How Neurons Communicate: The 20-Nanometer Gap Every Thought Must Jump

Introduction

There are roughly 86 billion neurons in your head. That number comes from a 2009 study in which Frederico Azevedo and colleagues dissolved four adult human brains into a uniform soup and counted the nuclei directly, which is a cruder and far more reliable method than the estimates it replaced [1]. Four brains is a small sample and the paper says so. It is still the best direct count anyone has.

None of those 86 billion cells touch each other in the way you probably imagine. At almost every point where one neuron passes a message to another, there is a gap. The gap is about 20 to 30 nanometers wide at synapses inside the brain, which is roughly the width of a virus particle, or about one four-thousandth the width of a human hair.

Everything crosses that gap. Reading this sentence. Remembering where you left your keys. The decision you will make in an hour without noticing you made it. All of it is chemical packets thrown across a space so small that the molecules arrive in well under a millisecond.

That much is in every explanation of how neurons communicate, and it is correct. What almost none of them tell you is that this tidy two-cell picture was essentially complete as a scientific account by about 1995, and the field has not stood still since. Neurons also talk directly to each other through channels that skip the gap entirely. They broadcast chemicals into the space between cells, where the signal reaches receptors nowhere near any synapse. They nudge their neighbours with raw electric fields. And at a great many synapses there is a third cell wrapped around the junction, listening and answering, that does not appear in the diagram at all.

This is the long version. It starts with an argument that two men had in public while accepting a Nobel Prize.

The Argument Settled At A Nobel Ceremony

In December 1906, Camillo Golgi and Santiago Ramón y Cajal shared the Nobel Prize in Physiology or Medicine for their work on the structure of the nervous system. They were not friends. They did not agree. And on 11 December, Golgi used his Nobel lecture to attack the theory that had won the prize, while Cajal sat in the audience.

The disagreement was about whether the nervous system is one thing or many. Golgi held to the reticular theory: nerve tissue is a single continuous network, a fused web through which signals flow the way water flows through pipes. Cajal held the opposite, that the nervous system is built from discrete cells that come very close to one another without merging. Wilhelm von Waldeyer had given those cells a name in 1891. He called them neurons.

The irony is almost too neat. The method Cajal used to prove Golgi wrong was Golgi's own invention. In 1873 Golgi had developed a silver-chromate stain, la reazione nera, the black reaction, which for reasons still not fully understood picks out a small random fraction of cells and blackens each one completely, dendrites and axon included. Against the pale background of unstained tissue, a single neuron stands out whole. Golgi built the tool. Cajal, working in Spain through the late 1880s, used it better, and what he saw were endings. Branches that stopped. Cells that approached and did not join.

Cajal took his slides to the German Anatomical Society congress in Berlin in 1889 and convinced Rudolf von Kölliker, the most respected histologist of his generation. That was the turning point. The neuron doctrine won.

Cajal later described the Stockholm ceremony with a certain amount of feeling, calling Golgi's lecture an extravagant lucubration and remarking on the cruelty of fate in pairing scientific adversaries of such contrasting character like Siamese twins joined at the shoulder. The shy sixty-three-year-old Italian and the combative fifty-four-year-old Spaniard never became friends.

Hold on to one detail, because it comes back near the end of this article. Golgi was wrong about the general case. He was not wrong about everything. There is a class of junction where neurons really are joined cytoplasm to cytoplasm, exactly the continuity he insisted on, and we will get to it.

Charles Sherrington supplied the missing word. In 1897, writing in Michael Foster's textbook of physiology, he introduced the term synapse, from the Greek for clasping together, for the junction where one neuron acts on the next. He had a name for the place before anyone knew what happened there.

Thought Is Slower Than You Think

Before anyone knew what a synapse was, someone measured how fast the nervous system runs, and the answer disturbed people.

Around 1849 and 1850, Hermann von Helmholtz measured the conduction velocity of a frog's motor nerve. He got roughly 27 to 30 metres per second. Fast by everyday standards, and shockingly slow by the standards of the time, because the prevailing assumption was that nervous signals were more or less instantaneous. Volition was supposed to be immediate. Helmholtz showed it takes measurable time to travel down a nerve, which meant that the will has a speed, and that speed can be written down.

Modern numbers give a wider range. Nerve fibres in a mammal conduct anywhere from about half a metre per second to about 120 metres per second, and the difference comes down to two things: how thick the fibre is, and whether it is wrapped in myelin. Joseph Erlanger and Herbert Gasser worked out the classification in the 1920s and 1930s and shared a Nobel Prize for it in 1944.

That table explains something you have felt. Stub your toe and there are two pains, a sharp one and then a duller one that arrives a moment later. The sharp signal travels on thinly myelinated A-delta fibres at up to 30 metres per second. The dull one travels on unmyelinated C fibres at walking pace. The gap between them is the table made physical.

