Hippocampus

Abstract

The hippocampus is a curved, bilateral structure in the medial temporal lobe and part of the limbic system, named for its resemblance to a seahorse. It is essential for forming new episodic and declarative memories, as the dense amnesia that follows its bilateral removal shows, even though such damage spares short-term retention and skill learning. It also carries a map of space: place cells fire at particular locations and, with entorhinal grid cells, compose the brain's navigation system. Its trisynaptic circuit, running from entorhinal cortex to dentate gyrus, CA3, and CA1, supports rapid encoding, pattern separation of similar experiences, and the long-term potentiation that stores them. This article sets out its anatomy, its memory and spatial functions, and the mechanisms that link them.

Keywords: episodic memory, spatial navigation, synaptic plasticity

Few structures in the brain have carried as much explanatory weight as the hippocampus. A single patient whose hippocampi were removed to treat epilepsy showed the world that the capacity to form new conscious memories is not diffusely spread through the cortex but depends on this one region (Scoville & Milner, 1957). A recording from a rat exploring a box showed that the same structure holds a map of where the animal is (O'Keefe & Dostrovsky, 1971). And a train of stimulation to one of its pathways left its synapses strengthened for hours, giving the first cellular model of how an experience might be stored (Bliss & Lomo, 1973). This article follows those three threads — memory, space, and plasticity — through the anatomy that ties them together.

What the Hippocampus Is

The hippocampus is a paired structure folded into the medial wall of each temporal lobe, one in each hemisphere, taking its name from the Greek for seahorse because a cross-section curls like one. In the standard anatomical scheme it belongs to the archicortex, a phylogenetically old three-layered cortex, and forms a central node of the limbic system alongside the amygdala, the fornix, and the mammillary bodies. Its gross form is an interlocking pair of C-shaped sheets — the cornu ammonis, divided into fields CA1 through CA3, and the dentate gyrus — rolled together and connected to the rest of the brain chiefly through the adjacent entorhinal cortex and, as an output, through the fornix.

What makes the structure a subject for cognitive psychology rather than only for anatomy is that this small piece of tissue turns out to be necessary for a specific and momentous mental function: laying down new memories of facts and events. The rest of this article treats the hippocampus through four questions that a century of research has made central — what it does for memory, how it represents space, how its synapses change, and how its input stage keeps similar experiences apart. Figure 1 lays out the internal wiring those questions share, and Table 1 previews the four roles and the classic evidence for each.

Types of Hippocampus

The National Library of Medicine's Medical Subject Headings files the hippocampus at two positions in its anatomical tree, under the limbic system and under the cerebral cortex, and hangs beneath the descriptor a set of narrower terms for its component subfields and pathways. Unlike the subtypes of a functional category, these narrower descriptors are not kinds of hippocampus but parts of one: the classification here decomposes the structure anatomically, naming the fields through which a signal passes and the tracts that carry it in and out. Table 2 lists the direct children of the descriptor in the MeSH tree; each is glossed from its role in the circuit, and none is yet a standalone article on this site.

Two cautions keep this list in proportion. It is a classification for indexing, built to organise the biomedical literature, not a theory of the mind's natural divisions, and its two parent placements — limbic system and cerebral cortex — reflect competing anatomical conventions rather than a contradiction. And because these are parts rather than kinds, they are complementary: a memory is not stored in CA3 or the dentate gyrus but in the coordinated activity that runs through all of them, which the trisynaptic circuit below traces in order.

A signal enters the hippocampus from the entorhinal cortex and passes through three synapses in turn. Step it forward to follow the pathway; each highlighted arrow names the synapse it crosses.

