47  The Diary on the Map

How Particular Events Keep a Predictive Map Current

An animal exploring while no need is urgent appears, from a narrow view, to be wasting time. It leaves a safe location, spends energy, and exposes itself to danger without immediately eating, drinking, mating, or escaping. Tolman’s rats did something like this in mazes. Animals allowed to traverse a maze before food was introduced later improved abruptly when food gave one destination value. The reward did not draw the map. It revealed what the animal had already learned [@tolmanhonzik1930reward; @tolman1948maps].

That distinction between learning the world and valuing one possibility now is the beginning of this chapter. Exploration reveals routes, barriers, refuges, resources, risks, and uncertain transitions before a particular deficit makes any one of them urgent. An animal may encounter water while hydrated, shelter while safe, food while sated, or an escape route before a predator appears. The information is acquired under one bodily condition and used under another.

This is why exploration is an allostatic investment. Homeostatic regulation cannot always wait for a deviation to become severe and then begin searching blindly for the means to correct it. Allostatic control reduces that delay by preparing for likely future demands [@sterling2012allostasis]. A body that has explored already knows where correction may be possible. It has purchased options in advance.

But a general map is not enough. Worlds change. The spring discovered last month may now be dry. The direct path may be blocked. A cache may have spoiled. A familiar shelter may have acquired a predator. A social partner who was reliable in one context may not be reliable in another. The controller therefore needs not only a structure of possible states and transitions, but a record of particular encounters that confirms, qualifies, or contradicts that structure.

A predictive map gives allostasis a geography. The diary keeps that geography current.

This chapter develops that claim from animal exploration to human episodic memory. It also returns, deliberately and critically, to H.M. and to Complementary Learning Systems. Those accounts did not merely add a few imperfect details to an otherwise settled story. The canonical interpretation of H.M.’s remote memory became a load-bearing premise for a theory in which old memories leave the hippocampus. When later work showed that H.M.’s remote autobiographical recollection was not intact in the way that premise required, the correction should have altered the center of the field. Instead, generations of students continued to learn the cleaner story.

That history exposes a recurring problem in cognitive neuroscience. A psychological taxonomy, a famous but anatomically broad lesion, and an elegant computational model can reinforce one another until the package is mistaken for a mechanism. The alternative route taken in this unit is to begin with the older animal, the older circuit, and the control problem. From that direction, the human diary is not a file cabinet added to a navigational organ. It is the event-specific evidence by which a predictive map remains useful to a regulated body in a changing world.

47.1 What exploration purchases

Tolman’s cognitive map was never simply a claim that rats know geometric layouts. It was a challenge to the idea that learning consists only of reinforced stimulus-response chains. A rat may learn relations among paths and destinations before a reward makes one path worth taking. When bodily state or outcome value changes, the learned structure can support a different route without being relearned from the beginning [@tolman1948maps]. O’Keefe and Nadel later placed the hippocampus at the center of this capacity: an animal can represent its location within an environment and use relations among places to reach a goal that is not currently sensed [@okeefenadel1978map].

The allostatic consequence is easy to miss if navigation is described only as getting from one point to another. Exploration enlarges the animal’s future control space. It discovers not just where things are, but what the environment permits:

• this opening can be crossed;
• this route leads around a barrier;
• this hollow protects the body from weather;
• this tree contains food at one season but not another;
• this location provides water after rain;
• this path exposes the animal to attack;
• this individual can be approached under some conditions and avoided under others.

James Gibson called such possibilities affordances [@gibson1979ecological]. An affordance is not merely an objective property of an object. Nor is it a reward stored inside the object. It is a relation between an organism and a situation. A gap affords passage only for a body that can fit through it. A branch affords perching only for an animal capable of grasping it. A spring affords drinking whether or not the animal is presently thirsty, but thirst changes the value of that possibility and the vigor with which the animal pursues it.

This makes it useful to separate several functions that ordinary language tends to collapse:

Component	Question answered
Map	What states, contexts, routes, and transitions exist?
Affordance	What can this body do in this situation?
Diary	What happened when that possibility was tested on one occasion?
Value	How much does the possible outcome matter under the current bodily state?
Policy	Which action should be selected now?

To say that exploration stores affordances in a cognitive map does not require a neural token labeled affordance. The map must preserve enough of the organism-situation relation to recover the possibility later: where the opportunity was encountered, which approach made it available, which barriers constrained it, what action was possible, and what outcome followed. Other systems supply the present need and assign current value. The map is not hunger, thirst, fear, or choice. It is the learned structure within which those states can select a destination and a route.

47.2 A map is not enough: it needs a diary

A map captures what generally holds. A diary preserves what happened this time.

