The Engram: Where a Memory Physically Lives

Jul 27, 2026 · #neuroscience #memory #neurotechnology

A memory is a physical object: a sparse pattern of neurons that can be tagged, silenced, switched back on, and rewritten every time you recall it.

The Engram: where a memory physically lives — An engram is not a place. It is a sparse subset of neurons whose connections were strengthened together.

In 1904 the German biologist Richard Semon coined a word: engram. His claim was simple and, for the time, reckless. An experience leaves a persistent material trace in the nervous system, and remembering is the reactivation of that trace by a fragment of the original situation. He called that reactivation ecphory. Almost nobody listened. The word survived; the idea was dismissed as mysticism.

Three decades later the American psychologist Karl Lashley decided to test it with a scalpel. He trained rats to run a maze, then removed pieces of cortex, region by region, and re-measured performance. His conclusion, published in 1950 in a paper that became famous under the title In Search of the Engram, is a magnificent admission of defeat: no local lesion erased the memory. Only the amount of tissue removed mattered, never the location. He wrote, with weary humour, that his results almost forced him to conclude that learning is simply not possible.

Lashley’s measurements were not wrong. He was looking for a place: a drawer, a slot, a file. An engram is not a place. It is a pattern: a sparse subset of neurons, spread across several regions, whose connections to one another were strengthened at the same time. Slicing a brain into regions to find a distributed pattern is like cutting up a newspaper to locate a sentence that has already been reprinted on ten different pages.

The three proofs you have to deliver

It took until 2012 and the arrival of optogenetics to turn the engram from a philosophical hypothesis into an experimental object. The field settled on a strict specification, and that specification is what makes the modern claim so strong. To say “I have found the engram of this memory”, you must show three things.

Persistence. You must identify the neurons that were active at the moment of encoding and show that they carry a lasting physical change: potentiated synapses, new dendritic spines.

Necessity. If you silence those specific neurons at the time of recall, the memory must fail to come back, while the rest of the animal’s behaviour stays intact.

Sufficiency. If you switch them back on artificially, with no external cue whatsoever, the animal must behave as though it remembers. This is the criterion that changed everything, because it is no longer correlational. It is causal.

How you catch a memory as it forms

The technique deserves a moment, because it is unusually elegant. When a neuron fires strongly it triggers the expression of an immediate early gene, the most famous being c-Fos. You take the promoter of that gene, the switch that flips on when the neuron discharges, and instead of the normal gene you have it drive the production of an opsin, a protein that makes the neuron respond to light.

One problem remains. If the construct is permanently active, you tag every neuron that fires over three weeks, which is to say everyone. Hence the tagging window. The animal is kept on doxycycline, which blocks the system. You withdraw the doxycycline a few hours before the experiment and the window opens. The animal lives through the event to be memorised, and only the neurons active at that moment equip themselves with opsin. Put the doxycycline back and the window closes; the tag is frozen. Days later, an optical fibre and a flash of blue light are enough to reactivate exactly that group.

The tag is not a marker you place on a memory. It is a time window you open. The amber neurons are the ones that were firing during the event, and days later those are exactly the ones natural recall reactivates.

Who gets recruited, and why them

Here is the part that surprises everyone. Recruitment is not determined by the content of the memory. It is determined by a competition. At the moment an event occurs, the neurons of a given region are not in identical states: some are transiently more excitable than others, typically those with a high level of the transcription factor CREB. Those win the race and are allocated to the engram. The others, which would have encoded the same information just as well, are excluded.

Put bluntly: the set of neurons that carries the memory of your wedding could have been a different set. It was selected by a transient excitability state, not by any semantic affinity. And that competition has a direct, measurable consequence. Two events experienced a few hours apart share a substantial fraction of their neurons, because excitability has not had time to reset, so the two become linked and recalling one recalls the other. Two events separated by a day recruit disjoint ensembles and stay independent. The co-allocation window is a clock that decides which memories get wired to which.

Recruitment is also strikingly sparse, on the order of a few per cent up to around twenty per cent of the neurons in a region depending on protocol. That is not a technical limitation. It is the solution to the capacity problem. The number of distinct patterns you can form by switching on k neurons out of n is the binomial coefficient, and the intuition behind the algebra is worth more than the algebra. Picture a wall of 100,000 switches, and for each memory you raise 5,000 of them. The number of distinct configurations you can build that way exceeds, by an enormous margin, the number of atoms in the observable universe. Turning on few neurons, but different ones, buys astronomical capacity and patterns that do not tread on each other. The brain does not make room by erasing. It makes room by being sparse.

The memory that is there but cannot be found

The most unsettling result in the field is the silent engram. You make an animal amnesic for a specific memory by blocking protein synthesis right after learning, which prevents consolidation. The classic outcome: no natural cue brings the memory back. The classic conclusion: it was never stored.

Except that if you optogenetically switch on the neurons tagged at encoding, the memory comes back perfectly. It was there. It was not accessible.

