Key facts
- What it is
- A long-lasting increase in synaptic strength after brief high-frequency stimulation
- First reported
- Bliss and Lomo, 1973, perforant path to dentate gyrus in the anaesthetised rabbit
- Transmitter
- Glutamate, acting on AMPA and NMDA receptors
- The trigger
- Calcium entering through the NMDA receptor once its magnesium block is relieved
- The expression
- More AMPA receptors in the postsynaptic membrane, and more conductive ones
- Two phases
- Early phase uses existing proteins; late phase requires new gene expression and protein synthesis
- The mirror process
- Long-term depression, which removes AMPA receptors and prevents saturation
- Status
- The leading cellular model of memory, and still a model rather than a proven mechanism
The problem LTP was invented to solve
Learning changes behaviour durably, and the brain is made of cells. Somewhere in the tissue there must be a physical change that outlasts the experience which produced it. That is not a philosophical point; it is a demand for a mechanism, and for most of the twentieth century nobody could produce one.
Donald Hebb sharpened the demand into a testable proposal in The Organization of Behavior in 1949. His postulate was that when an axon of one cell repeatedly or persistently takes part in firing another, some growth process or metabolic change occurs such that the first cell's efficiency in firing the second is increased. Hebb had no way to observe a synapse changing. He was writing a specification, not reporting a result.
The specification was demanding. Whatever mechanism satisfied it would have to be fast, because a single experience can produce a lasting memory. It would have to be durable, because memories last. It would have to be specific, because a memory has to be stored in some synapses and not in all of a cell's synapses at once. And it would have to be associative, because the whole point of a memory is that it links things that occurred together.
Long-term potentiation is, so far, the only well-characterised cellular phenomenon that meets every one of those requirements at once. That is why it dominates the field, and that is also why it is worth being careful about what it does and does not establish.
Long-term potentiation: a long-lasting increase in the efficacy of synaptic transmission, produced by a brief period of high-frequency stimulation of the presynaptic pathway. Operationally, the postsynaptic response to a standard test stimulus is measurably larger after induction than before, and stays larger.
1973: what Bliss and Lomo actually saw
The founding experiment is worth stating precisely, because it is usually described loosely.
Tim Bliss and Terje Lomo were working in Per Andersen's laboratory in Oslo, recording from anaesthetised rabbits. They placed a stimulating electrode on the perforant path, the great fibre bundle that carries input from the entorhinal cortex into the hippocampus, and a recording electrode in the dentate gyrus, where those fibres terminate. They delivered a standard test shock every few seconds and measured the size of the response the dentate granule cells produced. That gave them a stable baseline.
Then they delivered a brief train of high-frequency stimulation to the same pathway. Afterwards, they went back to the same test shock they had been using all along, and the response was bigger. Not for a few seconds, and not for a few minutes. It was still bigger hours later, and in some animals for days. The pathway had not been damaged, the test stimulus had not been changed, and nothing about the recording had been altered. The synapse itself had become more effective, and it had stayed that way.
The paper, published in the Journal of Physiology in 1973, is the origin of the field. What made it so consequential was not the phenomenon in isolation but the coincidence of three facts: it was in the hippocampus, a structure whose necessity for forming new memories had been made unavoidable by the case of the patient H.M.; it was induced in seconds and persisted for hours; and it was, on inspection, exactly the kind of thing Hebb had specified. A theoretical postulate from 1949 and an electrophysiological result from 1973 fitted together, and the modern neurobiology of memory begins at that junction.
A note on where LTP is usually studied today. Bliss and Lomo worked on the perforant path input to the dentate gyrus. Most modern work is done on a different hippocampal synapse: the Schaffer collaterals, the axons that run from CA3 pyramidal cells to CA1 pyramidal cells, generally in a slice of hippocampus kept alive in a dish. This synapse is convenient, orderly, and extremely well characterised, and it is the one described below. It is not the only place LTP occurs. LTP has been demonstrated throughout the hippocampus, in the amygdala, in the neocortex, and elsewhere, and the mechanisms are not identical everywhere. The mossy fibre synapse from dentate granule cells to CA3, for instance, produces a form of LTP that does not require the NMDA receptor at all and appears to be expressed presynaptically. The account that follows is the canonical case, not the universal one.
