TheEngram

In 1953 a surgeon removed most of both hippocampi from a 27-year-old man to stop his seizures. Henry Molaison, known for decades as H.M., lived 55 more years. His intelligence and personality were intact and he could hold a thought for a few seconds. But he could no longer form lasting memories of new events or facts. He saw the psychologist Brenda Milner for decades and never remembered meeting her.

Signals from the cortex arrive through the entorhinal cortex and loop through three relays (dentate gyrus, CA3, CA1), each with its own job, before leaving through the subiculum. The same circuits appear to map time, sequences and even abstract ideas, which may be why imagining a walk through a familiar house (the memory palace) is the oldest memory trick known.

• BIOHippocampus is Greek for seahorse. There is one in each temporal lobe, each a curled strip of cortex about 4 cm long.
• BIOIt acts as an index rather than a store: it binds the scattered pieces of an experience (who, what, where, when) into one pointer you can retrieve later.
• BIOH.M. (Henry Molaison) lost most of both hippocampi in 1953 at age 27. He lived 55 more years, intelligent and himself, but unable to form lasting memories of new events or facts.
• BIOLondon taxi drivers who memorized 25,000 streets had larger posterior hippocampi, and the size grew with years on the job (Maguire et al., 2000).
• BIOAlzheimer’s disease starts here. Tau tangles appear first in the entorhinal cortex, which is why the newest memories are the first to go.
• PHYSThe hippocampus hums at 4–8 Hz (theta). It times each relay: input arrives at one phase of the wave, recall at another.
• PHYSThe CA1 field potential shows theta near 8 Hz with faster gamma cycles nested inside each wave. The trace in step 7 is slowed about 9×.
• BIOEntorhinal cortex: layer II sends the perforant path to the dentate gyrus and CA3; layer III projects straight to CA1. Grid cells live in its medial part.
• BIODentate gyrus: about a million granule cells in a rat fire sparsely, only a few percent at a time, so similar experiences get very different codes. It is one of the few places where new neurons may be born in adulthood, though in humans this is still debated.
• BIOCA3: mossy fiber inputs end in huge “detonator” synapses able to fire a CA3 cell on their own, and recurrent collaterals let a fragment of a pattern reignite the whole. The next chapter simulates this network.
• BIOCA2, a narrow and long-overlooked strip, is now tied to social memory. Silencing CA2 in mice leaves them unable to recognize a mouse they have met before (Hitti & Siegelbaum, 2014).
• CHEMThe Schaffer collateral → CA1 synapse is the most studied synapse in neuroscience; much of what we know about long-term potentiation was measured there.
• BIOThe subiculum sends the processed signal back to the deep layers of the entorhinal cortex and on to the neocortex, and through the fornix to the mammillary bodies and thalamus.
• BIOJohn O’Keefe discovered place cells in the rat hippocampus in 1971. Each fires only in one patch of space, its place field.
• BIOIn 2005 May-Britt and Edvard Moser found grid cells in the medial entorhinal cortex, each firing at every corner of a triangular lattice. O’Keefe and the Mosers shared the 2014 Nobel Prize.
• BIOThe same circuits appear to map time, sequences and abstract ideas, which may be why the memory palace, imagining a walk through a familiar house, is the oldest memory trick known.

In 1982 John Hopfield showed that a network of on/off neurons with symmetric Hebbian links has an energy, borrowed from the physics of magnets:

Every stored memory digs a valley. Recall is the network sliding downhill from a partial or noisy cue into the nearest valley. This idea won Hopfield a share of the 2024 Nobel Prize in Physics.

Pick a memory, erase half of it or add noise, then press Recall. Click or drag on the neurons to draw and store your own picture. Raise the temperature and watch thermal noise melt the memories.