Memory and Amnesia

The modern understanding of the hippocampus begins with a surgical mistake. In 1953 the neurosurgeon William Scoville removed the medial temporal lobes on both sides of a young man, Henry Molaison, known for decades in the literature only as H.M., to relieve intractable epilepsy. The seizures eased, but H.M. was left unable to form new long-term memories of facts and events: he could hold a conversation, yet minutes later retain nothing of it, and he never learned the layout of the house he moved to or recognised the people who saw him daily for years. Brenda Milner's careful testing established the crucial dissociations. H.M.'s intelligence, language, and perception were intact; his memory for events well before the surgery was largely preserved; and his short-term retention over seconds was normal. What was lost was the conversion of momentary experience into durable memory, and the loss was specific to the medial temporal lobe, above all the hippocampus (Scoville & Milner, 1957).

A second dissociation deepened the picture. Milner found that H.M. could learn a new perceptual-motor skill — tracing a shape seen only in a mirror — improving day by day even though he had no memory of ever having done the task before. Memory, then, is not one faculty but several: the hippocampus is required for declarative memory, the conscious memory of facts and events that can be brought to mind and stated, but not for nondeclarative skill learning, which proceeds through other systems. Larry Squire drew this division into a systematic taxonomy and argued that the hippocampus and neighbouring structures form a medial temporal lobe memory system whose job is to bind the disparate elements of an experience into a single retrievable trace (Squire, 1992; Squire, Stark, & Clark, 2004). Howard Eichenbaum sharpened the account by proposing that the common thread is relational memory: the hippocampus specialises in representing the relations among items — this face in that place at this time — rather than the items themselves, which is why its loss so devastates memory for events, each of which is a bound set of relations (Eichenbaum, 2004).

The same evidence fixes the hippocampus's role in time as well as in kind. Because H.M.'s old memories survived while new ones could not form, the structure cannot be the permanent store; instead it is required to establish a memory and to hold it during a period of consolidation, after which the memory comes to depend on the neocortex and can survive hippocampal damage. This is the standard model of systems consolidation, and it explains the characteristic temporally graded amnesia — recent memories lost, remote ones spared — that follows hippocampal injury (Squire, 1992).

Place Cells and the Cognitive Map

While clinical work was defining the memory role, a parallel line in the rat was uncovering something that at first seemed unrelated. Recording from single hippocampal neurons in freely moving animals, John O'Keefe and Jonathan Dostrovsky found cells that fired only when the animal was in a particular part of its environment and fell silent elsewhere. Each such place cell has a preferred region, its place field, and across a population the fields tile the space, so that the animal's location can in principle be read from which cells are active (O'Keefe & Dostrovsky, 1971). O'Keefe drew the bold conclusion that the hippocampus implements a cognitive map, an allocentric representation of the environment that supports flexible navigation independent of any particular route.

The map is not built by the hippocampus alone. Upstream, in the medial entorhinal cortex, Edvard and May-Britt Moser and their colleagues discovered grid cells, neurons whose multiple firing fields form a regular triangular lattice covering the whole environment, as though the brain had laid down graph paper over the world. Grid cells provide a metric — a sense of distance and direction — that can update the animal's estimated position from its own movements, and their output, together with information about boundaries and landmarks, is thought to help generate the place fields seen one synapse downstream (Hafting, Fyhn, Molden, Moser, & Moser, 2005). Together the entorhinal grid system and the hippocampal place system constitute the brain's core spatial representation (Moser, Kropff, & Moser, 2008), a discovery recognised with the 2014 Nobel Prize in Physiology or Medicine.

That the spatial and mnemonic roles belong to one structure is not a coincidence but a clue. Eleanor Maguire's studies of London taxi drivers, who must memorise the city's street layout in exhaustive detail, found their posterior hippocampi to be enlarged relative to controls, and the more so the longer they had driven — direct evidence that intensive spatial learning reshapes the structure (Maguire et al., 2000). The place code is not a separate faculty bolted onto the memory system; it is the spatial face of a general capacity to bind the elements of experience into an organised representation.