Myelin is what buys the speed. Oligodendrocytes in the brain and spinal cord, and Schwann cells in the periphery, wrap themselves around axons in layer after layer of membrane, leaving small bare patches at intervals called nodes of Ranvier. The electrical signal effectively jumps from node to node instead of crawling along the whole membrane, which is why the process is called saltatory conduction, from the Latin for leaping. Klaus-Armin Nave's work on myelination shows that glia do far more than insulate here; they also support the metabolic integrity of the axon itself [2].

Two findings complicate the picture in ways almost no explainer mentions. First, myelin is not laid down as a uniform sleeve. Giulio Srubek Tomassy and colleagues showed in 2014 that its distribution varies substantially along individual axons in the cortex, with long stretches left bare in patterns that look deliberate rather than accidental [3]. Second, the small details matter more than you would guess. Lorena Arancibia-Cárcamo and colleagues found in 2017 that the length of the node of Ranvier alone, a structure about a micrometre long, can shift conduction velocity by around twenty per cent [4]. The wiring is tuned.

The Signal That Refuses To Fade

The thing travelling down the axon is not electricity in the sense a wire carries it. A copper wire passively conducts, and the signal weakens with distance. An axon regenerates its signal at every step, which is why a message can travel a metre down your leg without getting fainter.

At rest, the inside of a neuron sits at roughly minus 70 millivolts relative to the outside. That difference is maintained by pumps that push sodium out and potassium in, and by the selective leakiness of the membrane. Think of it as a loaded spring, or a reservoir held behind a dam.

Push the membrane potential up by about 15 millivolts, to somewhere near minus 55, and something abrupt happens. Voltage-gated sodium channels snap open, sodium floods in, and the potential rockets past zero to around plus 30 or plus 40 millivolts before potassium channels open and drag it back down. The whole event takes a millisecond or two. Then the membrane briefly overshoots in the other direction and there is a short period during which nothing can trigger another spike, which is what stops signals running backwards.

The critical property is that it is all or nothing. A stimulus that reaches threshold produces a full-sized action potential. A stimulus just below threshold produces nothing at all. There is no half-spike. Information is carried in how often spikes happen and in exactly when, not in how big they are.

The proof of all this came from a squid.

In 1936 J. Z. Young identified that a thick structure in the squid's mantle, which had been mistaken for a blood vessel, was in fact a single enormous nerve fibre. At up to about a millimetre across, it was big enough to push a wire inside. Alan Hodgkin and Andrew Huxley did exactly that at Plymouth, and in 1939 published the first recording of an action potential made from inside a nerve cell [5].

Then the Second World War stopped the work. Both men spent the war on military research. They came back to it afterwards with a new tool, the voltage clamp, which let them hold the membrane at a chosen voltage and measure the current required to keep it there. From that they reverse-engineered the behaviour of the channels.

The result was a series of five papers in the Journal of Physiology in 1952, ending with a quantitative model that predicted the shape, size and speed of the action potential from a handful of equations [6]. It is still the framework the field uses.

It was also computed by hand. There were no accessible computers, so the numerical integration was done on a hand-cranked mechanical calculator, more than a million operations carried to six decimal places, over a period of months. That is a large part of why three years separate the experiments from the publication. Hodgkin, Huxley and John Eccles shared the Nobel Prize in 1963.

The model is not considered complete. Purely electrical descriptions leave some phenomena unexplained, and there are minority proposals, including a thermodynamic or soliton account of the nerve impulse, that argue a mechanical wave accompanies the electrical one. Nothing has overturned Hodgkin and Huxley. The honest statement is that their framework is the gold standard and is acknowledged to be an incomplete description of everything an axon does.

The Gap Itself

The action potential arrives at the end of the axon and stops. It cannot cross. What happens next is a translation from electricity into chemistry and back again.

The space it has to cross is small and its size is worth being precise about. At synapses inside the central nervous system, the cleft is generally given as 20 to 30 nanometres. At the neuromuscular junction, where a motor neuron meets a muscle fibre, it is wider, around 50. Electron tomography refines this further: excitatory synapses show a fairly uniform cleft of about 18 to 20 nanometres, while inhibitory synapses narrow toward roughly 6 nanometres at their edges.

For scale, a typical virus particle is about 100 nanometres across. A synaptic cleft is a fraction of that. It is a gap so small that a molecule released on one side reaches the other by simple diffusion in a matter of microseconds.

That is why the crossing itself is not the slow part. When a signal takes time at a synapse, the time goes into the machinery on the near side: getting calcium in, getting a vesicle fused, getting transmitter out. Bernard Katz and Ricardo Miledi measured the delay directly at the frog neuromuscular junction in 1965 and found it to be roughly half a millisecond to a millisecond, with a floor somewhere around 0.3 to 0.5 milliseconds [7]. That figure, or a rounded version of it, is what almost every explanation of synaptic transmission still quotes.

