In the intellectual landscape of early twentieth-century American psychology, experimental inquiry was characterized by a rigorous adherence to observable phenomena. Dominated by the mechanical tenets of classical behaviorism, the discipline sought to expunge mentalistic constructs from scientific discourse, proposing instead that all animal and human behavior could be reduced to conditioned reflex arcs, associative bonds, and peripheral physiological states. Organisms were cast as passive recipients of environmental stimulation, driven by physiological deficits and shaped exclusively through direct reward or punishment. Within this paradigm, the internal experience of the subject—its capacities for deliberation, structural inference, and mental representation—was relegated to an inaccessible and scientifically illegitimate “black box.”
It was against this rigid orthodoxy that Edward Chace Tolman mounted a profound conceptual revolution. Through a series of elegant rodent maze experiments conducted at the University of California, Berkeley during the 1930s and 1940s, Tolman demonstrated that learning was not merely the mechanical stamping-in of stimulus-response connections via reward, but rather an active, cognitive process of environmental modeling. Tolman proposed that organisms construct internal, dynamic spatial layouts of their surroundings—theoretical constructs he formally christened cognitive maps. Crucially, his discovery of latent learning revealed that animals could acquire rich architectural knowledge of an environment in the total absence of reinforcement, storing this latent information until future motivational states rendered its expression advantageous.
Decades after Tolman formulated his purposive behaviorism, contemporary cognitive neuroscience has validated his visionary hypotheses. The cognitive map is no longer treated as a purely hypothetical construct; it has been identified in the neuroarchitectonics of the mammalian brain—most prominently within the hippocampal formation, the medial entorhinal cortex, and their associated limbic circuits. More remarkably, modern neuroscience has revealed that the very neural systems dedicated to spatial mapping also form the foundational architecture for mental time travel, counterfactual reasoning, and the simulation of the episodic future. The navigational mechanisms that allow a rodent to calculate an unvisited shortcut through a geometric maze are mathematically and functionally homologous to the prospective operations that allow human beings to project themselves into non-existent futures, simulate alternative outcomes, and deliberate across temporal horizons.
1. Historical Foundations of Behavioral Psychology and Tolmanian Divergence
1.1 The Dominance of Classical Stimulus-Response Behaviorism
The dawn of the twentieth century witnessed a radical methodological purge within psychological science, initiated as a conscious reaction against the subjective methodologies of Wilhelm Wundt and Edward Titchener’s structuralist introspectionism. Championed by John B. Watson in his foundational 1913 manifesto, behaviorism sought to transform psychology into a purely objective, experimental branch of natural science. Watson asserted that introspection formed no essential part of psychological methods, nor was the scientific value of its data dependent upon the readiness with which they lend themselves to interpretation in terms of consciousness. Drawing deep methodological inspiration from Ivan Pavlov’s discovery of conditioned salivary reflexes in canines, Watson argued that all complex behaviors, from emotional outbursts to intricate motor sequences, represented nothing more than chains of conditioned peripheral reflexes linked directly to external environmental stimuli.
Simultaneously, Edward L. Thorndike laid the empirical bedrock for operant conditioning through his formulation of the Law of Effect. Observing felines attempting to escape from puzzle boxes, Thorndike posited that responses followed by satisfying states of affairs were stamped into the nervous system, whereas responses accompanied by discomfort were stamped out. This conceptualization was rigorously non-cognitive: the animal did not deduce the mechanical logic of the latch; rather, the successful motor action was mechanically strengthened through hedonic feedback. Building upon these foundations, Clark L. Hull erected an elaborate mathematical formalization of behavior in the 1930s and 1940s known as drive-reduction theory. Hull asserted that an organism was propelled into activity by primary biological drives—such as hunger, thirst, or avoidance of pain—and that learning, conceptualized as “habit strength,” accrued exclusively when a stimulus-response pairing co-occurred with the immediate reduction of a physiological drive.
The epistemological limitations of this radical peripheralism soon became glaringly apparent. By treating the organism as a mechanical automaton governed exclusively by physical contact with stimuli and immediate reinforcement, Hullian and Watsonian behaviorism deliberately ignored unobservable internal states. The paradigm could not adequately explain spontaneous behavioral innovation, the flexible reorganization of movement trajectories in the face of novel barriers, or the acquisition of environmental knowledge occurring when no physiological drive was satisfied. By methodologically conflating overt behavioral performance with the underlying, covert acquisition of information, classical behaviorism erected an artificial barrier against the scientific study of representational systems and mental operations.
1.2 Edward C. Tolman and the Advent of Purposive Behaviorism
Edward Chace Tolman arrived at the University of California, Berkeley in 1918, harboring a deep dissatisfaction with both the sterile introspectionism of the past and the atomistic, mechanical reductionism of Watsonian behaviorism. Tolman sought to construct a psychology that remained firmly committed to empirical, objective observation while simultaneously recognizing the overarching, goal-directed nature of organismal activity. The theoretical system he articulated, known alternatively as purposive behaviorism or molar behaviorism, marked a radical departure from the molecular stimulus-response accounts of his contemporaries.
In his seminal 1932 work, Purposive Behavior in Animals and Men, Tolman differentiated between molecular and molar definitions of action. A molecular perspective parsed behavior into its constituent physiological fragments: the contraction of specific muscle groups, the firing of efferent nerve fibers, and the secretion of endocrine glands. Tolman argued that such fine-grained mechanical accounts, while physically accurate, failed to capture the emergent functional properties of behavior. In contrast, a molar approach recognized behavior as an integrated, emergent phenomenon characterized by distinctive holistic properties: it always aims at or away from an environmental goal, it exhibits flexible adaptation to physical obstacles, and it selects the shortest or least effortful path toward an objective (the principle of least effort). For Tolman, purpose and cognition were not mysterious, occult animisms residing in an untestable Cartesian soul; rather, they were objective, observable characteristics of molar behavior itself.
Central to Tolman’s theoretical innovations was the operationalization of “intervening variables.” Positioned mathematically and conceptually between the independent environmental stimuli ($S$) and the dependent behavioral responses ($R$), these intervening variables represented internal processes within the organism ($O$) that causally determined action. Tolman insisted that while internal cognitive events could not be directly observed through sensory inspection, their existence, structure, and functional states could be rigorously inferred through controlled experimental manipulations of the environment and precise measurement of subsequent molar performance. By introducing this $S-O-R$ architecture, Tolman legitimized the scientific investigation of animal mind, transforming teleological phenomena into rigorous, testable hypotheses regarding internal structural representation.
1.3 Philosophical Precursors to Internal Representation
Tolman’s conceptualization of animal cognition was profoundly shaped by intellectual movements occurring outside the perimeter of American behavioral orthodoxy. Chief among these was European Gestalt psychology, spearheaded by Max Wertheimer, Wolfgang Köhler, and Kurt Koffka, with whom Tolman interacted directly during his intellectual formation and sabbatical studies in Germany. The Gestaltists vehemently opposed elementarism, demonstrating that perceptual experience relies upon structural wholes, relational configurations, and contextual fields that cannot be decomposed into isolated sensory atoms without losing their essential psychological properties. Tolman absorbed this principle, recognizing that an organism navigating a maze does not perceive isolated walls, turns, or food cups as discrete stimuli, but rather perceives a dynamic, spatial-relational field—a unified Gestalt.
Further deepening this representational perspective was the epistemological legacy of Immanuel Kant, particularly Kant’s arguments in the Critique of Pure Reason regarding space as a synthetic a priori intuition. Kant asserted that external sensory impressions cannot be understood as an unstructured manifold; rather, the mind must actively organize perceptual inputs according to an inherent spatial framework that precedes experience. While Tolman naturalized and empiricalized this insight, the Kantian lineage remained unmistakable: an organism does not simply register point-by-point tactile and visual stimuli; it assimilates sensory experience into an overarching internal spatial framework. This internal framework endows environmental points with topological meaning, relational geometry, and metric coherence prior to the execution of any specific motor action.
Finally, Tolman was heavily indebted to American pragmatism, drawing heavily on the philosophical frameworks of Charles Sanders Peirce, William James, and John Dewey. Pragmatism posited that knowledge is inherently functional, an active tool forged through an organism’s practical transactions with an evolving environment. Tolman interpreted animal knowledge not as a passive, static repository of photographic facts, but as an active, predictive set of hypotheses, expectancies, and behavioral affordances. Organisms construct spatial models precisely because those models possess predictive utility, allowing them to anticipate the functional consequences of prospective actions. Tolman’s synthesis of Gestalt structuralism, Kantian spatial architecture, and pragmatist functionalism provided the robust epistemological scaffold necessary to conceptualize the cognitive map.
2. Theoretical Framework of Latent Learning and Cognitive Maps
2.1 Defining Latent Learning and Unreinforced Knowledge
The empirical and theoretical cornerstone of Tolman’s cognitive architecture was the phenomenon of latent learning. Historically defined, latent learning refers to the acquisition of environmental knowledge that occurs in the complete absence of primary biological reinforcement or drive satisfaction, remaining behaviorally unexpressed until an explicit motivational incentive is introduced. This concept struck at the heart of the prevailing Hullian and Thorndikian paradigms, which maintained as an inviolable axiom that learning could not occur without the immediate reinforcement of a stimulus-response connection via drive reduction.
Tolman severed the theoretical equation between performance and learning. In classical behaviorism, behavioral performance was assumed to be an isomorphic reflection of internal habit strength: if an animal made numerous navigational errors, it was inferred that minimal or no learning had taken place. Tolman demonstrated that this assumption conflated the structural acquisition of information with its behavioral execution. Animals exploring an unrewarded maze were learning its geometric structure, branch points, and environmental topography continuously, storing this knowledge in a latent, neural representational form. Learning, Tolman insisted, is a latent capacity that operates silently during unreinforced exploration.