Suppose an animal has learned that a spring lies beyond a ridge. That relation may remain useful across many visits. But control in a changing environment requires more than the timeless proposition water is there. The last visit may have occurred after heavy rain; the spring may be seasonal. A recent event may show that the direct route has collapsed. Drinking there may have been followed by an attack. The relevant memory is not merely the destination but a conjunction:

Under this bodily state, in this context, by this route, this action produced this outcome.

That conjunction must remain distinguishable from other encounters sharing the same place, objects, and actions. The same spring can be full in April and dry in August. The same shelter can be safe at noon and occupied at dusk. The same individual can cooperate during one interaction and defect during another. If the controller averaged these experiences immediately into one timeless relation, it would lose the very evidence needed to detect change.

A map without a diary predicts a stationary world.

The diary is therefore not a private video archive. It is a record of particular state-transition-outcome relations. It preserves the exceptions, failures, recent changes, and conditional outcomes that a general map would otherwise smooth away.

This operation did not begin with human autobiography. Scrub jays remember what they cached, where they cached it, and how long ago. They preferentially recover perishable food before it degrades and shift to more durable caches after the relevant interval [@claytondickinson1998episodic]. The experiment does not establish that a jay narrates the past to itself or experiences human autonoetic consciousness. It shows something older and more concrete: a particular resource encounter remains bound to place and elapsed time because those details determine whether returning is still worthwhile.

Rats likewise can combine information about what occurred, where it occurred, and its position within an event sequence. Hippocampal damage can spare performance based on individual elements while disrupting their combined use to distinguish one encounter from another [@ergoruleichenbaum2004whatwherewhen]. The biological problem is already present before there is language, a life story, or a word for memory:

One occurrence must remain distinguishable from other occurrences that share most of its parts.

The general map says that a resource can be found there. The diary says whether it was found there recently, under what conditions, and at what cost. Repeated episodes can eventually support a stable regularity. A surprising episode can overturn it. The diary supplies the dated evidence from which the map is learned and by which it is corrected.

47.3 The same world under hunger and thirst

Pamela Kennedy and Matthew Shapiro devised an experiment that joins the hippocampus, internal bodily state, memory retrieval, and goal selection more directly than almost any familiar human memory task.

Rats were trained in one unchanged external environment. Three goal boxes were visually distinct. One contained food, one contained water, and one was never rewarded. The boxes moved among locations so that the task could not be solved by approaching a fixed place. On hungry days, the rat had to select the food-associated object. On thirsty days, it had to select the water-associated object. The visible world was the same. What changed was the animal’s internal context [@kennedyshapiro2004internal].

The task required more than detecting hunger or thirst. It required the animal to use that state to retrieve a relation:

Control component	Kennedy–Shapiro task
Internal context	Hunger or thirst
External context	The same testing environment
Learned alternatives	Distinct food, water, and unrewarded goal objects
Current value	Food while hungry; water while thirsty
Memory demand	Retrieve which object predicts the currently needed outcome
Behavior	Approach the appropriate object despite its changing location

Selective hippocampal lesions or fornix transections reduced performance toward chance. Yet the rats still discriminated the visible objects, avoided the never-rewarded box, distinguished their deprivation states, and preferred the commodity they currently lacked. The lesion did not abolish hunger, thirst, perception, or simple object-reward learning. It impaired the flexible use of internal state to retrieve the appropriate relation among context, object, action, and outcome [@kennedyshapiro2004internal].

That distinction matters. The hippocampus was not the homeostatic controller. It did not create the deficit. It was required to let the current bodily state query a learned relational structure.

Kennedy and Shapiro later recorded hippocampal neurons while rats performed related tasks under hunger and thirst. Motivational state reorganized hippocampal activity most strongly when the internal state was needed to select among remembered alternatives in the same places. The same hunger and thirst states had much less influence during random foraging, when they were present but did not determine which memory should guide behavior. Prospective activity reflected the deprivation-defined target rather than simply the spatial trajectory [@kennedyshapiro2009motivation].

This is not a vague claim that motivation modulates memory. It is a division of labor within a controller:

• the body signals a regulatory condition;
• the learned map contains several possible outcomes;
• the internal condition selects which relation must be retrieved;
• valuation and action-selection systems choose among the retrieved possibilities;
• the resulting encounter supplies new evidence for the next decision.

The map can be acquired under one bodily state and read under another.

That is the allostatic advantage of exploration. An animal does not need to be thirsty while learning every route to water. It needs a system capable of preserving the relation and allowing thirst to make that relation matter later.

47.4 Human memory arrives last

We have arrived, at last, at ourselves.

For most of this unit we refused to begin with human amnesia, and the refusal was the argument. A medial-pallial system existed hundreds of millions of years before birthdays, autobiographies, and verbal reports. Its basal problem was not to maintain a personal archive. It was to organize states, contexts, and transitions so that past encounters could guide a body beyond the present cue.