This dissociation is worth sitting with, because it separates two things that were routinely conflated: is the information stored? and can the retrieval circuit find it? What the amnesic animal lacked was not the trace, but the synaptic strengthening of the inputs to the engram, the strengthening that lets a partial cue from the real world reignite the full pattern. The trace is the book; consolidation builds the index. The amnesia studied here is an index failure, not a burned book. The same logic is now being probed in early models of Alzheimer’s disease, where forced reactivation of tagged engrams restores memories presumed lost. The caveats are large, these are rodents and contextual fear memories rather than human autobiographical ones, but the conceptual frame has shifted: “lost” is no longer a synonym for “destroyed”.

Remembering is rewriting

The second big idea is reconsolidation. When a consolidated memory is recalled, it is not simply read out. Recall destabilises it: for a window of a few hours the trace becomes labile again, dependent on fresh protein synthesis to be restabilised. During that window it can be strengthened, weakened, or altered by whatever is happening at the time.

Recall opens a window during which the trace is soft. What restabilises is not a copy of the original but a new version, contaminated by the context of the recall.

Autobiographical memory, then, is not an archive. It is a palimpsest. The memory you are most confident about, the one you have told a hundred times, is statistically the most distorted, because you no longer remember the event but your last retelling of it. Subjective confidence and fidelity are largely uncorrelated variables. It is also why the most promising clinical protocols for post-traumatic stress do not try to erase a memory but to intervene inside the labile window to blunt its emotional charge while leaving it intact as information.

This inverts the usual intuition about learning. A memory you reread passively stays on its shelf. A memory you retrieve with effort enters the labile window and restabilises stronger. Spaced retrieval is not a study trick layered on top of biology; it is the direct exploitation of a biological rewriting mechanism. Testing yourself does not measure your memory. It modifies it.

Building a memory that never happened

If necessity and sufficiency hold, one consequence follows logically: it should be possible to construct a memory. It has been done. The protocol: tag the neurons of a completely neutral context A, an unremarkable cage where the rodent spends a quiet moment. The next day, place the animal in a different context B, deliver an unpleasant shock, and during that shock, use light to reactivate the context A neurons. The brain associates the fear with context A. Returned to A, where nothing bad ever happened to it, the animal freezes.

This is not confusion. It is a functionally complete memory whose content has no referent in the world. Lived experience and remembered experience have just been prised apart experimentally.

Where this stands in 2026

Three current threads are worth knowing about, and one caution.

An engram is turning out not to be a single block. A study in Nature Neuroscience used calcium imaging to tag neurons in hippocampal CA1 with far finer temporal resolution than classical tagging allows, slicing one fear-learning episode into successive phases. It identified four non-overlapping populations recruited at different moments of the same conditioning, and only two of them, when optogenetically reactivated, were sufficient to bring the memory back. Until now the field treated “the engram of event X” as one entity. This fragments it: an engram has internal structure, with load-bearing components and passengers. For any ambition of reading or transferring a memory, that moves the target. Catching the active neurons is not enough; you have to catch the right ones.

The synapse is moving back to the centre of the debate. A 2026 review in the Journal of Neurochemistry surveys “engram synapses” and the dynamics of synapse formation and elimination during learning. The functional correlate of memory is not a neuron’s membership in an ensemble but the potentiation of a specific subset of its synapses, often grouped in clusters on the same dendritic branch. That is a change of scale that makes copying dramatically harder: the unit of storage is not the neuron, of which a human has on the order of 10^11, but the weighted, positioned synapse, of which there are on the order of 10^14 to 10^15. Three orders of magnitude, to be measured as weights and not merely as presence.

Updating an old memory recruits a new engram. Work published in Neuron shows that revising a remote memory does not edit the original trace: the hippocampus recruits a fresh ensemble that receives the current context, incorporates it, and carries the revised version. The mental image to discard is in-place editing. The brain works by successive versions, with the most recent one taking over recall. Memory resembles a stack of layers more than a file you correct, and you only ever consult the top layer.

The caution concerns writing rather than reading. A long line of work on the hippocampal prosthesis uses a non-linear multi-input multi-output model: learn to predict CA1 activity from CA3 activity, then stimulate during encoding to impose the “correct” pattern. Tested in epilepsy patients already implanted for clinical reasons, stimulation during the encoding phase improves recall performance. Handle the numbers carefully: cohorts are very small, on the order of a dozen or so subjects, the population is not representative, and the improvement is on laboratory tasks rather than everyday memory. But the proof of principle is there. We are no longer only reading the engram.

What this does to the idea of copying a mind

If a memory were merely a graph of connections, a sufficiently fine connectome would suffice to carry it. What we have just seen says otherwise. A memory is: a sparse subset of cells; the precise weights of the synapses between them; the weights of the input synapses that let a cue reignite it, the index, without which it is silent; an intrinsic excitability state that decided recruitment and still governs its links to neighbouring memories; and a dynamic object, altered by every act of reading.

A static map of connections captures, at best, the second item. The rest is chemistry and state, not topology. A connectome scan would yield a library without an index, full of books nobody could open any more. This is not an argument against the possibility in principle. It is an argument about what would actually have to be measured, and about how far there is left to go.