• PHYSN neurons can hold about 0.138 N random memories. Past that the whole network collapses into confusion, a sudden transition like a magnet losing its order. Here: 256 neurons, about 35 memories.
• PHYSOverlapping memories can merge into spurious valleys: blends of two memories you never had. Switch to pure Hebb learning to find some.
• PHYSHopfield networks store every memory’s negative too: flip every neuron of a stored picture and you get another valley with exactly the same energy.
• PHYSRecall here uses Glauber dynamics: one neuron at a time adds up the votes w_ij s_j from its synapses and flips to agree. Each flip can only lower the energy. Temperature makes some flips random, so the state can climb back out of a valley.
• PHYSBy default this network learns with Hebbian learning plus decorrelation (the projection rule), which keeps overlapping pictures apart. “Pure Hebb” uses w = Σ ξ_iξ_j alone, where overlapping memories interfere.
• BIOEvery stored memory adjusts the same 32,640 synapses between these 256 neurons. Nothing is filed in its own drawer; each memory is spread across the whole network.
• BIOIn mice only about 2–6% of dentate gyrus neurons join a given engram. Switching just those cells on with light makes a mouse freeze in fear in a box where nothing bad ever happened (Liu, Ramirez & Tonegawa, 2012).

Alan Hodgkin and Andrew Huxley worked out the physics in 1952 using the giant axon of a squid, half a millimeter thick. Their equations treat the membrane as a capacitor in parallel with voltage-gated sodium and potassium conductances. The oscilloscope above runs those exact equations, live. They won the 1963 Nobel Prize.

Every ion has a voltage at which its chemical urge to diffuse is exactly balanced by the electrical force. That voltage is set by the Nernst equation, where chemistry (concentrations) meets physics (charge and temperature). Raise the outside potassium and the resting voltage creeps up until the neuron can no longer reset.

• PHYSThe membrane is only ~5 nm thick, so 70 mV across it is a field of about 14 million volts per meter. Air breaks down into a spark at about 3 million.
• PHYSVery few ions actually move. About 6,000 ions per square micron flip the voltage by 100 mV, a tiny fraction of the ions nearby.
• PHYSIn the model, each square centimeter of membrane is a 1 µF capacitor with up to 120 mS of sodium conductance, 36 mS of potassium conductance and 0.3 mS of leak, the values Hodgkin and Huxley measured.
• CHEMTetrodotoxin from pufferfish plugs sodium channels. One to two milligrams can kill an adult. Toggle TTX to see why: no sodium current, no spike.
• CHEMThe Na⁺/K⁺ pump spends one ATP to swap 3 Na⁺ out for 2 K⁺ in. Restoring these gradients uses much of the brain’s 20-watt budget.
• BIOMyelin, a fatty wrapping made by glial cells, lets the spike jump between gaps called nodes of Ranvier. It speeds conduction from about 1 m/s to over 100 m/s.

The key molecule is the NMDA receptor, a coincidence detector. It opens only when two things happen together: the sending neuron releases glutamate, and the receiving neuron is already depolarized. At rest a magnesium ion sits in its pore like a cork. Depolarization repels the positively charged Mg²⁺ and lets calcium in.

Calcium decides. A large, fast surge switches on kinases like CaMKII, which add AMPA receptors and strengthen the synapse (LTP). A modest, slower rise favors phosphatases, which remove AMPA receptors and weaken it (long-term depression, LTD). Calcium glows green here because neuroscientists image it with GCaMP, a sensor built from green fluorescent protein.

Try a presynaptic spike alone, then pair it with a postsynaptic burst. Try a 100 Hz tetanus, or 15 slow spikes. Then block NMDA receptors with AP5: LTP vanishes but ordinary transmission carries on.