The mechanism at the Schaffer collateral synapse
Everything that follows happens at a single glutamatergic synapse, and it turns on the fact that the postsynaptic membrane carries two different receptors for the same transmitter.
When the presynaptic CA3 terminal fires, it releases glutamate into the cleft. Glutamate diffuses across and binds to receptors on the CA1 spine. Two kinds of receptor matter here, and the entire logic of LTP lies in the difference between them.
AMPA receptors
These carry ordinary fast transmission. Glutamate binds, the channel opens, sodium flows in, and the postsynaptic membrane depolarises a little. This happens at every ordinary transmission event, it happens whatever the postsynaptic cell was doing, and on its own it changes nothing lasting. AMPA receptors are the current; they are also, as it turns out, the thing that gets modified.
NMDA receptors
These sit in the same membrane and bind the same glutamate, but they carry almost no current under ordinary conditions, because their pore is physically obstructed by a magnesium ion. They open only under a condition AMPA receptors do not care about, and when they do open they admit something AMPA receptors largely do not: calcium.
So the sequence for a single, ordinary, low-frequency transmission event is: glutamate released, AMPA receptors open, small depolarisation, NMDA receptors stay plugged, nothing changes. The synapse works, and it does not learn.
Now deliver a high-frequency burst. Glutamate is released repeatedly and rapidly, the AMPA-mediated depolarisations summate before they can decay, and the postsynaptic membrane is driven to a substantially depolarised state while glutamate is still bound. That is the condition that changes everything, and the next section explains why.
Why the NMDA receptor is the whole idea
At the resting membrane potential, the pore of the NMDA receptor is blocked by a magnesium ion sitting inside it. The magnesium is held there by the electrical field across the membrane: the inside of the cell is negative, and the positively charged Mg2+ is drawn into the channel and lodges there. Glutamate can bind to the receptor all it likes. The channel will not conduct, because it is plugged.
Depolarise the membrane, and the field that was holding the magnesium in place weakens. The Mg2+ is expelled. The pore is now clear. But an unplugged NMDA receptor still does nothing unless glutamate is bound to it, because glutamate is what opens the gate.
Put those two requirements together and the consequence is remarkable. The NMDA receptor passes current only when both of the following are true at the same time:
The presynaptic cell has fired
Glutamate has been released and is bound to the receptor. Without this there is no ligand, and the gate stays shut regardless of the membrane potential.
The postsynaptic cell is already depolarised
The magnesium block has been relieved. Without this the pore stays plugged, and the receptor conducts almost nothing no matter how much glutamate is bound.
Only then does calcium enter
With glutamate bound and the block relieved, the channel conducts, and critically it is permeable to Ca2+. Calcium floods into the spine. Calcium is the trigger, and it is the only signal in this sequence that carries the information that both conditions were met.
That is a molecular AND-gate. It is a single protein whose output reports the conjunction of presynaptic and postsynaptic activity, and it is difficult to overstate how neat that is. Hebb's postulate, written in 1949 as a piece of theoretical psychology, describes a rule that requires a cell to know whether its input contributed to its own firing. The NMDA receptor is a physical object that computes precisely that. The slogan "cells that fire together, wire together" is a crude compression of Hebb, but the reason the underlying idea survived is that a molecule turned out to implement it.
One complication is worth adding, because it changes who is in charge. Glutamate and depolarisation are necessary but not quite sufficient. The NMDA receptor also carries a separate co-agonist site, and unless that site is occupied, by glycine or by D-serine depending on the region, the channel stays shut however loudly the two neurons agree. Much of that co-agonist is supplied by the surrounding astrocytes. So the coincidence rule is a rule the neurons compute, but the permission to compute it at all is held, in part, by a cell that never fires.
Coincidence detector: any element that responds only when two or more conditions hold simultaneously. The NMDA receptor is the canonical biological example, and its coincidence requirement is not a metaphor. It is a direct consequence of a voltage-dependent block sitting in a ligand-gated channel.