The catalyst for the behavioral translation of this latent knowledge into observable performance is the introduction of an incentive. When food or water is suddenly introduced to an animal that has previously explored a maze without reward, the animal does not undergo a protracted, incremental process of habit formation. Instead, the incentive activates the latent representational map, prompting an immediate, dramatic realignment of behavior toward the goal. Reinforcement was thus dethroned from its status as the necessary causal agent of learning and relegated to its proper role as an executive motivator: reward does not forge the spatial knowledge; it merely provides the organism with a reason to utilize the spatial knowledge it already possesses.
2.2 The Mechanics of the Cognitive Map
To articulate the internal representational system that governed rodent navigation, Tolman introduced the metaphor of the cognitive map in his seminal 1948 essay, “Cognitive Maps in Rats and Men.” Tolman cautioned that the central nervous system must not be conceptualized as an oversimplified telephone switchboard, where external incoming calls (stimuli) are mechanically routed across switchboard jacks to outgoing motor lines (responses). Instead, Tolman asserted that the brain functioned far more like a central map room, wherein incoming environmental data are filtered, synthesized, and integrated into a comprehensive, dynamic cognitive map of the environment.
Within this framework, Tolman drew a critical theoretical distinction between what he classified as “strip maps” and “comprehensive maps.” A strip map represents a narrow, inflexible, and sequential representation of reality—functionally equivalent to a chain of motor reflexes or an egocentric set of directions: turn left at the first junction, walk twenty paces forward, and turn right. A strip map is fragile; should a single corridor along the path be obstructed, the entire navigational sequence breaks down because the organism lacks an understanding of the broader spatial matrix within which that path is situated. In contrast, a comprehensive cognitive map is a broad, allocentric representation encompassing the absolute and relative positions of environmental landmarks, corridors, open fields, and barriers.
A comprehensive map affords immense behavioral flexibility and navigational fluidity. Because the internal representation captures the overall topology of the environment, an animal possessing a comprehensive map can instantly compute detours around novel obstacles, formulate spontaneous shortcuts through previously unvisited routes, and navigate successfully toward a goal even if placed at an unprecedented starting location. The cognitive map is structured around environmental affordances—a term later popularized by ecological psychologist James J. Gibson—and field-expectancies. It represents not merely physical geometry, but an interconnected network of possibilities, indicating what spatial phenomena lead to what environmental consequences under specific behavioral conditions.
2.3 Expectancy Theory and Vicarious Trial and Error (VTE)
Complementing his spatial mapping hypothesis, Tolman developed expectancy theory to delineate the fine-grained cognitive processes operating during navigation. Departing from the peripheralist doctrine that behavior is propelled from behind by immediate sensory prods, Tolman posited that animals are pulled forward by internal expectancies of future states. He conceptualized these expectancies through the triad of the “sign-Gestalt”: a sign (an encountered environmental cue), a significate (an anticipated outcome or goal object), and a signified relation (the intervening spatial path or behavioral action linking the sign to the significate). Through repeated exposure to environmental regularities, organisms develop structural hypotheses: “If I move along this corridor (signified relation) marked by these visual textures (sign), I will encounter food (significate).”
Direct empirical evidence for the active, deliberative nature of these sign-Gestalt expectancies appeared in a behavioral phenomenon Tolman and his colleagues termed “Vicarious Trial and Error” (VTE). When a rodent arrives at a critical choice point or bifurcation within a maze, it does not immediately discharge a reflexive motor response. Instead, the animal frequently pauses, orienting its head back and forth between the competing corridors, visually and olfactorily inspecting each alternative path before committing to a physical traversal. Classical behaviorists dismissed this hesitation as mere behavioral ataxia or motor vacillation resulting from competing, equally balanced habit strengths.
Tolman, however, interpreted VTE as outward, observable evidence of internal cognitive deliberation. The animal was not physically sampling the pathways through trial and error; it was vicariously sampling them within its mental model. At the choice point, the rodent projects its attention down corridor A, internally testing the sign-Gestalt expectancy associated with that direction; it then turns its head and mentally tests the expectancy associated with corridor B. VTE demonstrated that before physical action occurred, an internal, cognitive evaluation was taking place. The frequency of VTE behaviors systematically peaked precisely when animals were actively learning the topological layout of a maze or confronting a sudden change in reward contingency, underscoring its role as a conscious, deliberative mechanism for updating and consulting the cognitive map.
3. The Landmark 1930 Tolman and Honzik Empirical Investigation
3.1 Experimental Architecture and Maze Geometry
In 1930, Edward C. Tolman and Charles H. Honzik published a landmark empirical study titled “Introduction and Removal of Reward, and Maze Performance in Rats” in the University of California Publications in Psychology. This experiment provided the definitive empirical proof of latent learning, systematically dismantling the Hullian assertion that reinforcement is an indispensable prerequisite for the acquisition of new environmental behavior. The investigation was meticulously designed to isolate the role of incentive from the process of information acquisition through rigorous experimental architecture and behavioral quantification.
The experimental apparatus consisted of a complex, fourteen-unit elevated T-maze. The maze possessed fourteen distinct choice points, each presenting the rodent with an authentic navigational branch: one arm led forward through a one-way swinging door toward the ultimate goal box, while the alternative arm led directly into a blind alley. To prevent the animals from back-tracking, which would introduce confounding variables into error metrics, each choice unit was fitted with mechanical one-way doors that permitted forward passage but closed gently behind the subject. The alleys were carefully painted a uniform gray, and ambient laboratory illumination was rigorously controlled to eliminate external directional light cues that might serve as extraneous beacons.
Tolman and Honzik implemented stringent experimental counter-controls. Olfactory confounding variables—a pervasive critique leveled by strict behaviorists against rodent spatial claims—were mitigated by regularly cleaning the maze pathways, rotating wooden floorboards between runs to scramble scent trails, and ensuring that blind alleys and correct corridors possessed identical scent profiles. Rodents were handled according to standardized protocols to minimize emotional reactivity, and hunger drives were strictly calibrated across all experimental cohorts through regulated feeding schedules. The subjects were divided into three fundamental experimental groups:
- Group 1: Regularly Rewarded (Continuously Reinforced): These rodents were placed in the maze daily and found food in the goal box upon every successful traversal. They served as the positive baseline for standard associative learning curves.
- Group 2: Non-Rewarded: These animals traversed the maze daily but never encountered food in the goal box. Upon reaching the end, they were simply placed into an empty goal box for a brief interval before being returned to their home cages, receiving their daily sustenance hours later under non-experimental conditions.
- Group 3: Delayed-Reward (Latent Learning Group): These subjects were treated identically to the non-rewarded group for the first ten days of the experiment, traversing the maze without finding any food in the goal box. On Day 11, however, without any advance warning, food was introduced into the goal box and remained there for all subsequent daily trials.
3.2 Observational Data and the Latent Learning Curve
The behavioral metrics recorded by Tolman and Honzik yielded results that directly challenged classical drive-reduction theory. For the first ten days, the performance of the three groups diverged in precise alignment with behaviorist expectations. The regularly rewarded cohort (Group 1) exhibited a steady, continuous, and predictable decline in both traversal times and navigational errors into blind alleys. Their learning curve presented the classic hyperbolic trajectory characteristic of incremental habit formation: errors diminished smoothly day after day as the rewarded associations were supposedly stamped in through drive reduction.
In stark contrast, the non-rewarded cohort (Group 2) and the delayed-reward cohort (Group 3) showed only marginal reductions in errors during the initial ten-day period. Because there was no food in the goal box, these animals ambled casually through the maze, entering numerous blind alleys, hesitating at intersections, and registering what behaviorists interpreted as a failure to learn. From a classical Hullian perspective, these animals possessed negligible habit strength because no stimulus-response bond had been reinforced by biological drive reduction. Under Thorndike’s Law of Effect, an organism could not systematically learn the correct path through fourteen consecutive T-junctions if that path terminated in an empty, neutral enclosure.
The paradigm-shifting revelation occurred on Day 11. Following the single introduction of food to Group 3 at the conclusion of their Day 11 run, their performance on Day 12 underwent an astonishing transformation. Rather than exhibiting a slow, incremental reduction in errors typical of an animal just beginning to learn a complex maze, Group 3 displayed an instantaneous, precipitous collapse in navigational errors. On Day 12 and subsequent days, the error curve of Group 3 plummeted immediately to match—and in some trials actually surpass—the performance of Group 1, which had been continuously rewarded for twelve consecutive days.
This dramatic shift demonstrated that the delayed-reward animals had not been idling aimlessly during their unrewarded traversals from Day 1 to Day 10. They had been actively mapping the spatial layout of the fourteen-unit maze, internalizing the topological relations between choice points, true corridors, and blind alleys. This vast store of spatial information had accumulated silently beneath the surface of overt behavior. The single introduction of the food reward on Day 11 did not create the learning; rather, it provided the motivational context that allowed the latent cognitive map to be instantly converted into peak, error-free behavioral performance.
3.3 Methodological Critiques and Experimental Counter-Controls
The publication of the 1930 Tolman and Honzik study provoked an intense intellectual counter-assault from orthodox behaviorists, led prominently by Clark Hull and Kenneth Spence. The behaviorists sought to defend the fundamental dogma of reinforcement theory by proposing alternative, non-cognitive interpretations of the latent learning data. The most significant critique was the “exploratory drive” hypothesis. Hullian theorists argued that exploratory behavior in rodents is not unrewarded, but is driven by an intrinsic exploratory drive, curiosity, or a homeostatic need to investigate novel stimuli. Under this view, reaching the end of a maze corridor provided minor drive reduction, meaning that Group 3 had technically received reinforcement all along, albeit of a weak and non-nutritive nature.