• BIOIn 1949 Donald Hebb proposed that when one cell repeatedly helps fire another, the link between them grows. In 1973 Tim Bliss and Terje Lømo found exactly that in the hippocampus and named it long-term potentiation, or LTP.
• PHYSGlutamate crosses the 20 nm cleft by pure diffusion in under a microsecond (t ≈ x²/2D).
• CHEMThe fraction of NMDA receptors free of Mg²⁺ follows B(V) = 1 / (1 + [Mg²⁺]/3.57 mM · e^{−0.062 V}) (Jahr & Stevens, 1990). Set Mg²⁺ to zero and a single spike can trigger LTP.
• CHEMCaMKII is a wheel of 12 subunits, two stacked rings of six. Once calcium/calmodulin activates neighbors, they phosphorylate each other at threonine 286 and stay on after the calcium is gone: a molecular switch that remembers.
• BIOA vesicle holds roughly 2,000–4,000 glutamate molecules. Many hippocampal synapses release nothing on most spikes. Watch for release failures.
• BIOAstrocytes wrap the synapse and pull glutamate back in with EAAT2 transporters within milliseconds. Left in the cleft, glutamate kills neurons.
• PHYSSynapse sizes fall into about 26 distinguishable steps, roughly 4.7 bits of information per synapse (Bartol et al., 2015).
• BIOAP5 sits in the NMDA receptor’s glutamate site, while AMPA receptors keep working. In 1986 Richard Morris showed that rats given AP5 could not learn where a hidden platform was in a pool.
• CHEMCalcium is imaged with GCaMP, a sensor built from green fluorescent protein, which is why calcium glows green throughout this page.

Eric Kandel traced this pathway in the sea slug Aplysia and shared the 2000 Nobel Prize in Physiology or Medicine for it. Its master switch is CREB, a transcription factor. Once phosphorylated it recruits enzymes that loosen the DNA’s packaging and read out memory genes.

Run the cascade, then add the inhibitors used in real experiments and see which stage breaks. Click any molecule, in 3D or on the map, for details.

• CHEMPhosphorylation, moving a phosphate group from ATP onto a serine or threonine, is the cell’s main on/off switch. CREB switches on at serine 133.
• BIOArc, a key memory gene, descends from an ancient retrovirus-like element. Arc protein still assembles into virus-like shells that carry RNA between neurons (Pastuzyn et al., 2018).
• CHEMDNA is spooled around histone proteins. Adding acetyl groups loosens the spool so genes can be read. Drugs that block histone deacetylases (HDACs) improve memory in mice.
• BIOSea slugs, flies, mice and people use the same core kit: cAMP, PKA, CREB. It is more than 500 million years old.
• CHEMSynaptic proteins are replaced within days, yet memories last decades. One proposed answer is a self-sustaining tag: the kinase PKMζ bound to KIBRA keeps marking strengthened synapses (Tsokas et al., 2024). How memories outlive their molecules is still an open question.
• BIOZIP, a peptide that blocks PKMζ, erases established late LTP and some memories. But mice lacking PKMζ still form memories, so what ZIP results mean is debated.
• CHEMCalcium is the deciding signal. In a spine it climbs from about 0.1 µM at rest to over 10 µM. Calmodulin grabs four Ca²⁺ ions, changes shape, and switches on kinases and phosphatases.
• BIOThe NMDA receptor opens only with glutamate AND depolarization, which drives out its Mg²⁺ plug: Hebb’s rule built from one protein. AP5 blocks it.
• CHEMCaMKII is a 12-subunit ring. Neighbors phosphorylate each other at threonine 286, so it stays on after Ca²⁺ falls. It is one of the most abundant proteins at the synapse.
• BIOEarly LTP lasts 1–3 hours and uses proteins already present: CaMKII phosphorylates GluA1 (Ser831) and the anchor protein stargazin, trapping more AMPA receptors in the synapse.
• CHEMAt 1 Hz, calcium rises only modestly. Calcineurin, a phosphatase with a higher affinity for Ca²⁺/calmodulin than CaMKII, wins and switches on a second phosphatase, PP1. Together they strip phosphates and pull AMPA receptors out: long-term depression (LTD).
• BIOSynaptic tagging and capture (Frey & Morris, 1997): a stimulated synapse carries a temporary tag for 1–2 hours and captures new proteins made elsewhere in the cell. Untagged neighbors do not.
• BIONovelty releases dopamine into the hippocampus from midbrain neurons. Through D1/D5 receptors it raises cAMP, which can turn a weak, fading potentiation into a lasting one.
• BIOLate LTP needs new genes and new proteins. Actinomycin D blocks transcription and anisomycin blocks translation; given around learning, they spare short-term memory but the memory is gone the next day.
• PHYSReal sizes: a spine head is about 1 µm across, a CaMKII ring about 20 nm, a nucleosome about 11 nm with 147 base pairs of DNA wrapped 1.65 turns. The scene enlarges molecules so they can be seen together.