Human episodic memory elaborates that operation, but it does not merely run the same machine for a longer distance. Human events are saturated with additions supplied by enlarged and interacting systems: language, self-knowledge, social roles, causal explanation, culturally organized calendars, narrative structure, and the awareness that this happened to me. A scrub jay recovering a perishable cache and a person recalling a childhood meal share event-binding problems. They do not possess identical memory systems distinguished only by timescale.

The continuity lies at the level of operation. A particular event binds what happened, where, when, in what order, under what bodily state, through which action, and with what result. The human form can place that event within a life story and describe it to someone else. But the biological reason for preserving particulars remains visible in the simpler animal: a dated encounter changes what the organism should expect and do next.

Human neuropsychology is therefore most informative when it arrives after the function has been defined. The famous patients reveal what disappears when a distributed event-organizing system is damaged. They do not, by themselves, tell us why the system evolved.

47.5 The man in the present tense

In 1985, the musician and musicologist Clive Wearing developed herpes simplex encephalitis. The infection damaged broad regions of medial and anterior temporal cortex on both sides. His lesion was not a selective removal of the hippocampus, and his impairment was not confined neatly to one textbook drawer. Formal testing found exceptionally severe episodic amnesia together with major semantic impairment [@wilsonetal1995wearing].

The detail that makes the case unforgettable is his diary. Wearing repeatedly wrote that he had only just become fully conscious, then crossed out the preceding entry as mistaken. The entries do not form a sequence for him. Each new moment appears to cancel the one before it because the earlier moment cannot be integrated into an event that remains available.

That diary should not be treated as a transparent view of what isolated hippocampal loss feels like. Wearing’s damage was broad. The investigators who described him interpreted his persistent belief that he had just awakened as a delusion rather than as a necessary consequence of amnesia [@wilsonetal1995wearing]. His preserved musical performance should not be reduced to one untouched box called procedural memory. Skilled performance draws on massively practiced perceptual, motor, affective, and semantic organization distributed across the brain.

Even after those corrections, the case remains devastating. Wearing can inhabit an ongoing act while its cues and structure sustain him. He can conduct music while the score, performers, and unfolding sequence support the performance. What he cannot normally do is bind the changing stream into a durable event and later use that event to locate the present within an extended personal history.

The usual metaphor is a failed recorder: experience occurs, but the tape does not retain it. The diary suggests a more useful description. The present cannot be placed within a retrievable trajectory. Earlier moments do not remain available as context for the current one. The result is not simply an empty archive. It is the repeated collapse of temporal continuity.

Wearing’s diary shows what happens when successive moments cannot be held together as one continuing event.

That is enough. We do not need to claim that loss of a future caused loss of the past, or that the diary proves one theory of hippocampal function. The case introduces the human problem in its starkest form: without durable event binding, the present repeatedly loses the context that would make it part of a life.

47.6 The cabinet and the cost of mistaking labels for mechanisms

The taxonomy of memory is useful. It is also where cognitive neuroscience can go wrong.

Long-term memory is conventionally divided into declarative and nondeclarative forms. Declarative memory includes flexible memory for facts and events that in humans can support recognition, conscious recollection, and report. Verbal declaration is an important expression of that memory, not its biological definition. Animals and nonverbal patients can demonstrate flexible relational knowledge without putting it into words.

Within declarative memory, Tulving distinguished episodic memory for particular events from semantic knowledge that is not tied to re-experiencing one occasion [@tulving1972episodic]. Outside this division lie several forms of learning expressed in performance: skills, habits, priming, conditioning, and other changes in behavior. These do not constitute one mechanism merely because they are classified as nondeclarative.

The categories describe real dissociations. A patient may improve at a task without recollecting the training episodes. Another may retain facts while losing the ability to re-enter particular personal events. Such contrasts establish that performance can come apart. They do not establish that the brain contains a separate anatomical cabinet for every psychological label.

This is the recurrent temptation. Begin with a noun supplied by psychology—working memory, episodic memory, semantic memory, attention, inhibition. Treat the noun as a natural object. Search for the brain region that contains it. A behavioral distinction becomes an anatomical claim before the neural operation has been identified.

Lesions can intensify the error. A biologically uncontrolled injury is described using one psychological contrast. The contrast is then assigned to the damaged structure as though the lesion respected the boundary of the construct. A computational model can complete the circle by giving the construct a formal mechanism. The case validates the model; the model makes the case look selective; the taxonomy gives both a familiar name.

The result can be a coherent story that is wrong at its center.

Memory taxonomy tells us what forms of performance can dissociate. It does not explain why an ancient medial-pallial system binds events, how bodily state queries those events, how particular encounters update a map, or why detailed recollection might remain dependent on hippocampal operations. Those questions require beginning with the circuit and the control problem, not ending the inquiry when a box has been named.