It is a frog measurement, taken at room temperature, at a specialised junction built for reliability rather than speed. Fast synapses in a mammalian brain, running at body temperature, are considerably quicker than that. Bernardo Sabatini and Wade Regehr examined the timing of transmission at fast central synapses and found the interval between the arriving spike and the postsynaptic response to be far shorter than the classical neuromuscular figure [8]. If you have been carrying the one-millisecond number around, it is worth knowing where it came from and what it does not describe.

The Experiment That Arrived In A Dream

Through the 1920s nobody knew whether transmission across that gap was electrical or chemical, and the question was decided by an experiment that its author said came to him while asleep.

Otto Loewi, working in Graz, told the story often enough afterwards that the details are well known. The idea for the experiment arrived in a dream. He woke, scribbled a note, and went back to sleep. In the morning he could not read his own handwriting. The following night the dream returned, and this time he got up at three in the morning and went straight to the laboratory.

The experiment was elegant. He took two frog hearts, one still attached to its vagus nerve, both bathed in saline. He stimulated the vagus of the first heart, which slowed it, as expected. Then he took the saline surrounding that first heart and applied it to the second heart, which had no nerve attached. The second heart slowed too.

Nothing electrical had passed between them. Something chemical had. Loewi called the substance Vagusstoff, vagus material. It turned out to be acetylcholine, which Henry Dale had isolated in 1914. The manuscript went to Pflügers Archiv in March 1921, and Loewi and Dale shared the Nobel Prize in 1936.

There is a detail that makes the story better rather than worse. The experiment does not work reliably at every time of year. Seasonal variation in the frogs' acetylcholinesterase and receptor levels means that at other points in the calendar the released acetylcholine is broken down too fast for the effect to show. Loewi later wrote that in the cold light of morning he would never have attempted anything so unlikely. He got the right answer partly because he did not stop to think about it, and partly because it was the right month.

The argument did not end there. Through the 1930s and 1940s the field split into what participants themselves called the soup and spark war. Dale led the chemical camp. John Eccles, a former student of Sherrington's, led the electrical one, and he was rigorous about it. Influenced by Karl Popper's insistence that a theory earns its place by surviving attempts to falsify it, Eccles set out to disprove chemical transmission. In 1951 his own intracellular recordings from motor neurons revealed the inhibitory postsynaptic potential, and the result pointed the other way. He changed his position publicly. It remains one of the cleaner examples of a scientist being talked out of a theory by his own data.

Packets, Not A Stream

The next surprise came from listening to a synapse doing nothing.

In 1952, Paul Fatt and Bernard Katz put an electrode into a frog muscle fibre at the neuromuscular junction and recorded with no stimulation at all. They expected silence. Instead they found small blips, roughly half a millivolt to a millivolt, arriving at random intervals [9]. They called them miniature end-plate potentials, minis for short. The synapse was leaking.

Two years later Katz and José del Castillo worked out what the leaks meant. A real end-plate potential, the kind produced when the nerve actually fires, is not a smooth continuous release of transmitter. It is built from whole-number multiples of the miniature unit. One mini, two minis, three, never one and a half [10]. Release is quantised. Transmitter comes in fixed packets, and the nerve impulse determines how many packets get thrown, not how much is in each one.

Katz's team also showed the release is probabilistic and depends on calcium. Lower the calcium in the bath, raise the magnesium, and the number of packets released per impulse falls in a way that follows Poisson statistics. Each release site is essentially flipping a weighted coin. Katz received the Nobel Prize in 1970.

Electron microscopy soon supplied the physical object. In the early 1950s George Palade, Sanford Palay, and independently Eduardo De Robertis and H. Stanley Bennett, saw small spheres clustered at the presynaptic terminal. Katz's statistical quantum had a body. It was a vesicle, and modern measurements put it at about 40 to 42 nanometres across, which Shigeo Takamori and colleagues catalogued in remarkable molecular detail in 2006 [11].

How many molecules are in one? This is a number worth handling carefully, because the figure everybody repeats is not a measurement. The commonly cited 5,000 glutamate molecules per vesicle is a modelling convention that entered the literature as a working assumption and then got quoted as fact. Yuanmo Wang and colleagues measured it directly in 2019 and found that an isolated single synaptic vesicle contains roughly 8,000 glutamate molecules [12]. The broader literature spans something like 2,000 to 10,000 depending on the synapse and the method. Dieter Bruns and Reinhard Jahn had earlier developed the approach of watching release from single vesicles in real time, which is what made this kind of question answerable at all [13]. Give the range, not the round number.

