Astrocytes and memory: the brain's second, slower trace

A mouse hears a tone, then feels a mild shock through the floor of its cage. One pairing is enough. Days later the tone alone makes it freeze. This is the most worked over experiment in the memory literature, and the standard way to read it is neuronal. A sparse set of cells in the amygdala and hippocampus was active during the pairing, those cells were physically changed by it, and artificially re-firing exactly that set brings the memory back without the tone. The set is called an engram, and the tools to tag it, silence it and reactivate it have been in circulation for fifteen years.

In October 2025 a group ran the tagging step on a different cell type. Instead of labelling the neurons that switched on during the experience, they labelled astrocytes: the star shaped glial cells that fill the space between neurons and that produce no action potentials at all. What came back was not a smear of activated tissue but a sparse ensemble, distributed across the brain, biased toward the regions where the neuronal engram sits. It was primed by the emotional salience of the experience, still identifiable days later, and re-engaged more strongly during recall than during the original learning. Interfering with it did not erase the memory. It left the memory unstable, which is a stranger and more specific result.

Six months later, in April 2026, Nature published a causal study of the same question in the basolateral amygdala. Calcium activity in local astrocytes tracked the animal’s fear state, and manipulating that activity changed both retrieval of the fear memory and its extinction. The part that matters is what happened upstream: with astrocyte signalling disrupted, the neurons themselves failed to hold the population pattern that normally accompanies retrieval. The claim is not that astrocytes replace the neuronal code. It is that the neuronal code does not assemble properly without them.

Which raises the question this article is about. What is a cell that cannot spike doing inside a memory trace?

The cell that does not spike

Start with numbers, because the folklore is wrong. Glia do not outnumber neurons ten to one in the human brain. Careful cell counting put the two populations in the same order of magnitude overall, with the ratio varying by region. Astrocytes are a large minority of cortical cells, not an overwhelming majority, and the corrected figure has not made them less interesting.

Each astrocyte holds its own territory and tiles with its neighbours, overlapping very little. Inside that territory it extends thousands of extremely fine processes, many of them thinner than the wavelength of visible light, and wraps them around synapses. That three way arrangement, presynaptic terminal, postsynaptic spine, astrocytic process, is called the tripartite synapse. How many synapses a single cell touches depends on species and counting method, and the estimates used in the modelling literature run from tens of thousands in rodents to the order of a million per cell in humans.

What an astrocyte does with those contacts is, uncontroversially, housekeeping of a very active kind. It clears glutamate out of the synaptic cleft after transmission, which is what stops a signal from lingering. It buffers the potassium that leaks out when neurons fire. It supplies metabolic substrate. It regulates local blood flow, which is the reason a functional MRI image exists at all. Every one of those functions gives it leverage over the transmission of every synapse it wraps.

The part that is not housekeeping is signalling. An astrocyte has no voltage spike. What it has is calcium: the concentration inside the cell rises and falls in transients that spread through the fine processes, and pass between neighbouring astrocytes through gap junctions. A neuronal action potential lasts about a millisecond. An astrocytic calcium transient lasts seconds, sometimes tens of seconds. The two cell types are sitting in the same tissue, coupled to the same synapses, running on clocks that differ by four orders of magnitude.

A slow signal is not a weak signal. It is a different kind of variable. A spike is an event, gone the instant it happens, and meaning has to be carried by which cells fire and exactly when. A calcium transient is closer to a value that is set and then held, established by activity in one branch of the cell and still readable elsewhere several seconds later. If a network of these is doing anything computational, it is not competing with neurons on timing. It is doing the thing neurons are worst at: holding a slowly varying context steady across a patch of tissue while the fast layer churns underneath.

Where this stands in 2026

The experimental case has moved quickly, and the reviews have moved with it. In January 2026 Nature Reviews Neuroscience ran a Perspective arguing for astro-neuronal engrams, on the grounds that activity dependent tagging now shows astrocyte ensembles recruited during learning, reactivated at recall, and in some preparations sufficient to drive recall on their own. The authors call the resulting object an astroengram, and the term is doing real work: it names the hypothesis that the trace is a joint object across two cell types rather than a neuronal object with glial support.

The modelling has followed. An April 2026 paper in Communications Biology extends the standard picture of short term plasticity to include an astrocytic term, and reports that the added interaction stabilises recurrent circuits and sharpens their response to input. That is a suggestive result and nothing more, since the model was built to include astrocytes and then found them useful.