Tolman and his collaborators robustly countered these challenges. If exploratory drive reduction was sufficiently potent to forge habits, they argued, then Group 2 (the never-rewarded group) should have exhibited the same rapid, dramatic error reduction as Group 1 over the course of the twenty days, since their exploratory drive was being equally satisfied. Yet Group 2 never exhibited a steep error drop; their performance plateaued at a high error rate until the experiment concluded. The sudden, discontinuous plunge seen exclusively in Group 3 upon the single introduction of food could not be explained by a generic, low-grade exploratory reinforcement operating uniformly across every trial.
Further methodological skepticism centered on potential experimental artifacts. Critics suggested that rodents might be tracking microscopic olfactory cues left on the maze walls, relying on kinesthetic proprioception, or being unconsciously guided by external auditory landmarks in the laboratory. Subsequent replication studies addressed these concerns with extraordinary experimental rigor. Researchers constructed complex mazes with completely interchangeable modular components, systematically rotating alleyways, doors, and floors between trials to thoroughly scramble any latent olfactory trails. Blind, deafened, and anosmic rats were tested in identical paradigms; remarkably, these sensory-deprived animals still exhibited clear latent learning, constructing robust internal cognitive representations despite the elimination of specific sensory modalities. The latent learning effect proved to be an exceptionally robust psychological reality, resilient across divergent rodent strains and varied maze geometries.
4. The Spatial Shortcut Paradigm and Mental Geometry
4.1 Tolman, Ritchie, and Kalish (1946): Sunburst Maze Experiments
Having established that learning could proceed latently in the absence of biological reinforcement, Tolman sought to definitively disprove the foundational behaviorist claim that animal navigation is governed by simple chains of peripheral motor habits. According to Hullian theory, an animal learns a maze by acquiring a fixed sequence of muscle contractions—a kinesthetic chain: run forward five feet, turn right 90 degrees, run forward two feet, turn left 90 degrees. Under this reflex-chain view, the animal has no understanding of where the food box is located in external space; it merely possesses an automated chain of motor reflexes triggered by sequential physical stimuli.
To shatter this kinesthetic chaining hypothesis, Tolman, along with his students B.F. Ritchie and D. Kalish, published a landmark series of experiments in 1946 centered on the famous “Sunburst Maze.” The experimental apparatus utilized an initial training phase followed by a radical geometric transformation in the testing phase. During the training phase, rodents were introduced to an apparatus consisting of a circular central table connected to an entry pathway. From the circular table, an indirect, convoluted path led away: it progressed into an alley, turned sharply right, turned left, and then extended across a long corridor to a final goal box where food was consistently provided. For several days, the animals were trained until they mastered this convoluted, winding route, reliably traversing the specific sequence of turns to reach the food reward.
Once this motor habit was thoroughly entrenched, the critical testing phase commenced. The original convoluted pathway was completely dismantled and removed. In its place, the experimenters attached a wide, radiating array of eighteen novel paths projecting outward from the circular table like the rays of a sunburst. These paths spanned a full 360-degree arc, extending into regions of the room that the animals had never physically traversed. The original path was blocked. If the animals were governed by kinesthetic chaining or simple stimulus-response bonds, they should have demonstrated total navigational disorientation, or alternatively, they should have selected the paths closest to the initial physical direction of their trained motor habit (which had begun with an immediate right turn).
The empirical results delivered a devastating blow to the stimulus-response habit model. When placed on the circular table, the rodents did not select the path that mimicked their previous motor responses. Instead, after engaging in extensive Vicarious Trial and Error at the center of the table, an overwhelming majority of the animals selected Path 6—a novel, previously unvisited radial corridor that pointed directly through Euclidean space toward the precise geographical location where the food box had been situated during the training phase. The rats had not memorized a chain of muscle contractions; they had deduced the allocentric spatial coordinates of the reward and mentally computed a direct, novel vector across unvisited physical space to reach it.
4.2 Place Learning versus Response Learning Paradigms
To further isolate cognitive spatial representations from motor reflexes, Tolman, Ritchie, and Kalish developed the definitive “Place Learning versus Response Learning” paradigm, utilizing an elevated cross-maze (a plus-shaped apparatus). In this experimental design, the conflict between an allocentric spatial representation (place learning) and an egocentric motor program (response learning) was brought into sharp, empirical opposition. The cross-maze possessed two starting locations—South and North—and two goal arms—East and West.
In the “Response Learning” condition, the rodents were required to learn a specific motor response regardless of their starting position. If started from the South, they had to turn right to find food (leading to the East arm); if started from the North, they were also required to turn right to find food (which led to the West arm). Thus, for the response-learning cohort, the spatial location of the food varied dynamically across trials, but the required egocentric motor action—executing a 90-degree right turn—remained perfectly constant. Under classical behaviorist theory, this condition should have been effortlessly acquired, as it reinforced a simple, stable muscle reflex.
In the “Place Learning” condition, the spatial location of the food remained completely invariant: the reward was always situated at a fixed geographical coordinate in the room (for example, the East arm). However, the starting positions alternated between South and North. Consequently, when an animal started from the South, it was required to execute a right turn to reach the food; when started from the North, it had to execute a left turn to reach that same food location. To succeed, the place-learning animals had to abandon a fixed motor reflex and navigate according to an allocentric representation of absolute space.
The experimental outcome was decisive: place learning was acquired with vastly superior speed, accuracy, and stability compared to response learning. The place-learning rodents mastered the task within a mere fraction of the trials required by the response-learning cohort, which struggled persistently to reconcile the shifting spatial locations of the food. Tolman demonstrated that organisms naturally learn *places*, not isolated motor actions. When rich visual and environmental landmarks are available, navigation is overwhelmingly dominated by an allocentric cognitive map rather than an egocentric motor habit.
4.3 Euclidean Representation within Non-Human Cognition
The findings of the Sunburst and Cross-Maze experiments forced a fundamental reconceptualization of the internal architecture of animal cognition, establishing that non-human minds are capable of generating Euclidean spatial representations. Euclidean navigation requires an internal computational engine capable of preserving metric distance, topological invariant relations, and directional angularity across continuous space. Rather than encoding the environment as a disconnected series of topological nodes, the cognitive map functions as a unified coordinate matrix.
This Euclidean capacity is evidenced most clearly in the phenomenon of vector computation. When an organism navigates along an irregular, circuitous route—such as the convoluted training path of the Sunburst maze—it continuously integrates information regarding its own directional velocity, angular turns, and linear displacements. Through internal path integration (dead reckoning), the mammalian brain synthesizes these heterogeneous sensory and proprioceptive inputs, extracting the underlying topological invariants of the environment. The animal calculates an internal representation of its current position relative to a global origin, allowing it to compute a novel resultant vector that points directly across unvisited terrain toward a previously located target.
By establishing that animals calculate spatial vectors rather than repeating rote motor chains, Tolman’s work laid the philosophical and empirical foundation for modern computational theories of mind. It demonstrated that cognitive operations involve the manipulation of internal symbols, geometric coordinate transformations, and predictive simulations of prospective physical actions. Animals do not merely react to the physical surfaces they touch; they maintain an internal, geometric model of the unobserved world, utilizing mental calculations of distance, angularity, and direction to navigate both immediate spatial challenges and prospective future goals.
5. Neurobiological Validation: From Theory to Neuroanatomy
5.1 O’Keefe and Nadel’s Discovery of Hippocampal Place Cells
For three decades following the publication of Tolman’s 1948 treatise, the cognitive map remained an abstract, purely psychological construct—a brilliant intervening variable that lacked an identified physiological mechanism within the central nervous system. Orthodox neurophysiologists, still deeply influenced by Cartesian reductionism, frequently dismissed the cognitive map as a mentalistic fiction. This theoretical impasse was shattered in 1971 when John O’Keefe and Jonathan Dostrovsky published their monumental discovery of location-specific firing patterns within the rodent hippocampus, followed in 1978 by O’Keefe and Lynn Nadel’s paradigm-defining monograph, The Hippocampus as a Cognitive Map.
Utilizing chronically implanted microelectrodes to record individual action potentials from pyramidal neurons within the CA1 and CA3 subfields of the hippocampus in freely moving rats, O’Keefe and Dostrovsky observed a phenomenon that defied classical sensory physiology. These neurons did not fire in response to specific, individual sensory stimuli—such as a flash of light, a pure auditory tone, or a localized tactile pressure. Nor did they fire in correlation with specific motor actions, such as running, sniffing, or grooming. Instead, an individual hippocampal pyramidal neuron fired at a high frequency if and only if the animal was physically situated within a specific, restricted territory of the experimental apparatus.
O’Keefe formally christened these neuroanatomical units place cells, and designated the specific physical zone in which a cell fired as its “place field.” A given place cell fires robustly whenever the animal enters its place field, regardless of the direction from which the animal approached, the motor postures it assumed, or the specific behavioral task it was performing. Crucially, the collective activity of an ensemble of place cells forms a continuous, comprehensive, and redundant mapping of the entire available environment. If an animal moves to a completely novel environment, the existing place fields dynamically remap, establishing an entirely new, stable coordinate configuration for that novel space within minutes. O’Keefe and Nadel explicitly recognized that they had uncovered the direct, biological implementation of Edward Tolman’s cognitive map within the archicortex of the mammalian brain.
5.2 Entorhinal Grid Cells and Metric Coordinate Systems
While the discovery of hippocampal place cells confirmed that the brain constructs an internal representation of specific locations, it left a profound neurocomputational question unanswered: how does the nervous system compute metric distance, direction, and spatial scale? Place fields appeared topologically flexible and varied in size depending on environmental context; they did not seem to possess the regular, crystalline geometric metric required to calculate precise Euclidean vectors across unvisited space. The resolution to this puzzle arrived in 2005 through the groundbreaking discoveries of Edvard Moser, May-Britt Moser, and their students at the Norwegian University of Science and Technology.