The Machinery

Knowing that vesicles fuse does not tell you how. That took another forty years and a different set of tools.

Calcium is the trigger. When the action potential reaches the terminal, voltage-gated calcium channels open, and calcium enters in a very brief, very local pulse. Erwin Neher's work established how tight those calcium microdomains are, and how the distance between a channel and a vesicle determines whether that vesicle releases [14]. Rodolfo Llinás demonstrated at the squid giant synapse that presynaptic calcium influx is the direct cause of release, not merely a correlate of it.

The fusion itself is done by a set of proteins called SNAREs. Thomas Söllner, working with James Rothman, identified them in 1993: synaptobrevin sitting on the vesicle, syntaxin and SNAP-25 on the target membrane [15]. The three wind around each other into a coiled bundle, and the winding physically drags the two membranes together until they merge. Thomas Südhof then identified synaptotagmin-1 as the calcium sensor that decides when this is allowed to happen, along with complexin, which acts as a clamp holding the machine cocked and ready. Südhof's account of the vesicle cycle as a cascade of protein interactions [16], his later work on the architecture of the presynaptic active zone [17], and his synthesis of what happens in the final millisecond of a vesicle's life [18] are the clearest routes into this literature. Randy Schekman, James Rothman and Thomas Südhof shared the Nobel Prize in 2013.

It is worth pausing on why this machine has to be so elaborate. Two lipid membranes do not want to merge. Doing it in under a millisecond, at a specific place, only when calcium says so, and then recycling the components fast enough to do it again immediately, is a genuinely hard engineering problem. The synapse solves it thousands of times a second.

What Happens On The Far Side

Transmitter crosses. Then it has to be read, and the reading comes in two very different flavours.

Ionotropic receptors are channels. The transmitter binds and the pore opens, directly, in well under a millisecond. Glutamate acting on AMPA receptors does this, and so does GABA acting on GABA-A receptors. Fast, brief, point to point.

Metabotropic receptors are not channels. Binding activates a G protein inside the cell, which sets off a chemical cascade that may open channels somewhere else, or change how the cell responds to other inputs, or alter gene expression. Slower to start, much longer lasting, and modulatory rather than declarative. Most dopamine and serotonin receptors work this way, which is a large part of why those systems feel like they set a mood rather than carry a message. The dopamine case in particular is worth its own treatment, and it gets one in dopamine and learning.

One receptor deserves separate attention because it is the hinge on which the second half of this article turns. The NMDA receptor is ionotropic, but it does not open on glutamate alone. At resting voltage its pore is physically plugged by a magnesium ion. Glutamate binding is not enough to move it. The cell also has to be depolarised at the same moment, which expels the magnesium and lets calcium through.

That makes the NMDA receptor a coincidence detector. It reports not that the presynaptic cell fired, nor that the postsynaptic cell was active, but that both happened together. Two keys, one lock. Geoffrey Lau and Suzanne Zukin's account of how these receptors are trafficked in and out of synapses is the standard reference for how that machinery is regulated [19]. Hold that thought.

The receiving cell is not a passive tally board either. A single cortical pyramidal neuron may carry many thousands of synapses spread across an elaborate dendritic tree, and where an input lands changes what it does. Jeffrey Magee's work on dendritic integration [20] and Nelson Spruston's on pyramidal neuron structure [21] describe dendrites that perform local computations before anything reaches the cell body. Yuriy Mishchenko and colleagues, reconstructing hippocampal tissue at electron-microscope resolution, gave hard numbers to how densely packed those connections are [22]. The figure often quoted, around 7,000 synapses per neuron, applies to neocortical neurons specifically. Averaged across every neuron in the brain the number is far lower, closer to a thousand, because cerebellar granule cells have only a handful of inputs each. Purkinje cells sit at the other extreme with well over a hundred thousand.

Some neuron populations have been studied so intensively that they have their own literature, and the clearest popular example is the one covered in mirror neurons.

A second simplification worth dropping: the idea that each neuron releases one neurotransmitter. Thomas Hnasko and Robert Edwards reviewed the evidence that corelease, a single neuron releasing more than one transmitter, is common rather than exotic [23]. And receptors are not scattered randomly across the postsynaptic membrane. Ai-Hui Tang and colleagues showed in 2016 that they are organised into nanocolumns aligned directly opposite the presynaptic release sites [24]. The transmitter is not thrown into a crowd. It is aimed.

Finally, a synaptic event does not have to stay local. Paul Greer and Michael Greenberg's work traces the route from calcium entering at a synapse to changes in gene transcription in the nucleus [25]. A signal that starts as a few thousand molecules crossing 20 nanometres can end as a change in which genes a cell is reading.

