal cortex, sensory systems, and the hippocampal formation. No single member contains need, map, value, and action. Their interaction turns information gathered earlier into regulation that can begin before a deficit becomes an emergency.
A predictive map gives allostasis a geography.
43.10 The diary on the map
A place code answers where the animal is. An event code distinguishes what happened here on this occasion. Human episodic memory combines these demands at extraordinary scale. It binds people, objects, actions, locations, temporal order, bodily state, and meaning into events that can later be reconstructed from partial cues.
Reconstruction is the important word. Episodic memory is not a video file replayed from one cortical address. Retrieval reinstates parts of a distributed pattern and fills gaps with semantic knowledge, schemas, current goals, and inference. This is why a remembered event can remain recognizable while some details change. It is also why two people can preserve the gist of a shared event while disagreeing about who said a particular sentence.
Constructive memory is not evidence that accuracy was biologically irrelevant. A cache system that routinely returned the wrong place would be disastrous. The controller needs enough fidelity to distinguish relevant events. But it also benefits from compression, generalization, and recombination. The machinery must preserve what makes one episode distinct while discovering structure shared with others.
That flexibility allows memory to participate in constructing possible events. Donna Rose Addis and colleagues showed that remembering past episodes and imagining future episodes recruit a broad overlapping network including hippocampal formation, medial prefrontal cortex, posterior cingulate and retrosplenial regions, and lateral parietal cortex [@addis2007remembering]. The overlap is consistent with shared processes for event construction from distributed elements; functional imaging alone does not show that the same computation is necessary in both tasks.
Hippocampal damage can expose one contribution to that construction. Demis Hassabis and colleagues asked amnesic patients to imagine commonplace scenes such as a beach or museum. Their descriptions were fragmented and lacked the spatial coherence that normally binds elements into one scene [@hassabis2007imagine]. The result fits the architecture developed here: a system that relates elements within a spatial and contextual structure contributes both to recollecting an event and to assembling a novel one.
The contribution is not required in the same way for every form of future thought. People with developmental amnesia and substantial bilateral hippocampal volume loss have sometimes produced rich fictitious and future scenes, probably reflecting developmental compensation, residual tissue, semantic knowledge, or task differences [@maguire2010developmental]. Abstract predictions, scripts, and verbal counterfactuals can also be generated without constructing a vivid scene. The secure conclusion is therefore more specific than “the hippocampus imagines the future.” It helps construct relationally and spatially coherent events within a larger network.
The chickadee barcodes now return in a new light. A sparse event index can distinguish one episode from another. A relational map can locate the event within spatial, temporal, and contextual structure. Distributed cortex supplies perceptual and semantic content. Frontal systems determine what question is being asked of the memory. Retrieval is the coordinated reconstruction of the event, not the opening of one stored file.
The title The Diary on the Map should therefore not imply that autobiography was pasted onto a completed navigation module. It names an evolutionary and computational continuity. Particular events are remembered through machinery that also relates states, contexts, and transitions. The diary is written in a system that never stopped being a map, and the map was never only geography.
43.11 What damage and time reveal
Henry Molaison enters here rather than at the opening because his syndrome is easier to understand once the system has a history and an architecture.
In 1953, neurosurgeon William Scoville removed tissue from both medial temporal lobes in an attempt to treat Molaison’s severe epilepsy. The operation damaged anterior hippocampal formation and neighboring medial-temporal structures, including most of the entorhinal cortex and much of the amygdala. Later imaging and postmortem reconstruction showed residual posterior hippocampal tissue and substantial sparing of posterior parahippocampal cortex. This was a broad lesion, not a selective removal of the hippocampus [@scovillemilner1957; @annese2014hm].
The behavioral result was nevertheless transformative. Molaison could converse, hold information briefly, retain much older knowledge, and acquire some new skills. What he could not do normally was form enduring new memories for the events and facts of his life. Each encounter with a person could become new again after attention moved elsewhere. His case established that bilateral medial-temporal damage can profoundly disrupt new declarative and episodic learning without abolishing perception, intelligence, immediate memory, or every kind of learning.
The spared abilities were as important as the deficit. Amnesic patients can improve on tasks such as mirror reading while lacking ordinary conscious memory of the training episodes [@cohensquire1980skill]. Memory taxonomy is therefore not an arbitrary filing cabinet. The nervous system genuinely separates forms of plasticity and retrieval. The error is to treat one taxonomy as a complete explanation of the circuits.
Time introduces a second problem. New memories depend strongly on hippocampal–cortical interaction, yet old knowledge can sometimes survive severe medial-temporal damage. One account, standard systems consolidation, proposes that hippocampal activity helps establish and reorganize distributed cortical representations until some memories can later be retrieved with much less hippocampal involvement. Replay during rest and sleep provides a plausible mechanism for repeated hippocampal–cortical coordination, and experiments can identify cortical engram populations that mature as memories age [@kitamura2017engrams].