Recording from the dorsomedial entorhinal cortex (MEC)—the principal cortical input structure projecting directly into the hippocampus via the perforant path—the Mosers discovered a radically different class of spatial neurons: grid cells. Unlike a hippocampal place cell, which possesses a single, discrete firing field within a given environment, an entorhinal grid cell fires at multiple, regularly spaced locations. Astonishingly, when an animal explores an open arena, the multiple firing fields of a single grid cell form a strikingly periodic, hexagonal, tessellating triangular lattice that tiles the entire physical surface of the environment, resembling an internal isometric coordinate sheet or hexagonal graph paper.
Grid cells provide the universal, metric coordinate framework that underpins Tolmanian mental geometry. They operate as an internal metric ruler, calculating path integration through the continuous mathematical integration of self-motion cues, including speed, angular velocity, and directional displacement. Grid cells are anatomically organized along a topographical gradient within the medial entorhinal cortex: cells located at the dorsal pole exhibit exceptionally tight, fine-grained spatial periodicities (with grid fields spaced mere centimeters apart), whereas cells situated progressively toward the ventral pole exhibit vastly expanded spatial scales, with grid vertices spanning several meters. This multi-scale, modular metric system provides the computational substrate that allows the brain to map environmental scale, compute Euclidean distances between disparate locations, and plan direct shortcut trajectories across novel terrain.
5.3 Head-Direction Cells and Boundary Vector Cells
The neural cognitive mapping system is not limited to place and grid cells; it comprises a highly integrated, multimodal navigational network involving several specialized neuronal cell types distributed across the hippocampal-parahippocampal-diencephalic axis. Fundamental among these are head-direction cells, initially discovered by James Ranck and rigorously characterized by Jeffrey Taube in the postsubiculum, anterior thalamic nuclei, and lateral mammillary bodies. Head-direction cells function as the brain’s internal compass: an individual head-direction cell fires at a maximum rate whenever the animal’s head is oriented in a specific azimuth within the horizontal plane, independent of the animal’s physical location or ongoing behavior. This internal compass orientation is calibrated by static visual landmarks but remains fully functional in complete darkness, driven continuously by vestibular inputs from the semicircular canals.
Working in close mechanical synergy with head-direction and grid cells are boundary vector cells (and border cells), discovered within the subiculum and entorhinal cortex by Colin Lever, Neil Burgess, and their colleagues. Boundary vector cells fire selectively when the organism is located at a specific distance and allocentric direction from an environmental barrier or geometric boundary, such as a maze wall, vertical drop-off, or physical partition. These cells provide the indispensable anchor that ties the internal metric coordinate system of the grid cells and the place representations of the hippocampus to the real physical architecture of the external world.
The collective synthesis of these neurobiological elements realizes Tolman’s vision of an integrated spatial mapping engine. Multi-modal sensory inputs (visual, tactile, auditory) are synthesized with internal idiothetic signals (vestibular, proprioceptive, motor efference copies) within the parahippocampal structures. The boundary vector cells and head-direction cells calibrate the entorhinal grid matrix, which in turn drives and stabilizes the high-capacity, contextual representations of the hippocampal place cells. The brain does not passively mirror external stimuli; it actively generates a self-sustaining, mathematically rigorous internal simulation of Euclidean space.
6. Bridging Spatial Cognition and the Episodic Future
6.1 Conceptual Overlap Between Spatial Maps and Temporal Projection
For decades following the validation of Tolman’s cognitive map by O’Keefe and Nadel, the hippocampus and its adjacent cortical structures were categorized almost exclusively as an internal GPS—a dedicated spatial navigational engine. In parallel, human neuropsychology, rooted in the study of amnesic patients, treated the medial temporal lobe primarily as an episodic memory system dedicated to the retrospective recall of personal biographical history. For decades, these two massive scientific literatures operated in virtual isolation from one another, separated by an artificial conceptual divide between spatial navigation in non-human animals and conscious memory in humans.
In the late 2000s, an extraordinary conceptual synthesis emerged, fundamentally transforming cognitive science. Theorists began to recognize that a cognitive map is not merely a tool for locating oneself in physical space; it is a general relational coordinate system for organizing information across all continuous domains, including time. Spatial navigation and episodic memory share identical computational requirements: both require an agent to orient its perspective, calculate relationships between disparate events or locations, and traverse a network of connected nodes. When we remember a past autobiographical event, we are mentally navigating backward through a coordinate matrix of subjective time, relocating our cognitive perspective to a specific spatio-temporal coordinate.
From an evolutionary perspective, however, memory did not evolve simply to allow organisms to reminisce about the past. Retrospection offers no adaptive advantage on its own; natural selection acts entirely on prospective behavior that enhances survival and reproductive fitness. The primary evolutionary driver for the development of an episodic memory system was almost certainly prospective: memory provides the raw informational substrate required to anticipate, plan for, and navigate the episodic future. Physical space and subjective time are fundamentally interleaved within the mammalian nervous system; mental time travel is essentially navigation through a spatiotemporal cognitive map.
6.2 Episodic Memory as the Substrate for Prospective Cognition
The mechanical bridge connecting retrospective memory to prospective forecasting was formalized by Daniel Schacter and Donna Rose Addis in their influential Constructive Episodic Simulation Hypothesis. Schacter and Addis addressed a long-standing paradox in cognitive psychology: why is human episodic memory so profoundly reconstructive, malleable, and prone to error? If the biological imperative of memory were merely to store an accurate, permanent recording of the past—akin to a digital video archive—the human brain would be judged an extraordinarily flawed apparatus.
The Constructive Episodic Simulation Hypothesis posits that memory is deliberately designed to be flexible, reconstructive, and combinatorially fluid precisely so that its constituent elements can be dynamically recombined to imagine, simulate, and plan for novel future scenarios that have never actually occurred. When we imagine a future scenario—for example, meeting a collaborator at an unfamiliar conference center next year—the brain does not generate an image from nothing. Instead, the hippocampal system engages in a combinatorial extraction process: it accesses memory traces from diverse past experiences (the face of the colleague, the layout of standard conference centers, previous conversations), deconstructs them into their structural components, and seamlessly recombines these fragments into a cohesive, novel prospective simulation.
The hippocampus functions in this prospective architecture as a master spatial-relational index. It acts as an internal coordinate generator that constructs a coherent, multidimensional spatial scene within which the prospective event can unfold. Without an intact hippocampal spatial scaffolding, the simulated fragments cannot be integrated into a unified mental image, and prospective thought collapses into disconnected, non-visual semantic facts. Tolman’s cognitive map is thus revealed to be the essential physical stage upon which the episodic future is mentally simulated and enacted.
6.3 Tolman’s Vicarious Trial and Error as Pre-play
This prospective reconceptualization of the cognitive map provides an extraordinary retrospective validation of Edward Tolman’s early behavioral observations. In the 1930s, Tolman described Vicarious Trial and Error (VTE) as the behavioral hesitation at a maze choice point where an animal looks back and forth between paths, mentally evaluating options before committing to a physical action. For eighty years, behaviorists dismissed VTE as a marginal motor habit, while mainstream neuroscience lacked the electrophysiological tools to probe what was occurring inside the nervous system during those crucial seconds of behavioral pause.
In a series of landmark studies beginning in 2007, neurophysiologists David Redish, Matthew Wilson, and their colleagues achieved the technological capability to record the simultaneous activity of hundreds of hippocampal place cells in real time as rodents performed spatial decision-making tasks. The empirical findings were breathtaking. When a rat arrives at an elevated choice point and engages in classic Tolmanian VTE behavior—pausing and turning its head toward an unchosen path—the hippocampal place cell ensemble does not remain static, nor does it merely represent the animal’s current physical position.
Instead, researchers observed rapid, transient, forward-sweeping sequences of place-cell activation that emanate outward from the animal’s current position, projecting sequentially down the physical arm of the maze that the animal is visually inspecting. Termed hippocampal forward replay or pre-play, this neural activity represents the direct, physiological realization of Tolmanian sign-Gestalt expectancy evaluation. The rat’s brain is literally running a fast-forward, prospective simulation of future spatial traversal: it projects its internal coordinate perspective down corridor A, anticipates the potential hazards or rewards situated at the terminal point, then sweeps its neural projection down corridor B, testing alternative prospective trajectories. VTE is not an outward symptom of behavioral confusion; it is the observable, molar manifestation of prospective neural simulation.
7. Neural Mechanisms of Episodic Future Simulation
7.1 The Default Mode Network and Prospective Thought
In the human brain, the neural machinery responsible for constructing the episodic future relies upon an extensive, highly interconnected macroscale functional architecture known as the Default Mode Network (DMN). First identified by Marcus Raichle and colleagues as a set of brain regions showing consistent metabolic deactivation during demanding, externally oriented perceptual tasks, the DMN was subsequently recognized by Randy Buckner, Daniel Schacter, and others as the core biological engine of internally generated cognition, self-projection, and mental time travel.
The DMN is anchored by a core network of anatomically interconnected structures:
- The Medial Temporal Lobe Subsystem: Centered on the hippocampus, parahippocampal cortex, and entorhinal cortex, this hub is responsible for the retrieval of episodic details, the generation of spatial scenes, and the assembly of relational representations.
- The Medial Prefrontal Cortex (mPFC): Functioning as an executive evaluator, the mPFC integrates personal preferences, goals, and affective valuation into the ongoing prospective simulation.
- The Posterior Cingulate Cortex (PCC) and Retrosplenial Cortex (RSC): Acting as crucial computational bridges, these regions mediate perspective transformations, translating allocentric spatial coordinate maps generated in the hippocampus into egocentric, first-person visual perspectives suitable for mental simulation.
Neuroimaging paradigms demonstrate an almost complete functional overlap between the neural regions recruited when an individual remembers an autobiographical event from their past and when they imagine a novel personal scenario occurring in their episodic future. When a participant is instructed in an fMRI scanner to imagine a realistic event occurring five years from now, the DMN exhibits robust, coordinated metabolic activation. Crucially, functional dissociation studies reveal that while semantic future planning (such as knowing that a train will depart at 5:00 PM tomorrow) relies primarily on lateral temporal and inferior prefrontal cortices, the rich, dynamic, experiential instantiation of an episodic future event—where one mentally projects oneself into an imagined spatiotemporal coordinate and visualizes the scene unfolding—strictly requires the synchronized engagement of the DMN and the hippocampal spatial mapping matrix.