Golgi Was Partly Right

Now the promised return to the 1906 Nobel ceremony.

Not every synapse is chemical. At an electrical synapse, two neurons are joined by gap junctions: paired channels, each built from two connexons, each connexon a hexamer of connexin protein, which form a continuous pore from the cytoplasm of one cell into the cytoplasm of the next. In neurons the principal protein is connexin-36. The membranes sit about 3 nanometres apart instead of 20 to 30, and current flows straight through with no chemistry involved.

Michael Bennett and Suzanne Zukin's review of electrical coupling in the mammalian brain is the best entry point to how much this matters [26]. Electrical synapses are fast, essentially instantaneous, and usually bidirectional. They are particularly good at making populations of neurons fire together, which is central to producing synchronised rhythms.

That is cytoplasmic continuity between neurons. It is exactly what Golgi insisted the nervous system had and Cajal insisted it did not. Cajal was right about the general architecture. Golgi was right that this kind of joining exists. Both of those statements are true, and the second one almost never gets said.

Talking Without A Synapse At All

The synapse is a private phone call. It is not the only way neurons communicate, and the alternatives are the part of this subject that popular explanations skip entirely.

**Volume transmission.** Luigi Agnati and Kjell Fuxe proposed in the mid-1980s that chemical signals also diffuse through the extracellular space to reach receptors that are nowhere near a synapse. This is broadcast rather than point to point. Glutamate and GABA spill out of the cleft and reach extrasynaptic receptors; tonic inhibition through extrasynaptic GABA-A receptors is a normal and important feature of cortical function; and much neuromodulator signalling, dopamine included, works partly this way. Giles Hardingham and Hilmar Bading showed that synaptic and extrasynaptic NMDA receptors can drive opposite outcomes in the same cell, which makes the distinction between the cleft and the space around it functionally real rather than merely anatomical [27].

**Ephaptic coupling.** Every active neuron generates an electric field, and that field extends beyond the cell. Costas Anastassiou and colleagues showed in 2011 that endogenous extracellular fields of the size that actually occur in cortex can entrain the spiking of nearby neurons [28]. In other words neurons influence each other simply by being near each other while active, without any synapse involved. Whether this matters for normal brain function or is a real but minor side effect is genuinely disputed, and this article does not pretend otherwise.

**Retrograde signalling.** Information also runs backwards across the synapse. Endocannabinoids made by the postsynaptic cell travel back to the presynaptic terminal and suppress further release, which is a feedback loop built into the junction. Nitric oxide, being a small gas, diffuses freely in any direction.

**Extracellular vesicles.** Neurons and glia release membrane-bound packages containing RNA and protein that are taken up by other cells. This is material transfer rather than signalling in the classical sense, and it is an active area of work.

None of these replace the chemical synapse. They sit alongside it, and any picture that has only the synapse in it is missing most of the channels.

The Third Cell

Here is the largest omission in the standard account.

For most of the twentieth century glia were regarded as support staff. The name itself means glue. They were understood to insulate axons, clean up debris, buffer the chemical environment, and otherwise stay out of the way of the interesting cells.

In 1999 Alfonso Araque, Vladimir Parpura, Rita Sanzgiri and Philip Haydon published a proposal they called the tripartite synapse: that at a great many junctions the correct unit of analysis is not two cells but three, because the astrocyte wrapped around the synapse both responds to transmission and modifies it [29]. Gertrudis Perea, Marta Navarrete and Araque extended the argument a decade later, describing astrocytes as processing and controlling synaptic information rather than merely reacting to it [30].

The proposed mechanism runs through calcium. Astrocytes do not fire action potentials, but they do produce internal calcium elevations, and those elevations are associated with the release of substances that act on neurons: glutamate, ATP, and D-serine. Aude Panatier and colleagues showed in 2006 that glia-derived D-serine acts as a required co-agonist at the NMDA receptor, meaning that the coincidence detector at the heart of synaptic learning needs a permission signal from a third cell to work properly [31]. Pascal Jourdain and colleagues reported that glutamate released from astrocytes can control synaptic strength directly [32]. David Stellwagen and Robert Malenka found that a glial signalling molecule, TNF-alpha, is required for a form of synaptic scaling [33].

Astrocytes also help build synapses in the first place. Nicola Allen and colleagues identified astrocyte-secreted glypicans that promote the formation of functional excitatory synapses [34], and Allen and David Lyons later framed glia as architects of nervous system formation rather than caretakers of it [35]. Alexei Verkhratsky and Maiken Nedergaard's review is the fullest modern reference on astrocyte physiology [36], and Baljit Khakh and Michael Sofroniew have documented how varied astrocytes are between brain regions [37]. Christian Giaume and colleagues described how astrocytes are themselves interlinked into gap-junction-coupled networks [38]. Ye Zhang and colleagues purified and characterised human astrocytes specifically, which matters because most of this work was done in rodents [39].