Space is not the only dimension the hippocampus maps. Recording while animals bridged an empty delay between two events, Eichenbaum and colleagues found time cells: neurons that fire at successive moments of a structured interval, so that across the population elapsed time is marked much as place cells mark location, tiling a remembered episode's temporal extent rather than its spatial one (Eichenbaum, 2014). Time cells give the when of an event a neural code alongside the where, and they reinforce the relational reading of the structure: the hippocampus lays down an organised representation whose axes can be spatial, temporal, or—as the newest work argues—more abstract still.

The demonstration below shows how a population of place cells encodes position and how that position can be recovered by combining their firing rates.

Four place cells have fields centred at 20, 40, 60, and 80 cm along a one-metre track. Move the animal (navy line): each cell fires by how close the animal is to its centre, and the population vector (gold) decodes position by averaging the centres weighted by firing rate.

Synaptic Plasticity and Storage

For the hippocampus to store an experience, something in it must change and stay changed. In 1973 Timothy Bliss and Terje Lomo found that change and gave it a name. Stimulating the perforant path — the fibres running from entorhinal cortex into the dentate gyrus — with a brief high-frequency train left the response of the target cells enlarged, and the enlargement lasted for hours and, in later work, days. This long-term potentiation (LTP) was a use-dependent, long-lasting increase in synaptic strength, exactly the sort of durable, activity-triggered change a memory mechanism would need (Bliss & Lomo, 1973).

LTP became the dominant cellular model of memory because its properties mirror the logic of association. It is input-specific, strengthening only the synapses that were active, so different memories can be stored on different synapses of the same cell. It is associative and cooperative, induced when pre- and postsynaptic activity coincide, which realises at the level of a synapse the old idea that cells that fire together wire together. The strongest expression of LTP, at the Schaffer collateral synapses onto CA1, depends on the NMDA receptor, a channel that opens only when the postsynaptic cell is already depolarised while glutamate is bound, making it a molecular coincidence detector for exactly the near-simultaneous activity that binding an event demands. That the same receptor is needed not merely for the plasticity but for memory itself was shown directly when blocking hippocampal NMDA receptors with the antagonist AP5 impaired rats' spatial learning while abolishing LTP, tying the synaptic mechanism to behaviour rather than leaving the connection a matter of analogy (Morris, Anderson, Lynch, & Baudry, 1986). The result is a plausible physical substrate for the relational binding the memory system performs: the co-activation of the elements of an experience potentiates the synapses that link their representations, so that later reactivating some elements can retrieve the rest.

Pattern Separation and Completion

A memory system faces two opposing demands. It must keep similar experiences apart, so that this morning's parking spot is not confused with yesterday's, and it must also fill in a whole memory from a fragment, so that a snatch of melody recovers the song. The hippocampal circuit is thought to meet the two demands with two different subfields. The dentate gyrus performs pattern separation: its very large population of granule cells, of which only a sparse few are active for any input, recodes overlapping entorhinal inputs into more distinct, less overlapping representations, so that similar events are stored on largely non-overlapping sets of cells. The recurrently connected CA3 field performs the complementary operation of pattern completion: its dense web of feedback connections lets a partial input settle, through the network's own dynamics, into the complete stored pattern, so that a cue can retrieve the whole (Leal & Yassa, 2018).

The two operations trade off, and the balance matters for behaviour. Too little separation and memories blur together; too little completion and a cue fails to retrieve what it should. Human neuroimaging has made the trade-off measurable with the mnemonic similarity task, in which people see objects and later judge whether test items are identical, similar lures, or new: correctly calling a lure similar rather than old requires pattern separation, and performance on this discrimination tracks dentate and CA3 activity and declines in ageing and disease (Leal & Yassa, 2018). The demonstration below makes the geometry concrete: it shows how a sparse expansion recodes two overlapping input patterns into outputs that share far fewer active units.

Two input patterns (8 entorhinal units, 4 active each) are recoded into 28 conjunction units, each firing only when both of its two inputs are active. Slide to set how many active inputs the patterns share; gold marks units active in both. Because a shared conjunction needs a shared pair, the sparse output overlap falls below the input overlap.