7.2 Hippocampal Sharp-Wave Ripples and Forward Trajectory Replay
At the microcircuit level, the primary neurophysiological mechanism supporting both memory consolidation and the prospective simulation of future trajectories is the Sharp-Wave Ripple (SWR) complex. Sharp-wave ripples are extremely high-frequency (150–250 Hz) oscillatory field potentials that originate primarily within the CA3 subfield of the hippocampus and propagate through CA1 to the subiculum, entorhinal cortex, and widespread neocortical networks. SWRs occur predominantly during periods of quiet behavioral immobility, consummatory pause, and slow-wave sleep—states during which the brain is liberated from processing immediate external sensory traffic.
During these brief, millisecond-scale SWR events, the hippocampus compresses and replays sequences of place-cell firing that correspond to physical trajectories through space. Remarkably, this replay is not merely retrospective; it is profoundly prospective. In “forward replay” or “pre-play,” the place cell sequences fire in a compressed temporal order that predicts the specific path the animal will physically traverse in the immediate future. Compressed by a factor of roughly ten to twenty relative to physical behavioral velocity, these SWR-mediated prospective sweeps allow the brain to mentally traverse a long spatial path in a fraction of a second.
Recent discoveries indicate that these prospective sharp-wave ripple events are actively modulated by value gradients. When an animal pauses before embarking on a complex navigational route, forward SWRs project selectively toward spatial locations that contain high-value goals or rewards, even if those paths are topologically complex. During these rapid forward projections, the hippocampus collaborates with downstream subcortical structures to read out the expected utility of the path. Sharp-wave ripples are not passive neurological echoes; they are high-speed computational simulations designed to optimize path selection, consolidate the structural topology of the cognitive map, and evaluate alternative prospective futures before committing bodily energetic resources to physical execution.
7.3 Prefrontal-Hippocampal Coupling in Scenario Evaluation
The generation of a prospective simulation within the hippocampal-entorhinal network is only the first stage of effective future-oriented behavior; the generated scenario must be dynamically evaluated, regulated, and translated into an adaptive behavioral decision. This critical executive control is achieved through dense, bidirectional neuroanatomical coupling between the hippocampal formation and the prefrontal cortex, specifically the medial prefrontal cortex (mPFC) and the orbitofrontal cortex (OFC), operating in intimate concert with the ventral striatum.
Electrophysiological recordings during prospective decision-making reveal precise theta-band (4–8 Hz) and gamma-band (30–80 Hz) phase synchronization between the hippocampus and the prefrontal cortex. As the hippocampus constructs an imagined forward trajectory, the orbitofrontal cortex and ventral striatum immediately assign an affective and hedonic value to the simulated outcome, essentially computing the prospective subjective utility of that specific future state. If a simulated path leads toward an anticipated hazard or a low-value reward, the prefrontal cortex exerts top-down inhibitory gating, terminating the hippocampal simulation of that pathway and prompting the network to mentally simulate alternative routes.
This prefrontal-hippocampal dialogue is the computational engine of model-based decision making. Rather than relying on simple stimulus-response habits stamped into the basal ganglia through past reinforcement, the prefrontal-hippocampal axis constructs an internal forward model of the world. It queries the cognitive map: “If I execute action $A$, what subsequent environmental state $S’$ will occur, and what is the hedonic value of $S’$?” Through this closed-loop iterative simulation, maladaptive prospective paths are pruned away internally, allowing the organism to select the optimal behavioral trajectory entirely within the safe domain of internal mental simulation.
8. Model-Based Reinforcement Learning as Contemporary Latent Learning
8.1 Model-Free versus Model-Based Computational Paradigms
In modern computational neuroscience and artificial intelligence, the historic theoretical debate between Clark Hull’s stimulus-response behaviorism and Edward Tolman’s purposive cognitive mapping has been formalized mathematically through the framework of reinforcement learning (RL). Contemporary computational theory bifurcates agent architecture into two fundamentally distinct computational paradigms: “model-free” and “model-based” reinforcement learning.
Model-free RL serves as the direct mathematical formalization of classical Hullian and Thorndikian behaviorism. In a model-free architecture, an agent does not possess an internal model of the environment; it maintains no representation of transition probabilities between states, nor does it store knowledge of what specific outcomes follow specific actions. Instead, it merely tracks a direct, cached value—a “Q-value”—associated with taking a specific motor action $a$ in a specific stimulus state $s$. Learning occurs incrementally via temporal-difference prediction errors, which slowly stamp in or stamp out the cached value of a state-action pair based on experienced reward. Model-free agents are computationally cheap and execute actions with extreme speed; however, they are profoundly inflexible. If the reward landscape changes or an environmental corridor is blocked, a model-free agent continues blindly executing its cached habits until dozens of failed experiences slowly degrade the stamped-in associations.
Conversely, model-based RL represents the exact mathematical formalization of Tolman’s cognitive map. A model-based agent explicitly constructs, updates, and stores an internal model of the world consisting of two foundational mathematical functions: a transition function $T(s, a, s’)$, which defines the probability of moving from state $s$ to a novel state $s’$ given action $a$, and a reward function $R(s, a)$, which specifies the reward expected from that transition. When confronted with a choice point, a model-based agent does not merely consult a cached habit; it executes a forward tree-search algorithm across its internal transition matrix. It mentally steps forward through time, projecting sequences of actions, anticipating subsequent states, and calculating the optimal policy dynamically. Model-based RL is computationally demanding and requires substantial working memory; however, it exhibits the exquisite flexibility, rapid adaptability, and spontaneous detour capacity that Tolman observed in his rodents nearly a century ago.
8.2 Successor Representations: Reconciling Space and Value
While model-based algorithms fully capture Tolmanian cognitive maps, pure forward tree search is computationally intensive and scales poorly in large, complex environments. In 1993, computational neuroscientist Peter Dayan proposed a brilliant mathematical synthesis known as the Successor Representation (SR), which has recently emerged as a dominant theoretical model for understanding hippocampal-entorhinal function. The Successor Representation elegantly reconciles the computational efficiency of model-free habit caching with the predictive, representational flexibility of a model-based cognitive map.
The mathematical formulation of the Successor Representation decouples the representation of the environment’s structural dynamics from its immediate reward values. Instead of storing the full step-by-step transition function $T(s, a, s’)$ or caching simple scalar values, an SR agent computes and stores a predictive matrix representing the expected discounted future occupancy of all states $s’$ given a current starting state $s$:
$$M(s, s’) = \mathbb{E}\left[\sum_{t=0}^{\infty} \gamma^t \mathbb{I}(s_t = s’) mid s_0 = s\right]$$
Here, $\gamma$ represents a temporal discount factor, and $\mathbb{I}$ is an indicator function that equals 1 when the agent occupies state $s’$. This $M$ matrix is literally a predictive map of the environment: it captures the geometric topology and structural connectivity of the maze by encoding how frequently future states will be visited from any current location, entirely independent of whether those future states contain food, water, or electric shock.
The deep convergence between the Successor Representation and Tolmanian latent learning is mathematically profound. Because the SR encodes the predictive spatial structure of an environment without reference to reward, it can be learned completely latently during unreinforced exploration. If an agent wanders through a fourteen-unit T-maze without food, its nervous system can continuously compute and update the structural matrix $M(s, s’)$. When a reward is suddenly placed in the goal box at state $s_g$, the agent does not need to slowly propagate cached values backward through trial and error. It simply multiplies its latently acquired predictive structural map $M$ by the newly discovered immediate reward vector $R$:
$$V(s) = \sum_{s’} M(s, s’) R(s’)$$
In a single computational step, the agent calculates the complete value function across the entire maze, instantly producing the precipitous drop in navigational errors observed on Day 11 of the Tolman-Honzik experiment. Recent neurocomputational work by Kimberly Stachenfeld, Matthew Botvinick, and Samuel Gershman has demonstrated that the firing fields of hippocampal place cells and entorhinal grid cells correspond mathematically to the eigenvectors and predictive rows of this Successor Representation matrix.
8.3 Reward Devaluation Experiments and Cognitive Flexibility
The empirical gold standard for dissociating model-free behavioral habits from model-based cognitive map representations in the laboratory is the “Reward Devaluation” paradigm, pioneered by Anthony Dickinson and Bernard Balleine. This paradigm exposes the deep cognitive differences between animals operating via automatic stimulus-response chaining versus those utilizing predictive internal representations.
In a prototypical devaluation experiment, an animal is trained to perform two distinct actions that yield two distinct food outcomes: pressing Lever 1 yields sucrose pellets, while pressing Lever 2 yields maltodextrin solution. Once this behavioral repertoire is firmly established, the critical devaluation phase occurs outside the testing apparatus. One of the rewards (for example, the sucrose pellets) is devalued—either by pairing its consumption with an injection of lithium chloride to induce a conditioned taste aversion, or by permitting the animal to consume sucrose to absolute sensory-specific satiety. Crucially, this devaluation takes place in the animal’s home cage; the animal has never experienced the levers in conjunction with the now-nauseating or devalued food.
The animal is then returned to the testing chamber and presented with the two levers in an extinction test, meaning that no food is actually delivered upon pressing. If the animal’s behavior is driven by a model-free, stimulus-response habit stamped into the sensorimotor striatum, the presentation of Lever 1 will automatically trigger the habitual motor response, because the mechanical stimulus-response bond has not been directly extinguished through unrewarded trials in the box. The animal should press Lever 1 and Lever 2 at equal, habitual rates.