This is spike-timing-dependent plasticity (Markram et al., 1997; Bi & Poo, 1998). The NMDA receptor again explains much of it. Glutamate binds and lingers for tens of milliseconds. If a spike travels back into the dendrite while glutamate is still bound, the Mg²⁺ cork pops and calcium surges. Reverse the order and the spike arrives at an empty receptor.

• PHYSThe window’s width is set by receptor kinetics: glutamate stays bound to NMDA receptors for tens to hundreds of milliseconds.
• BIOThe calcium story explains the strengthening side. The weakening side also relies on other messengers, such as endocannabinoids released by the receiving cell.
• BIOBecause a neuron that reliably fires first becomes the other’s predictor, STDP lets networks learn sequences: the next note of a song, the next turn on a route.
• PHYSThe curve here follows Bi & Poo (1998): strengthening decays with a time constant of about 17 ms (peak +80%), weakening with about 34 ms (peak −42%). Each value is the change after 60 spike pairings.
• CHEMThe Mg²⁺ block is voltage-dependent. In the model used here, about 4% of the NMDA receptor’s current gets through at −70 mV and about 78% at 0 mV.
• PHYSA back-propagating spike is the receiving neuron’s own action potential travelling from the cell body back into its dendrites. It is what tells each synapse “your cell just fired”.
• CHEMIn the calcium model, sender-first at +10 ms gives about 3.8 times the spine calcium of glutamate alone. Receiver-first at −10 ms gives about 1.6 times.
• BIOMarkram and colleagues (1997) found the effect with paired recordings from connected neocortical neurons. Bi and Poo (1998) mapped the full window in cultured hippocampal neurons.
• PHYSThe window is narrow. At ±50 ms the change is already under 10%, and at ±100 ms it is about 2% or less.

One elegant theory (Lisman & Idiart, 1995) puts items on brain rhythms. Each slow theta wave (about 6 per second) contains several faster gamma cycles (about 40 per second), and each item takes one gamma cycle. About seven gamma cycles fit in a theta cycle, which may be why George Miller found working memory caps out near 7 ± 2 items. Later work puts the true limit nearer four chunks.

• BIOWorking memory is barely written into synapses at all. Neurons in prefrontal cortex keep firing for exactly as long as you need the information.
• BIOMonkeys trained to remember where a light flashed show prefrontal neurons that keep firing through the empty delay until they respond (Funahashi, Bruce & Goldman-Rakic, 1989).
• PHYSTheta runs at about 6 cycles per second, gamma at about 40. 40 ÷ 6 ≈ 7 gamma cycles per theta cycle: in the model, one item each, replayed in order every theta cycle.
• PHYSGamma rhythms come from feedback loops between excitatory cells and fast inhibitory interneurons. The cycle length is set by how long GABA-A receptors stay open.
• BIOGeorge Miller (1956) found that people hold about 7 ± 2 items. Nelson Cowan (2001) argued the true limit is nearer 4 chunks once rehearsal and grouping are ruled out.
• BIOChunking beats capacity. FBICIANASA is ten letters but only three chunks.
• BIOWithout rehearsal, items fade from working memory within 15–30 seconds (Peterson & Peterson, 1959).
• BIODigit span: digits shown one at a time (here 750 ms on, 250 ms off), then typed back in order. Most adults land between 5 and 9.
• PHYSIn the model, capacity is the ratio of the two rhythms: a slower theta or a faster gamma fits more items per cycle. Try the theta slider.

In a slice of hippocampus kept alive in oxygenated salt water, an electrode stimulates the Schaffer collaterals and another records the response in CA1. One weak train of high-frequency pulses boosts the response, but it drifts back to baseline in about three hours. Three strong trains produce a boost that lasts as long as the slice survives. Add anisomycin, which blocks protein synthesis, and the strong trains act like the weak one.