Worked Example

The place-cell population code can be read out by hand, and the second demonstration reproduces the calculation. Suppose four place cells lie along a one-metre linear track, with place-field centres at 20, 40, 60, and 80 centimetres. Each cell's firing rate falls off as a Gaussian of the distance between the animal's position and the cell's centre, with a width (standard deviation) of 15 centimetres, so a cell fires most strongly when the animal is at its centre and progressively less as the animal moves away. The animal is in fact at 45 centimetres, and we ask what position the population implies.

Each cell's rate is the Gaussian evaluated at its own distance from 45 centimetres. The cell centred at 40 is only 5 centimetres away and fires near its maximum, at a relative rate of 0.95; the cell at 60 is 15 centimetres away and fires at 0.61; the cell at 20 is 25 centimetres away and fires at 0.25; the cell at 80 is 35 centimetres away and barely responds, at 0.07. To decode position we take the population vector — the average of the field centres weighted by how hard each cell is firing:

decoded position = (0.25 x 20 + 0.95 x 40 + 0.61 x 60 + 0.07 x 80) / (0.25 + 0.95 + 0.61 + 0.07)

The numerator is 5.0 + 38.0 + 36.6 + 5.6 = 84.5 (using the unrounded rates, 84.48), and the denominator, the total activity, is 1.88. Their ratio is 45.2 centimetres — within a fifth of a centimetre of the animal's true position of 45. The population has recovered the location that no single cell encodes exactly, because the weighted average pools the partial evidence each cell provides. The small residual, the decode landing at 45.2 rather than 45, is not noise but a slight pull toward the centre of the track: the cell at 20 has a little more of its tail inside the track than the symmetric cell at 80, so the estimate is drawn marginally inward. Placed exactly at the track's centre, 50 centimetres, the same four cells decode to 50 with no bias at all, because there the tails are symmetric.

Discussion

The hippocampus is the rare case where a structure defined by anatomy turns out to correspond to a natural joint of the mind. Three lines of evidence that began apart — the amnesia of a surgical patient, the firing of a cell in a moving rat, the strengthening of a synapse after a train of shocks — have converged on a single account in which memory, space, and plasticity are three views of one capacity. The capacity is relational binding: tying together the elements of an experience, whether the who-what-where of an event or the landmarks and distances of a place, into an organised representation that can later be retrieved from a fragment (Eichenbaum, 2004). Long-term potentiation supplies the synaptic mechanism, the dentate gyrus and CA3 supply the separating and completing operations, and the whole is held only temporarily before consolidation hands the memory to the cortex (Bliss & Lomo, 1973; Squire, 1992).

What remains genuinely unsettled is how to state the function at the right level of generality. Is the hippocampus fundamentally a memory device that also happens to map space, or a spatial mapper whose machinery was co-opted for memory, or is cognitive map too narrow a word for what is really a general-purpose system for representing the relational structure of any experience? A synthesis by several leading investigators laid the competing emphases side by side without dissolving them, a sign that the field still holds more than one organising idea at once (Lisman et al., 2017). The tension is productive rather than embarrassing: each framing predicts phenomena the others miss, and the structure is evidently doing something abstract enough that memory and navigation are both special cases of it.

Current Directions

The most active current idea generalises the cognitive map beyond physical space. If the hippocampal-entorhinal system can lay a metric over a room, perhaps it can lay one over any structured domain — a family tree, a set of social relationships, a space of concepts — representing knowledge of all kinds in a common map-like format that supports the same flexible inference navigation requires. On this view place cells and grid cells are special cases of a general machinery for organising knowledge for flexible behaviour, and the same coding principles should appear when people reason over non-spatial relational structures (Behrens et al., 2018). The proposal reframes the old memory-versus-space debate: both become instances of building and reading a relational map.