However, if the animal is navigating via a model-based cognitive map, it executes an internal prospective simulation: “Pressing Lever 1 leads to sucrose; sucrose is currently nauseating; therefore, do not press Lever 1.” In animals exhibiting intact model-based control, the execution of the devalued action drops instantly to zero on the very first trial, without requiring a single unrewarded exposure to the lever. This immediate, rational adaptation to altered reward states provides indisputable empirical proof of cognitive mediation. The animal updates the internal value vector within its structural world model, allowing the prospective simulation of future states to immediately overrule physical habits.
9. The Cognitive Map Beyond Physical Space: Abstract Knowledge Domains
9.1 Social and Relational Cognitive Maps
While Edward Tolman developed the cognitive map to explain rodent locomotion through physical mazes, he intuitively anticipated that the identical representational architecture must operate across abstract, non-spatial domains of human and animal thought. In his 1948 essay, Tolman explicitly warned against “narrow strip-maps” in human sociopolitical life, which produce prejudice, xenophobia, and ideological fanaticism, advocating instead for the educational cultivation of broad, comprehensive cognitive maps encompassing social and psychological relations. Over the past decade, cognitive neuroscience has revealed that Tolman’s metaphor is a literal biological truth: the hippocampal-entorhinal coordinate system is universally conserved to map non-spatial, relational, and social domains.
Groundbreaking fMRI investigations by Rita Tavares, Daniela Schiller, and colleagues have shown that humans track social hierarchies and affiliative relational networks utilizing the precise hippocampal-prefrontal circuits originally evolved for physical navigation. When human participants navigate a multi-agent social network, tracking agents across dimensions of social power (status/dominance) and social affiliation (intimacy/trust), the hippocampal formation activates dynamically. The brain treats social space as a continuous, two-dimensional geometric manifold: each social acquaintance occupies an allocentric coordinate determined by their power and affiliation vectors.
Furthermore, human participants performing transitive inference tasks—such as deducing that if individual $A$ is dominant over individual $B$, and $B$ is dominant over $C$, then $A$ must be dominant over $C$—rely on entorhinal-hippocampal circuitry to compute relational distances. Studies conducted by Christian Doeller, Jacob Bellmund, and Alexandra Constantinescu have demonstrated that when humans learn relationships between abstract stimuli defined by two non-spatial continuous dimensions (such as the length of a bird’s neck versus the length of its legs), the medial entorhinal cortex exhibits a characteristic hexagonal, grid-like blood-oxygen-level-dependent (BOLD) modulation as participants mentally traverse this abstract conceptual space. The human brain repurposes its ancient evolutionary metric grid to calculate distances, vectors, and shortcuts through abstract conceptual, relational, and social topologies.
9.2 Semantic Traversal and Conceptual Organization
The discovery that entorhinal-hippocampal networks map continuous non-spatial features provides a revolutionary neurocomputational framework for understanding human semantic memory and conceptual organization. Rather than storing linguistic knowledge and encyclopedic facts as an unstructured, static database of isolated lexical entries, the brain organizes semantic knowledge within low-dimensional geometric manifolds.
Within this semantic coordinate space, concepts are represented as topological points or vectors. Conceptual distance correlates directly with metric distance within the neural manifold: “robin” and “sparrow” are situated in close spatial proximity, while “robin” and “refrigerator” are separated by vast relational gulfs. When the human mind performs an analogy—such as the classic computational linguistic vector transformation “king is to queen as man is to woman”—it is executing an allocentric vector translation across an internal semantic cognitive map:
$$\vec{v}_{\text{queen}} \approx \vec{v}_{\text{king}} – \vec{v}_{\text{man}} + \vec{v}_{\text{woman}}$$
The cognitive map operates as the geometric matrix that makes rapid relational inference, metaphorical mapping, and conceptual synthesis computationally possible.
This spatialization of semantic architecture illuminates the neural basis of human problem solving and creative insight. When an individual confronts a novel, ill-structured intellectual problem, they do not engage in a random, brute-force search of all possible semantic associations. Instead, they navigate the conceptual manifold, utilizing cognitive map operations to compute shortcuts between previously disconnected semantic domains. Insight—frequently experienced as a sudden, discontinuous “Aha!” moment—is computationally equivalent to the discovery of a spatial shortcut in a Sunburst maze: the brain computes a novel vector that bypasses conventional, winding categorical pathways, bridging disparate conceptual territories to generate an entirely novel synthesis.
9.3 Temporal Mapping and Event Horizon Planning
In addition to mapping continuous spatial coordinates and semantic topologies, the hippocampal formation contains dedicated neurobiological machinery for mapping the pure, continuous passage of physical time. In 2011, Howard Eichenbaum, Christopher MacDonald, and their colleagues discovered hippocampal “time cells”. Recording from the CA1 subfield of rodents during paradigms where spatial position was held strictly invariant—such as running on a stationary treadmill or waiting during an imposed delay period—they identified individual pyramidal neurons that fired sequentially at specific, highly reproducible temporal moments during the interval.
Just as place cells tile a continuous physical surface with spatial fields, hippocampal time cells tile a continuous temporal interval with discrete temporal receptive fields. Time cells maintain an ongoing, internal metric of temporal duration and chronological event order, firing in a choreographed, relay-like cascade that spans the temporal bridge between an initial stimulus and a delayed outcome. Crucially, time cells do not operate in isolation from place cells; the very same hippocampal neurons multiplex both spatial and temporal coordinates, creating a four-dimensional spatiotemporal metric: an integrated coordinate manifold of subjective spacetime.
This spatiotemporal interleaving is what allows organisms to structure the episodic future into coherent, temporally distant narratives. When we plan for the future, we do not merely anticipate an abstract set of isolated events; we situate those events along an extended, metric temporal horizon: we envision what we must do in five minutes, what will occur next Tuesday, and where we hope to be in five years. The time-cell architecture provides the temporal metric that spaces out our prospective simulations, allowing the cognitive map to project coherent, goal-directed itineraries across both the physical geography of the world and the unfolding chronological corridor of human life.
10. Comparative and Evolutionary Dimensions of Prospective Navigation
10.1 Avian Spatial Memory and Episodic-Like Foresight
The evolutionary pressures that forged the cognitive map and its prospective temporal capacities are not restricted to mammals; they are vividly apparent across diverse taxa that possess complex ecological niches. Exceptional among these are food-caching avian species within the families Corvidae (jays, crows, ravens, nutcrackers) and Paridae (chickadees, tits). In harsh alpine and temperate climates, these birds face severe winter survival challenges that cannot be resolved through immediate foraging; they must cache tens of thousands of individual food items across square miles of heterogeneous territory during autumn and recover them months later beneath deep snowpacks.
In a groundbreaking series of experiments beginning in the late 1990s, Nicola Clayton and Anthony Dickinson demonstrated that western scrub-jays (Aphelocoma californica) do not recover their hidden caches through random search or simple olfactory tracking. Instead, they possess robust “episodic-like memory,” successfully fulfilling the rigorous “what-where-when” criteria. Scrub-jays were allowed to cache two types of food in distinct, visually identified ice-cube trays: highly preferred but perishable wax worms, and less preferred but non-perishable peanuts. When tested after a brief retention interval (4 hours), the jays selectively recovered the perishable wax worms. However, when tested after a long retention interval (124 hours)—by which time the wax worms had decayed and become unpalatable—the jays completely avoided the wax worm locations and selectively excavated the peanuts. They recalled not merely what food was cached and where it was located in space, but also how long ago the caching event had transpired.
Subsequent investigations by Clayton and colleagues revealed that corvid cognition extends into genuine prospective foresight. Western scrub-jays demonstrate the capacity to plan for future hunger independent of their current motivational state. When placed in an experimental paradigm where they learned through experience that a specific compartment would contain no food the following morning, the jays selectively cached abundant food in that specific “no-breakfast” room the evening before, even though they were completely sated at the moment of caching. This extraordinary behavioral flexibility is accompanied by profound neuroanatomical adaptations: food-storing corvids and parids exhibit massive, seasonal hypertrophy of the avian hippocampus, showing a substantial expansion of neurogenesis and total hippocampal volume directly correlated with the spatial complexity of their caching territories and the temporal horizon of their prospective planning.
10.2 Primate Foraging Strategies and Mental Time Horizons
In the primate lineage, ecological foraging demands within complex, heterogeneous tropical rainforest canopies served as the primary evolutionary catalyst for the co-expansion of the cognitive map and the episodic future horizon. Frugivorous primates, such as chimpanzees (Pan troglodytes) and baboons (Papio ursinus), subsist on ephemeral, patchy, and highly seasonal resources: fruit-bearing trees may be distributed across dozens of square kilometers, maturing unpredictably and remaining ripe for only a few days before rotting or being stripped by competitors.
Extensive field studies conducted by Christophe Boesch, Karline Janmaat, and colleagues in the Taï National Park, Côte d’Ivoire, have demonstrated that wild chimpanzees navigate these complex tropical forests utilizing sophisticated allocentric cognitive maps combined with long-range temporal projection. Chimpanzees do not forage via random walks or simple line-of-sight visual searches. Instead, they compute complex, multi-destination spatial itineraries, traveling in direct, purposeful, and geometrically optimized trajectories toward specific target trees that are kilometers away and completely obscured by dense jungle foliage.
Crucially, these foraging trajectories are temporally prospective and weather-dependent. Janmaat and colleagues documented that female chimpanzees systematically alter their morning departure times and travel speeds based on predictions of future fruit availability and competitive dynamics. On mornings when they plan to forage on high-value, highly competitive fig species, the chimpanzees depart their sleeping nests well before dawn, navigating through pitch darkness along direct spatial vectors to arrive at the fruit tree precisely as the sun rises, preempting competing primate groups. Furthermore, their routes account for the thermal trajectory of the coming day: if the forecast calls for intense afternoon heat, they adjust their morning route to visit exposed, sun-drenched feeding sites first, leaving shaded, protected valley sites for the midday heat. The mental horizon of wild primates is an integrated spatiotemporal simulation, continuously balancing metric distance against anticipated future states.