What happens next is the expression of LTP. Calcium in the spine activates a set of kinases, of which CaMKII (calcium/calmodulin-dependent protein kinase II) is the most prominent and the most studied. CaMKII, once activated, has the notable property of being able to sustain its own activity through autophosphorylation, which lets a brief calcium signal outlast itself. The downstream consequences are twofold, and both act on AMPA receptors:
AMPA receptor insertion
Additional AMPA receptors are trafficked from intracellular pools and inserted into the postsynaptic membrane. The synapse now has more channels available to respond to the next release of glutamate.
Increased single-channel conductance
AMPA receptors already in the membrane are phosphorylated, which increases the current each one passes when glutamate binds. The same number of channels now carries more current.
The result is exactly what was measured in 1973 and has been measured ten thousand times since: the identical presynaptic signal now produces a larger postsynaptic response. The synapse has been strengthened. Notice how the labour is divided. The NMDA receptor decides whether to change the synapse. The AMPA receptor is what changes. The NMDA receptor is the sensor, calcium is the trigger, and the AMPA population is the memory.
Early phase and late phase: the point at which the cell must build something
LTP is not one process but two, running on different timescales and with different requirements, and the boundary between them is the boundary between a change that will fade and a change that will last.
Early-phase LTP
Lasts on the order of an hour or two. It is produced entirely by modifying and moving proteins the cell already has: existing AMPA receptors are phosphorylated and additional ones are trafficked in from existing pools. It requires no new gene expression and no new protein synthesis. Block protein synthesis and early-phase LTP is entirely unaffected. Left alone, it decays.
Late-phase LTP
Lasts many hours, and in the intact animal considerably longer. It requires signalling cascades to reach the nucleus, activate transcription factors including CREB, and switch on genes whose protein products are used to construct lasting change. Block transcription or block translation, and late-phase LTP does not form, even though the early phase appears perfectly normal for the first hour and then simply falls away.
Late-phase LTP is accompanied by structural change. Dendritic spines that have undergone strong potentiation grow: they enlarge, their heads expand, new spines form on the dendrite, and some of the new spines go on to make functional contacts. This is the point at which a change in a number becomes a change in an object. A stronger synapse is, quite literally, a bigger one.
This two-phase architecture maps directly onto memory consolidation, and the correspondence is one of the strongest reasons the field takes LTP seriously as a memory mechanism. At the behavioural level, a protein synthesis inhibitor given shortly after learning leaves the animal remembering normally for an hour or two, after which the memory disappears. At the synaptic level, a protein synthesis inhibitor given shortly after induction leaves potentiation intact for an hour or two, after which it disappears. The same drug, the same window, the same time course, the same failure. Synaptic consolidation and late-phase LTP are, on the current understanding, two descriptions of one event.
Long-term depression, without which the system would saturate
A brain that could only strengthen synapses would be useless, and it is worth being blunt about why. If every episode of correlated activity pushes a synapse up and nothing ever pushes it down, then every synapse in the network eventually reaches its ceiling. A network in which everything is maximally connected to everything carries no information at all. It has forgotten by remembering too much.
Long-term depression, LTD, is the mirror process that prevents this, and the elegant part is that it uses the same machinery.
The determining variable is the amount of calcium, not merely its presence. Strong, high-frequency stimulation opens many NMDA receptors at once and produces a large, sharp rise in postsynaptic calcium. That level of calcium preferentially activates kinases, CaMKII foremost among them, and kinases phosphorylate AMPA receptors and drive their insertion. The synapse strengthens.
Weak, prolonged, low-frequency stimulation produces a modest and sustained calcium rise instead. That level preferentially activates phosphatases, which do the opposite of kinases: they strip phosphate groups off AMPA receptors, and AMPA receptors are removed from the postsynaptic membrane by endocytosis. The synapse weakens.
One signal, two outcomes. Calcium is not a switch that means "strengthen". It is a graded signal whose magnitude and time course determine which downstream enzymes win. A large fast rise favours kinases and yields potentiation. A small sustained rise favours phosphatases and yields depression. The same ion, entering through the same receptor, at the same synapse, produces opposite changes depending on how much of it arrives and how quickly. This is why the NMDA receptor can serve as the trigger for both halves of a bidirectional system.