H.M. is the clearest example of both the power of a lesion case and the danger of turning its first interpretation into an architecture.

47.7 H.M. and the remote memory that became a fact

In 1953, William Scoville performed a bilateral medial-temporal operation on Henry Molaison in an attempt to reduce severe epilepsy. The operation removed medial temporal polar cortex, most of the amygdala and entorhinal cortex, and anterior portions of the hippocampal formation. A substantial posterior segment of the hippocampus remained on each side, but it was largely disconnected from normal cortical input and output. Later examination also revealed pathology beyond the intended medial-temporal lesion [@scovillemilner1957; @anneseetal2014hm]. His seizures were reduced, not abolished, and he continued to receive anticonvulsant medication.

The operation produced a profound and enduring inability to form ordinary durable memories for new events and facts. H.M. could perceive, speak, reason within the available moment, and retain information while actively rehearsing it. Once attention moved elsewhere, the newly encountered material was usually lost. He also improved across repeated training on several perceptual and motor tasks despite lacking recollection of the training episodes [@milneretal1968hm; @cohensquire1980skill].

Those findings changed neuroscience. They established that the ongoing stream of experience does not become durable declarative memory automatically, that medial-temporal structures are critical to that process, and that learning can alter performance without producing conscious recollection of the learning episode.

None of those conclusions requires the textbook claim that caused the trouble.

The trouble concerned H.M.’s past. Early descriptions reported that his remote memories were preserved while memories from the years nearest the operation were impaired [@scovillemilner1957; @milneretal1968hm]. The finding was transformed into a much stronger proposition: old episodic memories had survived because they no longer required the hippocampus. The remote past had moved to cortex.

That proposition became standard before the relevant distinction had been measured adequately. Knowing facts about one’s childhood is not the same as re-experiencing one childhood event. Repeating a familiar anecdote is not the same as reconstructing a particular scene with its people, place, temporal order, sensations, actions, and contingencies. A generalized biography can survive after its episodes have been lost.

When H.M.’s remote autobiographical memory was later tested with methods designed to distinguish personal semantic knowledge from specific episodic re-experiencing, the canonical result did not hold. H.M. and another medial-temporal patient showed relatively preserved semantic knowledge but severe autobiographical impairment without a temporal gradient [@steinvorthetal2005remote]. The postmortem report, coauthored by Suzanne Corkin, states the same distinction plainly: H.M. retained semantic knowledge from before the operation but could not retrieve episodic autobiographical memories from that period [@anneseetal2014hm]. Corkin’s later book also makes clear how sparse, repetitive, and generalized many of the stories treated as remote memory had been [@corkin2013permanent].

The correction is precise. H.M. did not have no past. He retained language, facts, habits, skills, and a generalized biography. What he did not demonstrate was the intact, richly detailed remote episodic record needed to show that old episodes had become hippocampus-independent.

The image of H.M. with an intact remote autobiography was not merely incomplete. For the kind of memory used to support standard consolidation, it was probably wrong.

This matters because the claim was not ornamental. Preserved remote episodic memory in the absence of a functioning hippocampal system was the clinical fact that made a temporary hippocampal role appear necessary. If that fact collapses into preserved semantic knowledge and rehearsed personal facts, the mechanism no longer follows.

H.M. did not misdirect the field. The use made of H.M. did.

For decades, textbooks taught a clean sequence: the hippocampus stores new memories; time passes; the memories are transferred to cortex; old memories therefore survive hippocampal damage. The case and the theory became mutually reinforcing. H.M. seemed to prove the transfer. The transfer explained H.M. The distinction between an episode and the semantic residue of an episode disappeared inside the loop.

This was not a small correction delayed by the ordinary pace of science. It shaped what questions appeared worth asking. The dominant question became How long does a memory remain in the hippocampus before it leaves? The cognitive-map tradition asked a different question: What operation remains necessary whenever a detailed event is reconstructed? The field treated the second question as an exception to a story that the first question had already assumed.

47.8 An elegant model built on a brittle premise

Complementary Learning Systems gave the textbook sequence a computational reason.

McClelland, McNaughton, and O’Reilly proposed that the brain requires two learning systems with opposing properties [@mcclellandetal1995cls]. A hippocampal system learns rapidly, preserving the conjunctions of one event. Neocortex changes slowly across many experiences, extracting distributed regularities without allowing each new example to overwrite what was learned before. Hippocampal replay interleaves recent events with older knowledge so that cortical learning can proceed gradually.