A second front concerns pattern separation in the clinic. Because the dentate gyrus and CA3 are among the earliest sites affected in ageing and in Alzheimer's disease, the mnemonic similarity task and high-resolution imaging of hippocampal subfields are being developed as sensitive early markers of memory decline, and the separation-completion balance is a target for understanding why some memory failures take the form of confusing similar experiences rather than losing them outright (Leal & Yassa, 2018). Both directions share a method: treating the hippocampus not as a black box for memory but as a circuit whose specific computations can be measured, modelled, and, increasingly, read out.

Common Misconceptions

Memories are stored permanently in the hippocampus.

The hippocampus is required to form new declarative memories and to hold them during consolidation, but old memories survive its destruction, which is why hippocampal amnesia is temporally graded. The long-term store is distributed in the neocortex (Squire, 1992).

Hippocampal damage wipes out all memory.

It selectively impairs declarative memory for new facts and events while sparing short-term retention, remote memory, and nondeclarative skill learning, as patient H.M. demonstrated by learning a mirror-drawing skill he could not remember practising (Scoville & Milner, 1957).

The hippocampus is only a spatial map.

Spatial coding and memory are two faces of one function. The same structure that maps space binds the relations among the elements of any event, and current work extends its map-like coding to non-spatial conceptual structure (Eichenbaum, 2004; Behrens et al., 2018).

Glossary

CA3.

The recurrently connected hippocampal subfield whose autoassociative network supports rapid encoding and pattern completion of stored memories from partial cues.

Cognitive map.

An allocentric internal representation of an environment that supports flexible navigation independent of any specific route; proposed by O'Keefe as the hippocampus's function.

Declarative memory.

The conscious memory for facts and events that can be brought to mind and stated; the form of memory that depends on the hippocampus.

Dentate gyrus.

The input subfield of the hippocampus whose sparse granule-cell code performs pattern separation, and one of the few brain regions to generate new neurons in adulthood.

Entorhinal cortex.

The cortical region that is the main gateway between the hippocampus and the rest of the cortex, and the site of grid cells.

Grid cell.

An entorhinal neuron whose multiple firing fields form a regular triangular lattice over the environment, providing a metric for distance and direction.

Long-term potentiation.

A long-lasting, use-dependent increase in synaptic strength following brief high-frequency stimulation; the leading cellular model of memory storage.

Medial temporal lobe.

The region comprising the hippocampus and adjacent cortices whose bilateral damage produces profound anterograde amnesia.

Nondeclarative memory.

Memory expressed through performance rather than conscious recollection, such as skills and habits; it does not depend on the hippocampus.

Pattern completion.

The retrieval of a whole stored pattern from a partial or degraded cue, attributed to the recurrent CA3 network.

Pattern separation.

The recoding of overlapping inputs into more distinct representations so that similar experiences are stored separately; attributed to the dentate gyrus.

Place cell.

A hippocampal neuron that fires selectively when the animal occupies a particular region of its environment, the cell's place field.

Population vector.

A decoding method that estimates a represented value, such as position, as the average of many neurons' preferred values weighted by their firing rates.

Schaffer collaterals.

The axon collaterals projecting from CA3 to CA1, whose synapses are the classic experimental site for studying long-term potentiation.

Systems consolidation.

The process by which a memory initially dependent on the hippocampus gradually comes to be stored in and retrieved from the neocortex.

Time cell.

A hippocampal neuron that fires at a particular moment within a structured interval, so that a population of such cells encodes elapsed time much as place cells encode location.

Trisynaptic circuit.

The principal internal pathway of the hippocampus: entorhinal cortex to dentate gyrus, dentate gyrus to CA3, and CA3 to CA1.