Meanwhile the human question sits off to one side, largely untested. Human cortical astrocytes are not scaled up mouse astrocytes. They are more than three times the diameter, carry roughly ten times as many terminal processes, and include morphological types with no rodent equivalent. When human glial progenitor cells were grafted into mouse forebrain, the resulting animals showed faster calcium waves, enhanced long term potentiation and better learning across several tasks. That experiment is from 2013 and it is suggestive rather than decisive, but it is the closest thing available to a test of whether astrocytic complexity buys cognitive capacity.

Why a slow cell might make a good memory

The most concrete proposal for what astrocytes could actually store came from a theory group rather than a bench. In May 2025 a team published a model of neuron-astrocyte associative memory in PNAS, with a single structural move at its centre: stop treating an astrocyte as one computational unit. Treat each of its processes as a unit of its own. A cell with a million processes is then a million small elements that happen to share a cell body, and the calcium level in each one is a variable that can be written and read.

Plug that into the dense associative memory family, the modern descendants of Hopfield networks, where a stored pattern is a stable state the system falls into when given a partial cue, and the capacity arithmetic changes. In their model the number of patterns that can be stored grows faster with size than in a comparable network of neurons alone, because the storage is distributed across processes rather than across cells. The memories live in the pattern of calcium; the read path back to neurons is chemical, through molecules released at the tripartite synapse. A summary from MIT puts the motivation bluntly: human memory capacity looks larger than a neuron-only accounting can comfortably explain.

The same group had already noticed, in a 2023 paper, that a neuron-astrocyte loop can implement the core operation of a transformer block, and that their modelled calcium signal rises and falls almost exactly like the attention signal in the machine version.

That symmetry is worth enjoying and then distrusting. It is an existence proof in mathematics, not a finding in tissue: it shows that a network of this shape could implement such a memory, not that any brain does. And it belongs to a genre with a poor record. The brain has been described as clockwork, as a telephone exchange, as a digital computer, each time using the most impressive machine of the day, and each metaphor was eventually returned with regret. Attention is the impressive machine of this decade.

What breaks

Four things, in ascending order of seriousness.

The write path is contested. Both models above depend on gliotransmission, the release of signalling molecules by astrocytes onto neurons. A serious body of work argues that this is a pharmacological artefact rather than a physiological process, on the grounds that the manipulations used to demonstrate it drive calcium far outside its natural range. The rebuttal published alongside it argues the picture is not black and white and that the negative results turn on the specific transgenic lines used. That exchange dates from 2018 and has not been settled since. Every claim that astrocytes write into a memory trace inherits this dispute.

Necessary is not the same as storing. Silencing astrocytes degrades glutamate clearance, potassium buffering and metabolic supply all at once. Any of those alone would break a circuit. A result showing that a memory fails without astrocytic signalling is fully compatible with astrocytes being permissive infrastructure rather than a substrate holding content.

The resolution does not match the theory. The interesting hypothesis places storage in individual processes. The available tools act on whole cells, or on whole populations of cells. Nobody can currently read out, still less write, a calcium pattern process by process in a behaving animal, which means the specific claim is not yet falsifiable in the specific form that makes it interesting.

Almost all of it is rodent work. The behaviour tested is overwhelmingly fear conditioning, chosen because it is fast, reliable and easy to tag, and it is a narrow slice of what memory does. The cells that would matter most for the capacity argument, human astrocytes, are exactly the ones on which none of these causal experiments can be run.

None of this cancels the shift. What changed between 2025 and 2026 is not that a mechanism was demonstrated but that a question stopped being eccentric. For most of the century the fast, sparse, spiking layer was assumed to be the whole computational story, and everything else in the tissue was plumbing. That assumption is now the thing carrying the burden of proof. It is a useful thing to hold onto when looking at any system built from spikes alone, biological or otherwise: the layer that looks like the computer may just be the layer that was easiest to record from.

Further reading

• Astroengrams: rethinking the cellular substrate for memory, a January 2026 Perspective in Nature Reviews Neuroscience laying out the case and its gaps
• The astrocytic ensemble acts as a multiday trace to stabilize memory, the October 2025 tagging study
• Astrocytes enable amygdala neural representations supporting memory, the April 2026 causal study in Nature
• Neuron-astrocyte associative memory, the capacity model, open access