The fast-slow distinction is powerful. It captures a real problem: a system optimized to absorb one event immediately is not necessarily the system best suited to extract stable structure across thousands of events. But the original CLS account did not enter the literature as a neutral proposal about two learning rates. It began from the clinical premise that hippocampal damage disrupts recent memory while leaving remote memory intact, and it explained that pattern by making cortical independence the destination of consolidation [@mcclellandetal1995cls].

The premise and the model locked together. H.M.’s apparently spared remote memory gave the model biological authority. The model, in turn, made the clinical story seem mechanistically complete.

NoteDeeper dive: catastrophic interference is real; the historical conclusion does not follow

Connectionist networks store many relations in shared weights. If one set of associations is learned and the same weights are then changed rapidly to learn another, the second training can severely damage the first. McCloskey and Cohen called this catastrophic interference [@mccloskeycohen1989catastrophic].

Interleaving old and new examples reduces the problem. The network does not learn one block to completion and then overwrite it with the next. It alternates among them, allowing weight changes to settle on structure that accommodates several experiences.

CLS assigned those demands to different systems. The hippocampus rapidly captured individual events. Offline replay supplied recent events to a slower neocortical learner, mixed with older structure. The division explains how one-shot learning and gradual generalization can coexist [@mcclellandetal1995cls].

The computational problem is genuine. The leap from that problem to the historical and anatomical claim is not. A model showing that one architecture avoids interference does not establish that the hippocampus evolved to buffer cortex, that a detailed episodic memory is initially stored only in hippocampus, or that the episode later survives intact after hippocampal participation ends.

A solution in a model is not a history of a brain.

This is where the model became a misdirection. It encouraged the field to treat three claims as one:

1. rapid learning of particulars and slow learning of regularities are computationally complementary;
2. hippocampal-cortical interaction changes with experience and time;
3. detailed memories are eventually transferred out of the hippocampus and become independent of it.

The first claim remains compelling. The second is supported broadly. The third does not follow from either of them, and H.M.’s remote memory did not provide the evidence it was said to provide.

Replay is not proof of transfer. Cortical learning is not proof that the hippocampal representation has become dispensable. A remote memory that can be answered as a fact is not proof that its episodic form survived in cortex. Even the word consolidation can obscure the issue by making stabilization, network reorganization, and representational transformation sound like one process with one endpoint.

The model’s elegance made the misdirection durable. A weak clinical premise acquired computational necessity. Once the field had a formal reason why old memories must leave the hippocampus, contrary patients could be absorbed by extending the consolidation interval, questioning lesion selectivity, or redefining what counted as remote. The theory became difficult to disconfirm precisely where its central claim should have been most vulnerable.

Modern versions of CLS have become more flexible and have incorporated schemas, rapid cortical learning under some conditions, and continuing hippocampal-cortical interaction. Those developments may improve the framework. They do not retroactively validate the story generations of students were taught: that H.M.’s intact remote episodes proved that episodic memories are gradually handed from hippocampus to cortex.

As a theory of remote episodic memory, that story sent the field down the wrong road. It shifted attention from the operation performed during detailed reconstruction to the imagined location of a memory file at successive ages.

47.9 The alternative the field treated as an exception

The strongest alternative came from the cognitive-map lineage itself.

Lynn Nadel and Morris Moscovitch proposed that richly detailed episodic recollection continues to require the hippocampal complex no matter how old the event is [@nadelmoscovitch1997consolidation]. Each retrieval can establish additional traces and relationships, but time does not turn the original episode into an identical hippocampus-free cortical copy. What can become increasingly independent is a different representation: semantic knowledge, gist, schema, and a generalized account of what usually happened.

Multiple Trace Theory, and the later Trace Transformation account, therefore changed the question. The issue was not simply the age of the memory. It was the kind of information being demanded at retrieval [@moscovitchetal2005remote; @winocurmoscovitch2011transformation].

• A familiar fact about one’s childhood can be supported by distributed cortical knowledge.
• A repeated anecdote can survive as a verbal script.
• A generalized route through a neighborhood can remain available.
• Re-entering one particular occasion with its spatial, temporal, perceptual, and relational detail continues to call on hippocampal operations.

This was not an ad hoc rescue after H.M. became inconvenient. It made a risky prediction: remote episodic memory should remain vulnerable if it is tested as episodic memory rather than as personal fact. Later formal testing of H.M. produced exactly that pattern [@steinvorthetal2005remote].

The model is often summarized by saying that the hippocampus is “always necessary.” That phrase needs its object. The claim is not that every remote fact, habit, word, route, or feeling forever requires the hippocampus. It is that richly detailed, contextually situated episodic recollection continues to depend on hippocampal contribution for as long as it remains richly detailed and episodic [@moscovitchetal2005remote].

That distinction fits the map and diary developed here. Distributed cortex can learn what repeatedly holds across encounters: the stable layout, the familiar script, the general identity of a person, the usual relation between a cue and an outcome. The diary preserves what happened on one occasion and supports reconstruction of that conjunction later. Time can strengthen general knowledge without converting a particular event into an identical cortical file.