10.3 Evolutionary Fitness Advantages of Latent Environmental Modeling
The profound energetic expense of maintaining large mammalian and avian brains—tissues that consume a massively disproportionate share of the organism’s basal metabolic budget—presents an evolutionary enigma. Why did natural selection favor the emergence of complex, energy-demanding neural machinery for latent learning and prospective simulation over the simple, metabolically inexpensive architecture of Hullian stimulus-response conditioning? The answer lies in the massive, compounding survival advantages that an internal world model confers upon an organism navigating a volatile and hazardous natural world.
The primary evolutionary advantage of latent environmental modeling is the radical minimization of mortality risk. In the wild, trial-and-error learning is an exceptionally hazardous enterprise: an animal that must learn to avoid a stalking leopard or a sheer cliff face through direct, experiential reinforcement is extraordinarily unlikely to survive its initial educational mistakes. Latent learning allows an organism to explore an environment during periods of relative safety, systematically acquiring an accurate, detailed internal cognitive map of its topology, escape corridors, physical barriers, and hiding refuges without requiring any immediate drive satisfaction.
When a lethal predator suddenly attacks, the animal does not engage in random, frantic motor thrashing or wait for a slow, incremental associative habit to be stamped in. Instead, it instantly queries its latent cognitive map, computes a novel Euclidean escape vector through unvisited underbrush toward a safe rock crevice, and survives. Furthermore, prospective simulation provides massive energetic efficiency. As the philosopher Karl Popper famously observed, the possession of an internal representational world model allows our hypotheses and prospective plans to die in our stead. By mentally testing alternative foraging itineraries and spatial trajectories across an internal cognitive map, an organism selects the path of least resistance and highest energetic yield entirely within the virtual safety of its neural architecture, preserving precious calories and dramatically maximizing evolutionary fitness.
11. Clinical Pathologies of the Cognitive Map and Prospective Thought
11.1 Alzheimer’s Disease and Topological Disorientation
Because the biological substrates of Tolmanian cognitive mapping and episodic future simulation are localized precisely within the medial temporal lobe and its interconnected default mode networks, these cognitive capacities are exceptionally vulnerable to specific neurodegenerative cascades. Foremost among these is Alzheimer’s disease (AD), a progressive neurodegenerative disorder characterized histopathologically by the accumulation of extracellular amyloid-beta plaques and intracellular neurofibrillary tangles composed of hyperphosphorylated tau protein.
The neuropathological staging formulated by Heiko and Eva Braak demonstrates that the earliest manifestations of neurofibrillary tau pathology (Braak Stages I and II) do not arise randomly across the cerebral mantle; they emerge with devastating anatomical specificity within the transentorhinal and entorhinal cortices, precisely where the grid-cell metric system resides. As tau pathology progresses into the hippocampus proper (CA1 and subiculum) during Stages III and IV, the neural machinery responsible for generating both metric coordinate frameworks and allocentric spatial maps undergoes profound structural degeneration.
Consequently, topological disorientation is not a late, secondary symptom of Alzheimer’s disease; it is frequently the earliest, most sensitive behavioral indicator of nascent neurodegeneration, emerging years before overt deficits in declarative verbal memory are detected on standard clinical batteries. Patients with preclinical AD and amnestic Mild Cognitive Impairment (aMCI) exhibit profound impairments in path integration, an inability to calculate spatial shortcuts across virtual environments, and severe topographical disorientation when navigating both unfamiliar and historically familiar physical environments. Concurrently, these patients suffer a catastrophic collapse in their capacity for episodic future simulation. When asked to visualize a realistic, personal event occurring in their near future, individuals in the early stages of Alzheimer’s disease produce fragmented, spatially impoverished descriptions devoid of contextual spatial cohesion. The breakdown of the entorhinal grid matrix simultaneously obliterates their internal map of the physical world and their bridge to the episodic future.
11.2 Hippocampal Amnesia and the Inability to Imagine the Future
The most compelling neuropsychological evidence directly demonstrating that the spatial cognitive map provides the indispensable architectural substrate for prospective mental simulation emerged from the systematic study of patients with dense, bilateral hippocampal amnesia. For half a century, the medical world viewed Patient H.M. (Henry Molaison)—who underwent bilateral medial temporal lobectomy in 1953 to treat intractable epilepsy—primarily through the lens of retrospective memory loss: he exhibited profound anterograde amnesia and a temporally graded retrograde amnesia.
It was not until the early 2000s that neuropsychologists posed a deceptively simple, profound diagnostic question to amnesic patients: what are you going to do tomorrow? In a seminal 2007 paper published in the Proceedings of the National Academy of Sciences, Demis Hassabis, Dharshan Kumaran, Seralynne Vann, and Eleanor Maguire investigated five patients with well-characterized, bilateral hippocampal damage who exhibited classical dense amnesia. The patients were presented with hypothetical, novel future scenarios that were not temporally bound—such as “Imagine you are lying on a white sandy beach in a tropical cove”—as well as specific prospective personal scenarios occurring in their episodic futures.
The experimental findings were definitive. While the hippocampal amnesic patients could easily retrieve isolated semantic facts (they understood intellectually that beaches contain sand, ocean water, and sunshine), they were completely, profoundly incapable of constructing a coherent, integrated mental simulation of the scene. When attempting to imagine the future, their descriptions were fragmented, broken into isolated, floating items: a palm tree here, a beach towel there. Crucially, the patients explicitly reported that they could not create a spatial frame or stage upon which the elements could reside; the mental image continuously disintegrated into nothingness. Patient H.M., when asked about his future, famously remarked that his internal state was like waking from a dream, describing his prospective world as “just a blank.” Without the hippocampal spatial mapping engine to assemble an allocentric spatial coordinate frame, the generation of the episodic future is rendered neurocomputationaly impossible.
11.3 Psychiatric Conditions and Maladaptive Prospective Simulation
The profound interdependence between the cognitive map and the episodic future is not confined to neurodegenerative or focal structural lesions; it constitutes a primary operational nexus across severe psychiatric and mood disorders. When the neural mechanisms that regulate prospective simulation undergo functional dysregulation, the cognitive map of subjective future spacetime becomes a landscape of clinical pathology.
In Major Depressive Disorder (MDD), this prospective capacity undergoes a devastating, characteristic truncation. Clinical and fMRI investigations demonstrate that depressed individuals exhibit what cognitive psychologists term “overgeneral prospective memory.” When prompted to imagine positive future scenarios, the prefrontal-hippocampal coupling fails to activate properly: the patient cannot simulate rich, detailed, spatially situated future scenes featuring positive outcomes. Instead, their episodic future horizon collapses into a bleak, impenetrable temporal wall, dominated by negative semantic abstractions: “Nothing will ever improve.” The dynamic, model-based capacity to simulate alternative rewarding futures and execute prospective spatial shortcuts around current life difficulties is paralyzed, driving profound feelings of hopelessness and clinical immobility.
In Post-Traumatic Stress Disorder (PTSD), the pathology represents a terrifying failure of the spatiotemporal cognitive map to properly bind and timestamp traumatic memories within their historical coordinate boundaries. During an intrusive flashback, the traumatic memory is not recalled as a past event situated safely at a distant, historical temporal coordinate; rather, the hippocampal contextual-indexing mechanism fails, and the traumatic trace completely overwhelms the present egocentric perspective. The patient literally experiences the historical horror as unfolding *here and now* within their current physical coordinate space. Concurrently, in schizophrenia, structural and functional microcircuit abnormalities within the hippocampal CA3-CA1 recurrent collateral networks disrupt both spatial path integration and forward predictive coding. The internal forward model of the world—the model-based engine that predicts the sensory and spatial consequences of future actions—decouples from physical reality, spawning spatial disorientation, fragmented relational maps, and the hallucinatory intrusion of aberrant prospective simulations into waking consciousness.
12. Epistemological Synthesis: Tolman’s Enduring Paradigm for Modern Cognitive Science
12.1 The Resolution of the Hull-Tolman Debate in Modern Neuroscience
Viewed through the perspective of twenty-first-century cognitive neuroscience, the historic intellectual war waged between Clark Hull’s stimulus-response habit behaviorism and Edward Tolman’s purposive cognitive mapping does not conclude with the absolute triumph of one paradigm over the other. Rather, modern neuroscience has achieved an elegant, dialectical synthesis that reconciles both perspectives within a dual-system neurocomputational architecture.
Contemporary neuroscience recognizes that the mammalian nervous system evolved two distinct, complementary navigational and behavioral control systems that operate in continuous, dynamic tension:
- The Habitual, Model-Free System (The Hullian Engine): Anatomically localized primarily within the dorsolateral striatum and the sensorimotor basal ganglia, this system learns slowly via dopamine-mediated temporal-difference prediction errors. It automates repetitive motor routines, encoding rigid stimulus-response reflexes that execute rapidly and with minimal metabolic expenditure.
- The Cognitive, Model-Based System (The Tolmanian Engine): Localized within the hippocampal formation, the medial entorhinal cortex, and the prefrontal cortex, this system constructs, maintains, and navigates flexible, internal cognitive maps of the environment, computing novel shortcuts, detours, and prospective episodic simulations at high energetic cost.
The resolution of the Hull-Tolman debate lies in the computational arbitration mechanisms that determine which system commands behavior at any given moment. Seminal computational models developed by Nathaniel Daw, Yael Niv, and Peter Dayan demonstrate that the brain employs an internal, Bayesian uncertainty arbiter—likely mediated by the anterior cingulate and prefrontal cortices—that dynamically calculates the subjective uncertainty of both systems. When an environment is novel, volatile, or structurally disrupted, the habitual Hullian striatal system exhibits massive uncertainty, prompting the arbiter to hand executive control to the Tolmanian model-based hippocampal system. Conversely, when an animal traverses a highly stable, invariant environment over hundreds of trials, the Hullian striatal system masters the cached motor habits, and the brain smoothly shifts control from the expensive, deliberate cognitive map to the rapid, automated stimulus-response habit. Tolman was correct in his core epistemological assertion: learning is inherently an active, representational construction; habit is merely an optimized, downstream compression of deeply consolidated experience.