In August 2025, Lucas Benoit, Sandra Hristovska, Fabrice Liaudet and colleagues published work in Cell that changes the shape of the picture again [40]. Combining volumetric high-resolution electron microscopy with two-photon calcium imaging, they characterised astrocytic leaflets, the extremely fine processes, 250 nanometres or less, that wrap around synapses. Inside those leaflets they found minuscule saccules of endoplasmic reticulum carrying IP3 receptors, with IP3R1 identified as the driver of local calcium signals, and no mitochondria. Crucially, the leaflets are interconnected through gap junctions into functional domains that integrate signals across multiple synapses at once.

That is a different sort of object from the one in the diagram. Not a passive wrapper around a single junction, but a structure that samples many synapses and coordinates a response across them.

Microglia, the brain's resident immune cells, belong in this section too. Beth Stevens and colleagues showed that the classical complement cascade, borrowed from the immune system, tags synapses for elimination [41], and Dorothy Schafer and colleagues showed microglia physically engulfing those tagged synapses in an activity-dependent way [42]. Communication between neurons is being actively edited by cells that are not neurons. If you want the developmental version of that story, it is the subject of how the adolescent brain deletes half its synapses.

One caution, because this is a place where enthusiasm has run ahead of certainty in some of the popular coverage. Whether gliotransmission is a normal physiological signalling channel in an intact brain, or is largely an artefact of the strong manipulations used to study it, is a live argument. Thomas Fiacco, Cendra Agulhon and Ken McCarthy have argued the sceptical case forcefully. And Wei Sun and colleagues found in 2013 that glutamate-dependent neuroglial calcium signalling differs between mice and humans, which is a real problem for generalising rodent results [43]. The tripartite synapse is an important and productive idea. It is not settled consensus.

When A Synapse Learns

Everything so far describes a message being sent. The reason any of it matters for you personally is that the sending changes the sender.

In 1966 Terje Lømo, working in Per Andersen's laboratory in Oslo, noticed something while stimulating the perforant path of a rabbit hippocampus. After a burst of high-frequency stimulation, the response to a single test pulse was larger than it had been before, and it stayed larger. Timothy Bliss joined him, and their 1973 paper reported the effect formally: potentiation lasting anywhere from thirty minutes to ten hours after a brief tetanus [44]. They called it long-term potentiation.

The mechanism runs straight through the coincidence detector described earlier. Strong activity depolarises the postsynaptic cell enough to expel the magnesium block from NMDA receptors while glutamate is present. Calcium enters. Calcium activates CaMKII, which drives more AMPA receptors into the postsynaptic membrane. The synapse is now physically better at responding to the same input. That is early LTP, and it takes minutes.

Late LTP is different in kind. Sustained changes require new protein synthesis and new gene transcription, with CREB as a central player. This is the difference between turning up a dial and rebuilding the dial. Bliss and Graham Collingridge's 1993 review is what made LTP the standard model of synaptic memory [45], and Eric Kandel's 2001 synthesis of the molecular biology of memory storage remains the single clearest account of how a synaptic event becomes a lasting one [46].

The reverse also happens. Long-term depression weakens synapses, and the deciding factor appears to be the shape of the calcium signal rather than its presence. Large, brief calcium elevations strengthen. Smaller, more prolonged ones weaken.

Then the timing turned out to matter more precisely than anyone expected. In 1997 Henry Markram, Joachim Lübke, Michael Frotscher and Bert Sakmann recorded from pairs of connected layer-5 pyramidal neurons in the neocortex and varied the order of presynaptic and postsynaptic spikes [47]. Guo-qiang Bi and Mu-ming Poo did the equivalent in hippocampal cultures the following year [48]. What they found is now called spike-timing-dependent plasticity, and it is startlingly sharp.

If the presynaptic neuron fires shortly before the postsynaptic one, within roughly 20 milliseconds, the synapse strengthens. If the order reverses, so that the postsynaptic cell fires first, the same synapse weakens. The transition between the two happens over a window of a few milliseconds. Bi and Poo's later review connects this directly back to Donald Hebb's 1949 postulate, usually paraphrased as cells that fire together wire together [49], and Natalia Caporale and Yang Dan's account treats it as a general learning rule [50]. Hebb's version was a hunch stated in words. This is the same idea with a millisecond-resolution graph attached, and it corrects the popular paraphrase: firing together is not enough. Order matters, and the wrong order actively weakens the connection.