The uncomfortable lesson is not that one theory won a polite contest. The field elevated an interpretation of H.M. that directly favored a temporary hippocampal role, then treated a model predicting permanent involvement in detailed episodic recollection as the dissident view. When H.M. was finally tested in a way that separated episodic re-experiencing from semantic personal knowledge, the dissident model fit the result better.

The correction should be taught as a correction, not as an optional nuance appended to the old diagram.

47.10 A map without its particulars

The patient K.C. makes the distinction visible in another form.

After a motorcycle accident, K.C. had widespread damage that included large bilateral hippocampal lesions. His anterograde amnesia was severe. Across his earlier life, he retained substantial semantic knowledge about the world and about himself but could not recollect personally experienced events [@rosenbaumetal2005kc].

His spatial knowledge was especially revealing. K.C. could answer many questions about the general layout of a familiar neighborhood, judge major directions, and navigate using broad relations. Yet he lost many less salient landmarks and fine details of places he had known [@rosenbaumetal2000spatial]. The broad structure survived better than the particular encounters through which it had been learned.

K.C. retained much of the map and lost the diary entries.

His lesion was broad, so the case does not supply a clean hippocampus-only test of every remote-memory theory. That limitation should not be turned into the opposite error of treating lesion evidence as intrinsically weak. A lesion provides causal information that functional imaging cannot. Its constraint is anatomical: naturally occurring damage rarely respects the borders of a theoretical system.

K.C.’s importance is conceptual. Knowing the layout of a neighborhood is not the same as remembering one walk through it. Knowing that a family trip occurred is not the same as re-entering one scene from the trip. General relations can remain useful after the particular occasions that established them have disappeared.

This is precisely the difference between a map and its diary.

47.11 What memories become with time

The map and diary are not independent stores. They learn from one another.

Each particular encounter supplies evidence. Across repeated encounters, distributed cortical systems can extract stable regularities: this route usually remains open; this kind of place contains food; this person usually responds in a particular way. Those regularities make control efficient because the animal need not reconstruct every decision from the full history of individual episodes.

But efficiency creates a risk. Averages can conceal change. A general schema that a spring contains water is useful until one recent visit shows that the spring has dried. The controller must preserve enough event-specific evidence to override the average when conditions change.

The division of labor is therefore not simply fast storage followed by slow transfer. It is a continuing exchange between particulars and regularities:

• rapid hippocampal-cortical binding keeps overlapping encounters distinct;
• distributed cortical learning integrates what recurs across encounters;
• schemas guide what is noticed, encoded, and reconstructed in later events;
• surprising events revise the schema;
• recent exceptions can govern behavior before the long-run average changes.

Animal work on systems reorganization also undermines the image of an initially empty cortex waiting for a completed memory to arrive. In a contextual fear paradigm, prefrontal ensembles associated with the memory were established during initial learning and matured over time through interaction with hippocampal and other circuits [@kitamuraetal2017engrams]. That experiment does not settle human remote autobiographical memory. It does show that network reorganization is not well described as an intact packet first stored in hippocampus and later shipped to cortex.

Several processes commonly gathered under consolidation should therefore be separated:

1. Synaptic stabilization: molecular and cellular changes that make a newly altered circuit less fragile.
2. Systems reorganization: changing interaction among hippocampal, cortical, and subcortical populations over time.
3. Representational transformation: loss of some event-specific detail and increased reliance on schema, gist, or semantic structure.

These processes can occur together, but they do not have the same endpoint. Stabilization does not imply hippocampal independence. Cortical participation does not imply transfer. A schematic remote memory is not the same representation as the event from which it was abstracted.

Cortex preserves what repeatedly holds. The diary preserves what happened this time.

That sentence is deliberately schematic. Detailed memory is distributed, and cortex participates from the beginning. The hippocampus does not contain a complete episode by itself. Its contribution is to bind and later reconstruct the relations that make one event this event rather than another. The essential contrast is not one anatomical box versus another. It is slow integration of regularities versus preservation of particulars and exceptions within an interacting system.

47.12 Remembering one event and constructing another

Remembering a particular event and constructing a possible event recruit overlapping systems. In functional imaging, autobiographical recollection and imagination of future events engage a distributed network that includes the hippocampus, medial prefrontal cortex, posterior cingulate and retrosplenial regions, lateral temporal cortex, and other association areas [@addisetal2007remembering]. The overlap is expected if both tasks require elements to be bound into a coherent situation not currently present.

Hassabis and colleagues asked patients with hippocampal amnesia to imagine novel scenes. Most produced fragmented descriptions lacking the spatial coherence of control participants [@hassabisetal2007imagine]. The result is important because it moves beyond retrieving a stored event. The hippocampal system contributed to constructing a particular relational scene.