Then there is synaptic tagging and capture (Frey & Morris, 1997). A weak stimulus sets a temporary tag at its synapse. If another synapse on the same neuron gets strong stimulation within about an hour, the new proteins are captured by the tagged synapse too. This may be why a dull detail sticks when it happens next to something important.

• PHYSThe recording electrode picks up the field EPSP: the summed current of many synapses near its tip. The slope of its rising edge is the standard measure of synaptic strength, plotted here as a percent of the baseline before training.
• PHYSA slice is a few hundred micrometres thick, thin enough for oxygen to diffuse in from the bath, so its neurons stay alive and wired for many hours.
• BIOEarly LTP (one weak train) lasts minutes to about three hours. It runs on proteins already at the synapse: kinases switch on and extra AMPA receptors move in.
• BIOLate LTP (three strong trains) lasts as long as the slice survives. It needs new mRNA and new proteins, made in the hours after the stimulus.
• CHEMAnisomycin blocks protein synthesis at the ribosome. Given around the strong trains, it leaves early LTP intact but turns late LTP into the weak, fading kind.
• BIOSynaptic tagging and capture (Frey & Morris, 1997): a weak train sets a temporary tag. Strong stimulation elsewhere on the same neuron within about an hour supplies proteins the tag can capture, and its LTP lasts too.
• CHEMEmotional arousal releases noradrenaline in the amygdala, which boosts consolidation. The beta-blocker propranolol cancels that boost for emotional stories (Cahill et al., 1994).
• BIONovelty releases dopamine into the hippocampus, which helps trigger the protein synthesis lasting LTP needs.
• BIOThe same logic shows up in behaving animals as "behavioral tagging": exploring a new place around the time of weak learning makes that memory last, and the boost needs protein synthesis (Moncada & Viola, 2007).

Three rhythms lock together. Slow oscillations (below 1 Hz) sweep the cortex between active up-states and silent down-states. Sleep spindles (12–15 Hz) from the thalamus ride the up-states. Hippocampal sharp-wave ripples (around 150–250 Hz) nest in the spindle troughs, and each ripple carries a compressed replay of a place-cell sequence.

Deep slow-wave sleep dominates the first half of the night; dreaming REM sleep dominates toward morning. A full cycle through the stages takes about 90 minutes.

• BIOHippocampal cells that fired together while a rat explored fired together again during the sleep that followed (Wilson & McNaughton, 1994).
• BIOThe replay keeps the order of the day but runs it fast: place-cell sequences recur during slow-wave sleep compressed roughly twentyfold (Lee & Wilson, 2002).
• BIODisrupting ripples impairs spatial memory in rats (Girardeau et al., 2009). Boosting the coupling between ripples, spindles and slow waves improves it (Maingret et al., 2016).
• BIOPresent an odor while people learn, present it again during deep sleep, and that memory gets stronger (Rasch et al., 2007).
• PHYSDuring sleep the space between brain cells expands by about 60%, so cerebrospinal fluid can flush out waste such as amyloid-β faster (Xie et al., 2013).
• BIOAfter a night without sleep, people encode new memories about 40% less well (Yoo et al., 2007).
• PHYSEach slow oscillation is a traveling wave. Most start over the front of the brain and sweep toward the back at a few meters per second (Massimini et al., 2004).
• PHYSThe rhythms nest by frequency: slow oscillation below 1 Hz, spindles 12–15 Hz, ripples about 150–250 Hz in rodents (slower in humans, roughly 80–140 Hz). Awake and in REM the hippocampus runs on theta instead, roughly 4–10 Hz.
• CHEMIn REM sleep acetylcholine runs high while noradrenaline and serotonin fall nearly silent, and brainstem circuits block the motor neurons, so dreams are not acted out.

In 1885 Hermann Ebbinghaus memorized thousands of nonsense syllables and measured how fast he lost them. Most forgetting happens in the first day. Each well-timed review flattens the curve. Retrieval strengthens memory most when it takes effort: reviewing just as a memory starts to fade does more than reviewing while it is fresh.