Two more pieces complete the picture. Uwe Frey and Richard Morris showed in 1997 that a weakly stimulated synapse can be marked, or tagged, and later capture plasticity-related proteins manufactured because of strong activity elsewhere in the cell [51]. This is why a trivial event that happens near an important one can end up being remembered. And Jaime Grutzendler and colleagues showed by imaging the same dendrites in living adult cortex over time that many spines, the small protrusions where excitatory synapses sit, are stable for months [52]. Lorenzo Cingolani and Yukiko Goda described the actin machinery that lets those structures change shape while remaining stable [53].

So there is a physical object with the right properties. Small enough to be numerous, modifiable in minutes, and stable for months. Eve Marder and Jean-Marc Goaillard's work is a useful corrective here, showing that neurons and circuits maintain their function through constant compensation rather than by holding every parameter fixed [54]. The brain is not a set of frozen settings. It is a system that keeps re-tuning itself and stays recognisable anyway.

If you want the version of this story told from the memory side rather than the synapse side, how the hippocampus decides what to remember covers the structure where LTP was found, and reconsolidation covers what happens to a memory each time it is retrieved.

Why Spacing And Testing Work

This is the part that has practical consequences, and it needs to be stated with its limits attached.

Two findings from learning research are about as solid as behavioural science gets. First, spreading study sessions out beats cramming them together. Hermann Ebbinghaus noticed it in 1885 and wrote that a suitable distribution of repetitions over a space of time is decidedly more advantageous than massing them at a single time. Second, testing yourself beats rereading. Henry Roediger and Jeffrey Karpicke demonstrated this repeatedly, first in 2006 [55] and then in a 2008 Science paper showing that repeated retrieval, not repeated study, is what produces durable learning [56].

Both findings are usually presented as bare facts about behaviour. They have a molecular story underneath them.

Paul Smolen, Yili Zhang and John Byrne reviewed that story in 2016 [57]. The pathways that convert synaptic activity into lasting change, particularly the kinase cascades involving PKA and ERK and the CREB-driven transcription that follows, do not respond linearly to repetition. They have their own timing. After a bout of activity these pathways go through a period of elevated activity and then a refractory period during which further stimulation accomplishes much less. Space the training bouts so that each one lands while the machinery is receptive and you get durable late-phase LTP. Mass them together and later repetitions land in the refractory window, doing far less work than the effort suggests.

That is the reason cramming feels productive and is not. The work is going into a system that is temporarily unable to use it.

Retrieval practice has a parallel story. Recalling something is not a read-only operation. It reactivates the trace and puts it briefly back into a modifiable state, which is why retrieval both strengthens the memory and creates an opportunity to change it. That process is covered in more detail in retrieval practice and the brain.

Sleep belongs here as well. Chiara Cirelli and Giulio Tononi's synaptic homeostasis hypothesis proposes that waking experience potentiates synapses across the board, and that sleep scales them back down to preserve signal-to-noise and stay within the brain's energy and space budgets. Luisa de Vivo and colleagues provided ultrastructural evidence for it in 2017, measuring synapse size in mouse cortex and finding a net decrease after sleep [58]. There is more on that in the sleeping brain and memory.

Now the limit, stated plainly. Nothing cited above measures long-term potentiation in a human being who is using flashcards. The cellular mechanism and the behavioural finding point the same direction, and the molecular account of spacing gives a coherent reason why they should. That is an inference, and a good one, but it is an inference. Anyone telling you that a study schedule causes LTP in your hippocampus is describing a model, not a measurement.

What is safe to say is this. Synapses change with use. The change depends on timing at the scale of milliseconds and on spacing at the scale of hours. Study methods that respect both of those do better than methods that do not, and now you know what is happening underneath.

What Is Still Argued About

Popular explanations of this subject present it as finished. It is not, and the open questions are more interesting than the settled ones.

**Does gliotransmission matter in a living brain?** Covered above. Araque and Haydon's camp says astrocytes actively shape transmission and behaviour. Fiacco, Agulhon and McCarthy argue much of the supporting evidence relies on manipulations that do not occur naturally. The species difference Sun and colleagues found makes it harder still [43]. Genuinely unresolved.

**Is ephaptic coupling functional?** Anastassiou and Christof Koch showed real entrainment by endogenous fields [28]. Sceptics say the effect sizes under physiological conditions are too small to matter. Unresolved.

**Is LTP actually the mechanism of memory?** It remains the leading model and the evidence is substantial. It is also true that demonstrating LTP in a slice is not the same as demonstrating that natural learning uses it at the same synapses, and there are dissociations in both directions. Optogenetic work on engram cells, tagging and later reactivating the specific ensembles that store a memory, has strengthened the case considerably without closing it.

**Is memory stored in synapses at all?** A minority position, argued seriously by Charles Randy Gallistel among others, holds that the real storage may be intracellular and molecular rather than synaptic, on the grounds that synaptic weights are a poor medium for the kind of precise quantities animals demonstrably remember. This is not the consensus. It is not a fringe position either.