LTD is not a failure mode or a curiosity. It is the mechanism that keeps synaptic weights within a usable range, that allows the specific weakening of connections which turn out not to predict anything, and that makes the network's total capacity finite rather than exhausted. Any account of LTP that does not mention it is describing half a system and calling it a whole one. The bidirectionality is the point: memory requires the capacity to weaken exactly as much as it requires the capacity to strengthen, and both are forms of neuroplasticity.
Three properties that make LTP look like a memory mechanism
LTP has three canonical properties. They are usually listed. They are much less often explained, and each one exists because a memory mechanism would be broken without it.
Input specificity
Only the synapses that were actually stimulated are potentiated. Other synapses on the very same postsynaptic neuron, which did not receive the burst, are unchanged.
Why it matters: a single cortical or hippocampal pyramidal cell carries many thousands of synapses. If potentiating one of them potentiated all of them, then storing one memory would strengthen every connection the cell has, and the memory would smear across the entire cell. Every subsequent memory would overwrite the last. Input specificity is what allows a single neuron to participate in many different memories without those memories bleeding into one another. It is the reason the synapse, and not the neuron, is the unit of storage.
Associativity
A weak input, one which by itself is too feeble to induce LTP, will be potentiated if it is active at the same time as a strong input onto the same cell.
Why it matters: this is the cellular basis of association, and it falls straight out of the NMDA mechanism. The weak input releases glutamate onto its own synapses but cannot depolarise the cell enough to relieve the magnesium block. The strong input, arriving elsewhere on the same neuron, supplies that depolarisation. The weak synapse now satisfies both conditions at once, its NMDA receptors open, and it is potentiated. A stimulus that means nothing on its own acquires strength because it happened to coincide with one that already matters. That is what learning an association is, expressed in a single dendrite.
Cooperativity
There is a threshold. A single weak input cannot induce LTP on its own; a sufficient number of inputs must be active together to depolarise the cell past the point at which the magnesium block is relieved.
Why it matters: a threshold is what prevents the network from learning noise. Neurons fire spontaneously and randomly all the time. If any single presynaptic spike could potentiate its synapse, the network would spend its life encoding accidents. Cooperativity requires evidence in the form of convergent, coordinated input before a lasting change is made. It is the signal-to-noise filter of the whole system.
Notice that all three properties follow from one fact: the NMDA receptor requires coincident presynaptic glutamate and postsynaptic depolarisation. Specificity follows because glutamate is only present at the synapses that fired. Associativity follows because the depolarisation is a property of the whole cell and can be supplied by other inputs. Cooperativity follows because it takes a certain amount of summed input to reach the depolarisation the block requires. Three separate memory-like behaviours, one molecular cause.
The evidence tying LTP to learning, and where it stops
The case that LTP underlies memory is built from several independent lines of evidence, and it is genuinely strong.
Pharmacological. Infusing an NMDA receptor antagonist into the hippocampus impairs spatial learning in rodents. Animals treated in this way are markedly worse at learning the location of a hidden platform in a water maze, while remaining able to swim, see, and locate a visible platform normally. The deficit is in acquiring the new spatial memory, which is what the mechanism predicts, and not in the sensory or motor capacity to perform the task.
Genetic. Manipulations that impair the molecular machinery of LTP impair learning. Knocking out or disabling components of the NMDA-CaMKII-CREB pathway produces animals with deficient LTP and deficient hippocampus-dependent learning. Conversely, some manipulations that enhance LTP, such as increasing the proportion of the NMDA receptor subunit associated with longer channel opening, produce animals that learn certain tasks somewhat better. The correlation runs in both directions, which is considerably more persuasive than a one-sided knockout result.
Occlusion and detection. If learning uses LTP, then learning ought to leave LTP-like traces at the relevant synapses, and there is work indicating exactly that: after training, synapses in the circuits engaged by the task show the electrophysiological and molecular signatures of potentiation, and in some preparations prior learning reduces the amount of LTP that can subsequently be induced at those synapses, as if the mechanism had already been partially used up.
Convergence. The two-phase structure of LTP matches the two-phase structure of memory consolidation, with the same protein-synthesis dependence and the same time course. The three canonical properties of LTP match the three things a memory mechanism would need. The anatomy is right: it is most robust in the hippocampus, and the hippocampus is the structure whose loss abolishes the formation of new declarative memories.