The hippocampus learns in one shot. The neocortex learns slowly, weaving each new memory in among the old ones so nothing gets overwritten. Through nightly replay, a memory’s pieces grow direct links in the cortex and lean less on the hippocampal index. That is why H.M. kept his childhood but lost the years just before his surgery.

• BIOThe spacing effect, better retention from reviews spread out in time than from the same reviews crammed together, is one of the most replicated findings in psychology. It is how flashcard apps schedule your cards.
• BIOForgetting may be partly active. Dopamine signals in flies, and new neurons in the mouse dentate gyrus, appear to erase or overwrite old traces.
• BIOThe division of labor between a fast-learning hippocampus and a slow-learning neocortex is called complementary learning systems (McClelland, McNaughton & O’Reilly, 1995).
• BIOIn 1953 surgeons removed the inner temporal lobes on both sides, including most of both hippocampi, from Henry Molaison (H.M.) to stop his seizures. For the remaining 55 years of his life he formed almost no new memories of events, yet he kept memories from his childhood.
• BIOThéodule Ribot described the gradient in 1881: after brain damage, the newest memories are the most fragile and the oldest survive. It is now called Ribot’s law.
• BIONot everyone agrees the hippocampus is ever released. Multiple trace theory argues that vivid episodes always need it, while the cortex keeps only the gist (Nadel & Moscovitch, 1997).
• BIOAdults recall almost nothing before age 3 or 4. Rapid neurogenesis in the infant hippocampus may rewire the circuits that held those early memories (Akers et al., 2014).
• PHYSThe forgetting model here: recall probability R = (1 + t / 9S)⁻¹ with starting stability S = 0.12 days. A review at recall R multiplies S by 1.3 + 5(1 − R). With no review, recall on day 30 is about 3%; with reviews on days 1, 3, 7 and 14 it is about 81%.

Memories become fragile again each time they are used. In 2000 Karim Nader showed that recalling a fear memory in rats made it unstable: block protein synthesis right after recall and the memory disappeared. This is reconsolidation. A recalled memory has to be stored again, and it can be changed in the process.

Every chapter you touched on this page recruited a few more neurons into this page’s engram, a sparse ensemble among many. In real experiments, mice carry a light-sensitive channel (channelrhodopsin) in only the cells active during learning. Shine 473 nm blue light through an optical fiber and those cells fire, and the memory returns.

• BIOThe list test is the Deese–Roediger–McDermott paradigm. People recalled the never-shown word (here, sleep) about as often as words that were actually on the list, and often felt sure they had seen it (Roediger & McDermott, 1995).
• BIOIn Elizabeth Loftus’s “lost in the mall” study, about a quarter of adults came to remember a childhood event that never happened (Loftus & Pickrell, 1995).
• BIORecalling a fear memory in rats made it unstable: blocking protein synthesis in the amygdala right after recall erased it, while the same block without recall did not (Nader, Schafe & LeDoux, 2000).
• BIOIn 2013 researchers planted a false memory in mice by switching on engram cells for one room while the mice were shocked in another. The mice then froze in the room where nothing had happened (Ramirez et al., 2013).
• BIOLight alone was enough to bring a memory back: dentate gyrus cells tagged during fear learning made mice freeze in a safe room when switched on with light (Liu et al., 2012).
• BIOA memory “lost” to a protein-synthesis blocker given right after learning may still be stored: in mice, switching its tagged cells on directly with light still brought the fear back, though natural cues no longer could (Ryan et al., 2015).
• CHEMNeurons with more active CREB at the moment of learning are more excitable, and they win the competition to join the engram (Han et al., 2007).
• CHEMTagging works through an activity gene: cells that fire during learning switch on c-fos, which (in engineered mice, while the drug doxycycline is withheld) turns on the gene for channelrhodopsin fused to yellow fluorescent protein.
• PHYSChannelrhodopsin-2 comes from green algae. A photon of blue light (about 470 nm) flips its retinal molecule and opens a pore for positive ions within milliseconds, enough to make the neuron fire. Lasers at 473 nm are the standard source.
• BIOOnly a small fraction of dentate gyrus granule cells, a few percent, are active in any one place. That sparseness helps keep memories of similar places apart.