**How much of the classical account is complete?** The Hodgkin and Huxley framework has not been overturned and probably will not be. It is nonetheless acknowledged to be an incomplete description of some things axons do.

**The serotonin question.** This one has escaped the laboratory, so it needs care. In 2022 Joanna Moncrieff and colleagues published a systematic umbrella review in Molecular Psychiatry concluding that the main areas of serotonin research provide no consistent evidence of an association between serotonin and depression, and no support for the hypothesis that depression is caused by lowered serotonin activity or concentrations [59]. A 2024 comment in the same journal, led by Andrew Smith, argued the review was methodologically flawed and attacked a version of the theory that specialists had already abandoned [60]. Both papers exist and both are worth reading. What can be said without taking a side is that the chemical imbalance explanation, in the simple form it usually reaches the public, is not supported by the evidence, and that this is a separate question from whether any particular treatment works. This article makes no claims about treatment and nobody should change anything they are doing on the basis of it.

**And the numbers.** Even the 86 billion figure carries uncertainty. It came from four brains, the range across them ran from about 79 to 95 billion, and a 2025 commentary in Brain revisits how confident anyone should be. It is still far better than the 100 billion neurons and ten times more glia that circulated for decades, which Azevedo's team pointed out had no supporting references at all [1]. If you see those older figures quoted confidently, you are looking at a repeated claim rather than a measured one.

Counting Every Connection

The last few years have added something the field never had: complete wiring diagrams.

In 2024 Sven Dorkenwald and a large collaboration published the neuronal wiring diagram of an adult fruit fly brain, reconstructing 139,255 neurons and 54.5 million synapses between them [61]. Every connection in an entire adult brain, mapped.

The mammalian equivalents are partial but enormous. The MICrONS Consortium published functional connectomics spanning multiple areas of mouse visual cortex in 2025, combining a cubic-millimetre-scale reconstruction with prior recordings of the same neurons' activity, so that structure and function can be compared in the same tissue [62]. And Alexander Shapson-Coe and colleagues reconstructed about a cubic millimetre of human temporal cortex at nanoscale resolution in 2024, containing roughly 57,000 cells and 150 million synapses, in a dataset of about 1.4 petabytes [63]. That fragment turned up things nobody predicted, including rare axons making unusually powerful connections of up to about fifty separate synapses onto a single target cell.

A cubic millimetre. The human brain is roughly 1.2 million cubic millimetres.

The Cost Of Running It

One number puts the rest in perspective. The brain is about two per cent of body weight and consumes about twenty per cent of the body's resting energy.

David Attwell and Simon Laughlin worked out where that goes in 2001, building an energy budget for signalling in grey matter [64]. In their accounting, action potentials took about 47 per cent, postsynaptic effects of glutamate about 34 per cent, maintaining the resting potential about 13 per cent, and recycling glutamate about 3 per cent. Communication is not a side activity of the brain. Communication is most of what the brain spends its energy on.

That budget was then revised by measurement. Henrik Alle, Arnd Roth and Jörg Geiger recorded directly from hippocampal mossy fibre boutons in 2009 and found that mammalian action potentials use only about 1.3 times the theoretical minimum charge, against roughly four times for the squid axon that the original calculations were based on [65]. Real spikes in a real mammalian brain are considerably more efficient than the model predicted, which shifts the balance of the budget further toward synaptic processes.

There is a design logic visible in that. Energy is the constraint, and much of the architecture described in this article, the all-or-none spike, the quantal packet, the myelin sheath, looks like a set of solutions to the problem of moving information without spending more than the body can supply.

Conclusion

Start again from the gap.

Twenty to thirty nanometres, at hundreds of trillions of junctions, crossed constantly by packets of a few thousand molecules each. That much of the story is a century old, was hard won, and is correct.

What has changed is everything around it. The synapse is one channel among several. Neurons also connect directly through gap junctions, in exactly the way Golgi insisted and Cajal denied, so the argument at that 1906 ceremony resolved less cleanly than the textbooks suggest. They broadcast into the extracellular space. They influence each other through fields. And at a great many synapses there is a third cell whose fine processes wrap the junction, sense what is happening, answer, and, on the 2025 evidence, coordinate that answer across many synapses at once.

The strength of every one of those connections is adjustable, on a timescale of milliseconds for the timing rule and hours for the spacing rule. That adjustability is the leading physical account of how anything is remembered, and it is why the way you distribute your effort over time changes what survives.

The honest summary is that we have an excellent description of one channel, a growing description of several others, and complete wiring diagrams for a fly and for roughly a millionth of a human brain. The question of how neurons communicate is not closed. It has been reopened, several times, by people looking at the parts of the picture that had been left out.