And now the caveat, which is the intellectual point of this page. LTP, as it is studied, is an artificial induction protocol. A brief train of stimulation at a frequency and intensity chosen by the experimenter is delivered through an electrode to a bundle of axons in a slice of hippocampus in a dish, or in an anaesthetised animal. Nothing in that procedure is a memory. It is a demonstration that the tissue is capable of a lasting, specific, associative change in synaptic strength, which is precisely the capability a memory mechanism would require. The inference from that capability to the claim that this is how memories are stored is an inference, and it should be named as one. It is supported by pharmacology, by genetics, by anatomy, by time course, and by the absence of any competing candidate that comes close. It is held by very nearly everyone who works on the problem, and this page holds it too. But an inference supported by convergent evidence is not the same thing as having watched a memory be written, read back, and identified in the synapses that hold it. The honest statement of the field's position is that LTP is overwhelmingly the leading candidate mechanism for memory storage, and that this remains a candidacy.
It is worth saying why this distinction is not pedantry. Almost every popular account of LTP presents it as settled: the brain stores memories by strengthening synapses, mechanism known, question closed. That framing throws away the interesting part. The genuinely hard problems, how a distributed pattern of synaptic weights constitutes a particular memory, how such a pattern is read out during recall, how it survives the continuous molecular turnover of the proteins that hold it, and how many synapses across how many regions a single remembered event actually occupies, are not solved by LTP. They are the questions LTP makes it possible to ask.
What LTP is routinely claimed to have shown, and has not
Claim: "cells that fire together, wire together" is a proven law of the brain.
It is a slogan, coined decades after the fact, compressing a postulate that Donald Hebb published in 1949 in The Organization of Behavior. Hebb had no evidence for it; he proposed it because a learning brain would need something like it. The slogan is also a poor compression: Hebb specified that cell A must repeatedly take part in firing cell B, which is a causal and ordered relationship, not mere simultaneity. LTP matters precisely because it later supplied a plausible mechanism for the postulate, in the form of the NMDA receptor's coincidence requirement. The correct order of the story is: Hebb proposed a rule the brain would need, and physiology later found a molecule that could implement it. Presenting the slogan as an established law inverts that, and hides the fact that it was a bet which happened to pay off.
Claim: LTP has been shown to be a memory.
It has not. LTP has been shown to be a lasting, input-specific, associative, cooperative increase in synaptic strength that can be induced experimentally, that depends on molecular machinery whose disruption also disrupts learning, and that shares the protein-synthesis dependence and time course of memory consolidation. Every one of those is a statement about a mechanism the brain possesses. None of them is the observation of a memory being stored in identified synapses and later read out from them. The gap is small, the bridge across it is strong, and it is still a bridge.
Claim: LTP works the same way everywhere in the brain.
It does not. The account given on this page is the canonical NMDA-receptor-dependent, postsynaptically expressed form found at the Schaffer collateral input to CA1. The mossy fibre synapse onto CA3, only a short distance away in the same structure, shows a form of LTP that does not require NMDA receptors and is expressed presynaptically as an increase in transmitter release. Cerebellar synapses show LTD as their dominant plasticity. "LTP" names a class of phenomena unified by their outcome, a lasting increase in synaptic strength, not a single universal molecular pathway.
Claim: because LTP strengthens synapses, more LTP means a better memory.
Unrestrained potentiation would saturate the network and destroy its capacity to store anything new, which is the reason long-term depression exists and the reason homeostatic mechanisms scale synaptic weights as a whole. A memory system is defined by the pattern of which synapses are strong and which are weak. Raising all of them is not an improvement; it is the erasure of the pattern.
Sources
- Bliss TVP, Lomo T. Long-lasting potentiation of synaptic transmission in the dentate area of the anaesthetized rabbit following stimulation of the perforant path. Journal of Physiology. 1973;232(2):331-356.
- Hebb DO. The Organization of Behavior: A Neuropsychological Theory. Wiley; 1949.
- Kandel ER, Koester JD, Mack SH, Siegelbaum SA. Principles of Neural Science. 6th ed. McGraw-Hill; 2021.
This page is an educational reference. It is not medical advice and does not diagnose or treat any condition.