The phrase “the hippocampus trains cortex and then bows out” is too simple. What remains after time depends on what is being tested. A patient with profound medial-temporal damage could reconstruct the broad layout of a neighborhood learned many decades earlier [@tengsquire1999places]. A former London taxi driver with bilateral hippocampal lesions retained much of the city’s major-road framework but failed when navigation required detailed, flexible local relations [@maguire2006taxi]. Functional imaging can also distinguish specific remote autobiographical memories in hippocampal subfields [@bonnici2013remote].
These findings motivate a transformation account rather than a uniform transfer [@winocurmoscovitch2011transformation]. Over time, memories can become more schematic, semantic, or cortex-supported, while vivid contextual reconstruction may continue to recruit hippocampal operations. A familiar fact, a generalized route, and the detailed re-experiencing of one journey are not the same retrieval demand. Treating spared remote knowledge as proof of spared remote episodes obscures precisely the distinction we need to explain. Chapter 47 will examine that error and the influence it had on models of consolidation.
Damage and consolidation therefore reveal the same principle. The hippocampal formation is especially important when behavior requires rapid binding of relations, differentiation of overlapping events, and reconstruction of a context from partial cues. Other systems can slowly acquire regularities, habits, facts, and schemas. Memory changes as those systems interact across experience and time.
43.12 The four problems ahead
The unit is organized around four problems, each approached through comparative anatomy, physiology, behavior, and human neuropsychology.
Chapter 44 asks how a hippocampal system can be recognized across vertebrate brains whose adult pallia look radically different. Development, gene expression, cell types, connectivity, and lesion effects will be used together. The chapter will also test the simple navigation-first story against evidence that fish hippocampal pallium contributes to temporal association and relational inference.
Chapter 45 asks how movement and sensation become neural structure. Place, grid, direction, boundary, speed, and event-related signals will appear as interacting population codes rather than a set of isolated labeled neurons. The chapter will examine path integration, landmark correction, remapping, sequence activity, and the relation between spatial geometry and learned transitions.
Chapter 46 asks when map-like operations extend beyond physical space. Time, continuous sensory dimensions, task states, concepts, and relational inference provide test cases. The central question is not whether cognition can be described with spatial metaphors, but whether the underlying representation preserves relations that support flexible traversal and novel inference.
Chapter 47 begins with exploration, bodily need, and the particular encounters that keep a learned map useful. It then turns to Clive Wearing, H.M., and other patients to distinguish personal episodes from retained knowledge. Its critical argument concerns what the evidence actually established about remote memory, what consolidation models assumed, and how event reconstruction can remain useful as cortical knowledge changes.
These chapters are not an evolutionary staircase from fish navigation to human autobiography. Nor are they serial processing stages. They are four views of a recurrent hippocampal–cortical system: its deep history, its spatial codes, its extension into relational domains, and its contribution to particular remembered and imagined events.
43.13 Coda: learned worlds, inherited worlds
Unit VI asked how something absent can govern present action. This unit supplies part of the answer. Experience changes the structured world available to control. A remembered route, inferred relation, or reconstructed event can become a state through which the animal acts although the relevant place or consequence is not currently visible.
The predictive map is neither an archive nor an oracle. It is a learned model of relations. Its predictions are constrained by where the animal has gone, what it has sampled, which transitions it has experienced, and how the world has changed. Exploration enlarges that model. Need and value reweight it. Frontal systems use it to compare alternatives. Action tests it against reality and supplies new evidence.
A solitary animal must acquire much of this structure by moving through the world itself. A social animal can also learn from where another animal goes. A human can inherit routes, categories, warnings, histories, and imagined destinations from people who died before that individual was born. Language and artifacts allow one person’s experience to change another person’s map without requiring the second person to repeat the first person’s exploration.
That is the bridge to Chapter 48. The next unit turns from the learned world within one brain to worlds built, altered, and remembered by many brains. A predictive map allows the past to reorganize one animal’s future. Culture allows the past of other animals to enter the map.
Established findings. Hippocampal and related medial-pallial systems contribute to flexible spatial and relational behavior across vertebrates. Hippocampal–entorhinal populations represent location, direction, boundaries, temporal position, task variables, particular events, and behavioral sequences. Brief population sequences can represent paths that are not currently being traveled. Bilateral medial-temporal damage can profoundly impair new episodic and declarative learning while sparing immediate memory and several forms of skill learning. Remembering past events and constructing possible events recruit overlapping distributed networks.
The organizing interpretation of this unit. Spatial navigation makes visible a more general capacity to learn states, contexts, events, and transitions. Those learned relations can support retrieval, inference, route generation, and construction of situations not supplied by current sensory input. Human episodic memory elaborates this architecture; it does not sit outside its evolutionary history.
The proposal developed across this book. Exploration is an allostatic investment in future controllability. By learning the structure of the world before a particular need becomes urgent, an organism enlarges the set of reliable actions available when bodily state or environmental conditions later change.
Open questions. We do not yet know whether spatial navigation preceded every nonspatial relational function, whether successor-like coding provides the best general account of hippocampal representation, how event barcodes contribute causally to prospective cache selection, which internally generated sequences directly guide a particular decision, or when richly detailed remote memories remain dependent on hippocampal reconstruction. These questions constrain the synthesis without weakening the architecture on which it rests.