12.2 Artificial General Intelligence and Representational World Models
As artificial intelligence strives to bridge the immense chasm separating narrow, specialized systems from true Artificial General Intelligence (AGI), the field has run squarely into the fundamental computational limitations that doomed Watsonian and Hullian behaviorism a century ago. Contemporary deep reinforcement learning systems that rely exclusively on model-free architectures—such as deep Q-networks—can achieve superhuman performance in fixed, static environments like Atari video games or chess; however, they require millions of trials to learn, exhibit brittle failure modes when confronted with minor environmental perturbations, and fail completely when asked to generalize knowledge to novel contexts without extensive retraining.
To overcome this limitation, the vanguard of AGI research has executed a decisive pivot toward Tolmanian principles, focusing intensely on the engineering of “World Models.” Pioneered by researchers such as David Ha, Jürgen Schmidhuber, and Yann LeCun, world-model architectures explicitly construct an internal, generative model of the task environment. An artificial agent equipped with a world model does not merely react to visual pixel inputs; it continuously trains a compact, latent spatial representation of the world, learning the predictive transition dynamics that govern how actions transform states over time.
Systems such as DeepMind’s MuZero represent the structural computational realization of Tolmanian cognitive mapping within deep neural networks. MuZero constructs an internal latent model of the world that predicts three critical quantities: the value of a current state, the reward of an action, and the prospective transition to the next latent state. When deciding upon an action, the agent executes an internal Monte Carlo Tree Search across its own generative latent model—simulating thousands of prospective future trajectories entirely in its “imagination” before executing a single physical move in the real environment. By coupling these generative world models with episodic memory buffers that store past trajectory vectors, artificial agents achieve zero-shot and few-shot learning, planning shortcuts through unvisited state spaces precisely as Tolman’s rodents did in the Sunburst maze of 1946. The road to artificial general intelligence is paved directly with the conceptual materials of Tolmanian cognitive mapping.
12.3 Towards an Integrated Science of Spatial and Prospective Cognition
A century after Edward Chace Tolman began his quiet rebellion in the basement laboratories of the University of California, Berkeley, his intellectual legacy stands as one of the most transformative in the history of cognitive science. Tolman took what the scientific establishment of his era dismissed as untestable, unscientific metaphysics—purpose, expectancy, internal representation, cognitive deliberation—and operationalized these phenomena with exquisite, unyielding empirical rigor. In doing so, he erected the theoretical foundations that would eventually bloom into modern cognitive science, neurobiology, and computational intelligence.
Today, the arbitrary disciplinary boundaries that once separated the study of animal spatial navigation, human declarative memory, artificial intelligence, and temporal forecasting have dissolved. They are revealed to be unified facets of a single, deeply conserved biological imperative: the construction and navigation of internal representational models. The profound insight of the modern synthesis is that our grandest, most uniquely human cognitive faculties—our capacity to write alternative histories, to dream of distant futures, to formulate scientific theories, to weigh moral counterfactuals, and to mentally journey across unobserved temporal horizons—are not mysterious, ethereal powers detached from our biological origins.
Rather, these vast temporal capacities represent the evolutionary expansion of an ancient navigational engine. The mammalian brain took the spatial coordinate matrices that first allowed a primitive creature to calculate a physical shortcut through an ancient forest, scaled them across higher-order associative cortices, and directed them inward upon the landscape of subjective time and abstract thought. We are, at our biological core, perpetual navigators. Whether we are maneuvering a vehicle through a bustling modern city, tracing an analogy across abstract conceptual manifolds, or closing our eyes to simulate the unobserved terrain of our own episodic future, we are traveling across the boundless, dynamic topography of Edward Tolman’s cognitive map.
References
- Addis, D. R., Wong, A. T., & Schacter, D. L. (2007). Remembering the past and imagining the future: Common and distinct neural substrates during event construction and elaboration. Neuropsychologia, 45(7), 1363–1377. https://doi.org/10.1016/j.neuropsychologia.2006.10.016
- Balleine, B. W., & Dickinson, A. (1998). Goal-directed instrumental action: contingency and incentive learning and their cortical substrates. Neuropharmacology, 37(4-5), 407–419. https://doi.org/10.1016/s0028-3908(98)00033-1
- Bellmund, J. L., Gärdenfors, P., Moser, E. I., & Doeller, C. F. (2018). Navigating cognition: Spatial codes for human thinking. Science, 362(6415), eaat6766. https://doi.org/10.1126/science.aat6766
- Buckner, R. L., & Carroll, D. C. (2007). Self-projection and the brain. Trends in Cognitive Sciences, 11(2), 49–57. https://doi.org/10.1016/j.tics.2006.11.004
- Clayton, N. S., & Dickinson, A. (1998). Episodic-like memory during cache recovery by scrub jays. Nature, 395(6699), 272–274. https://doi.org/10.1038/26228
- Constantinescu, A. O., O’Reilly, J. X., & Behrens, T. E. (2016). Organizing conceptual knowledge in humans with a gridlike code. Science, 352(6292), 1464–1468. https://doi.org/10.1126/science.aaf0941
- Daw, N. D., Niv, Y., & Dayan, P. (2005). Uncertainty-based competition between prefrontal and dorsolateral striatal systems for behavioral control. Nature Neuroscience, 8(12), 1704–1711. https://doi.org/10.1038/nn1560
- Dayan, P. (1993). Improving generalization for temporal difference learning: The successor representation. Neural Computation, 5(4), 613–624. https://doi.org/10.1162/neco.1993.5.4.613
- Hassabis, D., Kumaran, D., Vann, S. D., & Maguire, E. A. (2007). Patients with hippocampal amnesia cannot imagine new experiences. Proceedings of the National Academy of Sciences, 104(5), 1726–1731. https://doi.org/10.1073/pnas.0610561104
- Hull, C. L. (1943). Principles of Behavior: An Introduction to Behavior Theory. Appleton-Century-Crofts.
- Janmaat, K. R., Polansky, L., Ban, S. D., & Boesch, C. (2014). Wild chimpanzees use tree species analysis to formulate cognitive maps and plan future foraging routes. Proceedings of the National Academy of Sciences, 111(17), 6347–6352. https://doi.org/10.1073/pnas.1323539111
- Johnson, A., & Redish, A. D. (2007). Neural ensembles in CA3 transiently encode paths forward of the animal at a decision point. Journal of Neuroscience, 27(45), 12176–12188. https://doi.org/10.1523/JNEUROSCI.3761-07.2007
- MacDonald, C. J., Lepage, K. Q., Eden, U. T., & Eichenbaum, H. (2011). Hippocampal “time cells” bridge the gap in memory for temporally separated events. Neuron, 71(4), 737–749. https://doi.org/10.1016/j.neuron.2011.07.012
- Moser, E. I., Kropff, E., & Moser, M. B. (2008). Place cells, grid cells, and the brain’s spatial representation system. Annual Review of Neuroscience, 31, 69–89. https://doi.org/10.1146/annurev.neuro.31.061307.090723
- O’Keefe, J., & Dostrovsky, J. (1971). The hippocampus as a spatial map: Preliminary evidence from unit activity in the freely-moving rat. Brain Research, 34(1), 171–175. https://doi.org/10.1016/0006-8993(71)90358-1
- O’Keefe, J., & Nadel, L. (1978). The Hippocampus as a Cognitive Map. Oxford University Press.
- Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, D. A., & Shulman, G. L. (2001). A default mode of brain function. Proceedings of the National Academy of Sciences, 98(2), 676–682. https://doi.org/10.1073/pnas.98.2.676
- Schacter, D. L., & Addis, D. R. (2007). The cognitive neuroscience of constructive memory: Remembering the past and imagining the future. Philosophical Transactions of the Royal Society B: Biological Sciences, 362(1481), 773–786. https://doi.org/10.1098/rstb.2007.2087
- Stachenfeld, K. L., Botvinick, M. M., & Gershman, S. J. (2017). The hippocampus as a predictive map. Nature Neuroscience, 20(11), 1643–1653. https://doi.org/10.1038/nn.4650
- Taube, J. S. (2007). The head direction signal: Origins and sensory-motor processing. Annual Review of Neuroscience, 30, 181–207. https://doi.org/10.1146/annurev.neuro.29.051605.112842
- Tavares, R. M., Mendelsohn, A., Grossman, Y., Williams, C. H., Shapiro, M., Trope, Y., & Schiller, D. (2015). A map for social navigation in the human brain. Neuron, 87(1), 231–243. https://doi.org/10.1016/j.neuron.2015.06.011
- Thorndike, E. L. (1898). Animal intelligence: An experimental study of the associative processes in animals. The Psychological Review: Monograph Supplements, 2(4), i–109. https://doi.org/10.1037/h0092987
- Tolman, E. C. (1932). Purposive Behavior in Animals and Men. Century Company.
- Tolman, E. C. (1948). Cognitive maps in rats and men. Psychological Review, 55(4), 189–208. https://doi.org/10.1037/h0061626
- Tolman, E. C., & Honzik, C. H. (1930). Introduction and removal of reward, and maze performance in rats. University of California Publications in Psychology, 4(19), 257–275.
- Tolman, E. C., Ritchie, B. F., & Kalish, D. (1946). Studies in spatial learning. II. Place learning versus response learning. Journal of Experimental Psychology, 36(3), 221–229. https://doi.org/10.1037/h0060262
- Watson, J. B. (1913). Psychology as the behaviorist views it. Psychological Review, 20(2), 158–177. https://doi.org/10.1037/h0074428