Key Researchers

Howard Eichenbaum (1947-2017). Director of the Center for Memory and Brain at Boston University; he argued that the hippocampus supports relational and declarative memory beyond spatial mapping and discovered hippocampal time cells that encode the temporal structure of experience. Faculty Page - Google Scholar - Wikipedia - Wikidata - ORCID

Eleanor A. Maguire (1970-2025). Professor at the Wellcome Centre for Human Neuroimaging, University College London; her London taxi-driver studies demonstrated experience-dependent hippocampal plasticity and showed the hippocampus supports imagining future and fictional scenes. Wikipedia - Wikidata - ORCID

Brenda Milner (b. 1918). Professor at the Montreal Neurological Institute, McGill University; her study of patient H.M. established that the hippocampus is essential for forming new declarative memories while sparing skill learning, founding the cognitive neuroscience of memory. Faculty Page - Wikipedia - Wikidata

Edvard I. Moser. Director of the Kavli Institute for Systems Neuroscience at the Norwegian University of Science and Technology; he co-discovered grid cells in the entorhinal cortex, revealing how the entorhinal-hippocampal circuit builds a metric map for navigation and memory. Faculty Page - Google Scholar - Wikipedia - Wikidata - ORCID

May-Britt Moser. Professor and founding co-director of the Kavli Institute for Systems Neuroscience at the Norwegian University of Science and Technology; she co-discovered grid cells, revealing the spatial coordinate system that works with the hippocampus to support navigation and spatial memory. Faculty Page - Google Scholar - Wikipedia - Wikidata - ORCID

John O'Keefe. Professor at the Sainsbury Wellcome Centre, University College London; he discovered hippocampal place cells in 1971, establishing the hippocampus as the brain's spatial map and earning a share of the 2014 Nobel Prize in Physiology or Medicine. Faculty Page - Wikipedia - Wikidata - ORCID

Larry R. Squire. Professor at the University of California San Diego and the VA San Diego Healthcare System; he identified the medial temporal lobe memory system and the declarative-nondeclarative distinction, establishing the standard model of systems-level memory consolidation. Faculty Page - Google Scholar - Wikipedia - Wikidata - ORCID

Frequently Asked Questions

What does the hippocampus do?
The hippocampus is essential for forming new memories of facts and events and for representing space through place cells. It binds the elements of an experience into a memory that can later be retrieved, and it holds that memory during consolidation before the neocortex takes over (Squire, 1992).

What happens if the hippocampus is damaged?
Bilateral hippocampal damage causes anterograde amnesia, the inability to form new declarative memories, as in the patient H.M. Short-term retention, memory for the distant past, and skill learning are relatively spared, showing that memory is made of several systems (Scoville & Milner, 1957).

What are place cells?
Place cells are hippocampal neurons that fire only when an animal is in a particular location, the cell's place field. Across a population the fields cover the environment, so the animal's position can be decoded from which cells are active (O'Keefe & Dostrovsky, 1971).

How are place cells related to grid cells?
Grid cells sit one synapse upstream in the entorhinal cortex and fire in a repeating triangular lattice across space, providing a metric for distance and direction. Their output helps generate hippocampal place fields, and together the two systems form the brain's spatial map (Hafting et al., 2005; Moser et al., 2008).

What is long-term potentiation?
Long-term potentiation is a lasting increase in the strength of a synapse after brief high-frequency activity, first shown in the hippocampus. Because it is durable, input-specific, and triggered by coincident activity, it is the leading cellular mechanism for memory storage (Bliss & Lomo, 1973).

What is the difference between pattern separation and pattern completion?
Pattern separation, attributed to the dentate gyrus, recodes overlapping inputs into more distinct representations so similar events do not blur. Pattern completion, attributed to CA3, restores a whole memory from a partial cue. The two operations balance each other (Leal & Yassa, 2018).

Does the hippocampus store memories permanently?
No. It is needed to form and temporarily hold declarative memories, but over time consolidation transfers them to the neocortex, where they survive hippocampal damage. This is why recent memories are lost after hippocampal injury while remote ones remain (Squire, 1992).

Is the hippocampus only about memory and space?
Current work suggests its map-like coding extends to non-spatial relational structure, such as social or conceptual spaces, casting memory and navigation as two instances of a general system for organising knowledge for flexible behaviour (Behrens et al., 2018).

References