It should not be inflated into the claim that the hippocampus is required for every thought about the future. People can predict from semantic knowledge, follow scripts, make verbal plans, estimate trends, and choose among familiar options without constructing a richly situated episode. Developmental-amnesia cases have also shown that some individuals with severe episodic impairment can construct plausible fictitious or future scenes, presumably through residual tissue, semantic scaffolding, developmental compensation, or different task strategies [@maguireetal2010developmental].

The secure conclusion is narrower and more useful:

Hippocampal-cortical interaction contributes when a particular event—remembered or possible—must be assembled as a coherent conjunction of place, time, entities, relations, actions, and outcomes.

This capacity has an obvious allostatic use. Before taking an expensive or dangerous action, the controller can construct a possible encounter: approach the spring by the exposed route, arrive after dark, find the resource absent, or encounter a predator. Event construction allows consequences to be sampled internally before they are paid for by the body.

The constructive character of memory also produces error. Schemas fill gaps. Present goals bias which elements are recovered. Repeated retrieval alters the event. Pieces from different occasions can be combined. Schacter and Addis emphasized that the flexibility used to build possible events is also a source of memory distortion [@schacteraddis2007constructive].

But error is not proof that literal accuracy was never important. A cache-recovery system that routinely confuses locations fails. An animal that mistakes the dry spring for the full one may die. The same architecture can be both adaptive and fallible.

Distortion is a cost of flexible reconstruction, not evidence that fidelity is biologically irrelevant.

The diary need not preserve every detail. It must preserve the distinctions that change prediction and control.

47.13 Development without an ordinary diary

Developmental amnesia provides a different view of the division between particulars and regularities.

Vargha-Khadem and colleagues described three young people whose early injuries produced bilateral hippocampal pathology. They had profound difficulty remembering the episodes of everyday life, yet attended mainstream schools and acquired language, literacy, and substantial factual knowledge in the low-average to average range [@varghakhademetal1997developmental].

The result is often presented as a clean anatomical separation between episodic and semantic memory. It is stronger and more interesting when stated accurately.

There were three intensively studied cases, not a population-scale controlled experiment. The damage was not a surgical isolation of one structure. Residual hippocampal tissue and associated pathology differed. Early injury allowed years of developmental compensation. Repeated exposure, recognition, external scaffolding, and slow cortical learning could support knowledge acquisition even when ordinary one-trial event memory was profoundly impaired.

The cases therefore show that substantial semantic knowledge can develop without normal episodic memory. They do not show two independent filing cabinets. The children learned facts through a slower and more supported route while struggling to retain the daily encounters through which other people ordinarily build those facts.

Their lives make the control problem visible. A child can learn that a bus route reaches school, what multiplication means, and the names of familiar people, yet fail to remember what occurred earlier that day. General structure accumulates. Particular events do not remain available in the usual way. Family members, teachers, schedules, notebooks, and repeated routines become an external diary that compensates for the unreliable internal one.

This is not a minor impairment hidden beneath preserved test performance. The map can become broad enough to support education while remaining dangerously slow to incorporate recent change. A person may know the rule and lose the occasion on which the rule failed.

Developmental amnesia thus supports the chapter’s central distinction without turning it into a set of isolated modules:

General knowledge can be built gradually from repeated structure. Ordinary life depends on retaining what happened on this occasion.

47.14 The diary inside the controller

The hippocampus is not the diary by itself, just as it is not the cognitive map by itself. Events are reconstructed through interaction among hippocampal, cortical, subcortical, interoceptive, valuation, and action systems. The division of labor can be stated concretely:

System contribution	Role in allostatic use of memory
Brainstem, hypothalamic, and interoceptive systems	Signal bodily condition, regulatory error, and anticipated need
Hippocampal–entorhinal and distributed cortical systems	Preserve states, contexts, transitions, and the relations that distinguish particular encounters
Amygdalar and orbitofrontal systems	Contribute affective significance and expected outcome identity
Striatal systems	Learn action values, vigor, and selection tendencies
Frontal systems	Specify current goals, questions, rules, and alternatives to be evaluated
Motor and autonomic systems	Implement the selected response
New sensory and interoceptive evidence	Confirm, qualify, or contradict the remembered relation

No row contains the whole behavior. Thirst does not specify a route. A remembered route does not determine whether water currently matters. Value does not reconstruct the last encounter. A frontal goal does not create the map it queries. The organism acts through recurrent interaction among these systems.

The Kennedy–Shapiro experiments make this architecture unusually clear. The rats remained hungry or thirsty after hippocampal damage. They could discriminate the goal objects and prefer the missing commodity. What failed was the use of internal condition to retrieve the relevant relation in a context where several learned alternatives competed. The body supplied the question. The hippocampal system helped recover the structured answer. Other systems selected and executed the response.

The diary changes the answer over time. A spring once associated with water may acquire a recent dry episode. A refuge may acquire a threat episode. A social partner may acquire a betrayal. The map still contains the destination and route, but the event-specific evidence changes the predicted outcome and therefore the policy.

This is why episodic memory belongs in a unit about predictive maps rather than in a detached chapter on storage. Particular events preserve the conditions under which an affordance succeeded or failed. They allow the map to be revalued by a new bodily state and revised by a changed world.

47.15 Coda: the past changes what can happen next

Return to the animal that found water while it was hydrated.

Exploration established a route to the spring before thirst made the destination urgent. Later dehydration revalued the learned world. The spring became a goal. But the controller did not consult a timeless map alone. It also retrieved what happened on recent visits: water was present after rain; the direct route was blocked yesterday; a predator was encountered there at dusk. Those particulars determined whether the old affordance could still be trusted.

The map preserves possibilities. The diary preserves the evidence that keeps those possibilities honest.

Human episodic memory elaborates this architecture enormously. We bind events to a self, place them within calendars and narratives, describe them to others, and use them to construct remote personal futures. Language and culture add something more consequential than a longer private forecast: one person’s diary can alter another person’s map. Testimony, instruction, written records, photographs, schedules, and public histories allow events to outlive the nervous systems that experienced them. That is where the next unit begins—not because the cognitive map fails at cycles or recursion, but because culture makes maps and diaries shareable.

The famous patients now enter the unit in a different role. Clive Wearing shows the disintegration of continuity when successive moments cannot be retained as events. H.M. shows that medial-temporal damage can devastate the ordinary incorporation of new experience while sparing much perception, established knowledge, immediate retention, and several forms of learning. K.C. shows that general structure can survive after particular lived events have disappeared.

H.M. also leaves a warning. A compelling patient, a useful taxonomy, and an elegant model can become a closed explanatory loop. The field taught preserved remote episodic memory as a fact before it had distinguished episodes from personal semantic knowledge. CLS then supplied a mechanism for that fact, and the mechanism made the fact harder to question. Later evidence did not merely decorate the story with nuance. It removed a premise that had held the story together.

The evolutionary and control-theory approach does not solve this problem by replacing one psychological cabinet with another. It asks what the system contributes to an organism. The oldest map organized states and transitions beyond the present cue. Exploration expanded the animal’s future options. Bodily need later selected among those possibilities. The diary preserved what happened when one possibility was tested, allowing the general map to be confirmed, corrected, or abandoned.

The brain preserves past events because they alter present and future control. That claim does not reduce memory to a forecast, and it does not make accuracy optional. It gives the diary a biological job.

A predictive map gives allostasis a geography. The diary keeps that geography honest.

TipEstablished findings, the chapter’s interpretation, and open questions

Established findings. Exploration can produce relational learning before a currently valued reward reveals that learning. Nonhuman animals preserve event-specific conjunctions involving content, location, sequence, and elapsed time. In rats, hunger and thirst can serve as internal contexts that retrieve different learned goal relations in the same external environment; hippocampal or fornix damage disrupts that flexible retrieval while sparing deprivation discrimination, visible-cue discrimination, and simpler reward preferences. Bilateral medial-temporal damage can profoundly impair rapid formation of durable memories for new events and facts while sparing immediate retention under sustained attention, established knowledge, and several forms of learning. Formal later testing of H.M. found severe impairment of remote autobiographical re-experiencing without a temporal gradient despite relatively preserved semantic knowledge. Detailed remote episodic retrieval, generalized remote knowledge, and broad spatial structure can dissociate.

The chapter’s organizing interpretation. Exploration is an allostatic investment that learns possible resources, refuges, routes, risks, and other affordances before one bodily need becomes urgent. Current internal state revalues that learned structure. Particular event memories preserve the conditions, actions, outcomes, failures, and recent changes that determine whether a generally learned affordance remains useful. Distributed cortical systems integrate recurring structure; hippocampal–cortical interaction preserves and reconstructs the particulars that keep the structure calibrated. The canonical use of H.M. and the original CLS transfer story misdirected the field by treating remote personal facts as intact remote episodes and by turning a computational solution to interference into a biological history of memory.

Open questions. The field has not fully established how interoceptive state is incorporated into hippocampal representations, which features of remote events continue to require hippocampal operations under different retrieval demands, how schemas and particular events update one another, or how developmental compensation supports semantic learning without ordinary episodic memory. Nor has it identified one fixed timescale or one endpoint for systems reorganization. These are questions about the changing division of labor within a distributed controller, not about the location to which a memory file is eventually moved.