Take a slime mold, a single cell the size of a dinner plate with no neurons anywhere in it, and make it cross a bridge coated in quinine. It hates quinine. It slows down, hugs the edges, and takes a long time about it. Do this every day. By roughly the sixth day it crosses at normal speed, having apparently concluded that the quinine is unpleasant but survivable. Leave it alone for two days without quinine and the caution comes back, on a schedule. Then take that trained cell and fuse it with a naive one, which slime molds do routinely as part of ordinary life, and the naive cell crosses the bridge at speed on its first attempt.
A memory was formed, maintained, allowed to decay, and transferred to another organism by merging bodies. There is no nervous system anywhere in this story.
The reflexive response is to call that surprising, and it is worth resisting, because the surprise depends entirely on an assumption that will not survive the rest of this. Memory is not a thing brains do. Memory is a thing matter does when it is organized a particular way, and brains are one implementation, arrived at late, by one branch of the tree. Studying memory without a brain is not a tour of exceptions. It is the only way to see what the general problem actually is, and what the neural solution specifically bought.
It is also the part of comparative neuroscience with the worst signal-to-noise ratio, which is the other reason to work through it carefully. The subject attracts overclaiming from one direction and reflexive dismissal from the other, several of its foundational results were wrong, and the field has recently spent considerable effort auditing its own evidence. Sorting the solid cases from the discredited ones is most of the work, and the solid cases are more interesting than the discredited ones ever were.
What memory without a brain actually requires
Strip the concept to its load-bearing parts and it needs three things. Some internal state has to change as a result of experience. That change has to persist after the experience ends. And the persisting change has to alter what the system does later.
That is the whole specification. Nothing in it mentions neurons, synapses, brains, or consciousness. Anything with a state variable that can be written, held, and read will satisfy it, which is why memory keeps turning up in places that make people uncomfortable.
Discipline matters here, though, because the specification is loose enough to admit things nobody should count. Fatigue satisfies it, technically: a muscle that has been worked is in a different state and behaves differently. So does damage. So does a rusty hinge. The distinctions the field uses to separate memory from mere state change are whether the response can recover, whether it is specific to the stimulus rather than general, and whether the system still responds normally to other things while ignoring the one it has learned about.
That last criterion is the one that does the work. In habituation, the simplest genuine form of learning, an organism stops responding to a repeated harmless stimulus while remaining fully responsive to novel ones. That rules out exhaustion, since an exhausted system cannot respond to anything, and it rules out damage, since damage is not stimulus-specific. Habituation also shows spontaneous recovery after a rest period, which is a signature of memory rather than depletion, and it shows dishabituation, where a strong novel stimulus restores the original response.
Those criteria are old, precise, and testable in a petri dish. Run them across the tree of life and the results are not close.
There is a further distinction worth carrying, between memory that is merely persistent and memory that is retrievable. A scar is persistent state caused by experience and it changes future behavior, in the sense that scarred tissue behaves differently. Nobody calls it a memory, and the reason is that nothing in the organism can consult it as information. The question of whether a given trace is being read as information or is simply a lingering physical consequence turns out to be the hardest question in this entire subject, and most of the disputes in the literature on memory without a brain are versions of it.
The slime mold, and memory written in plumbing
Physarum polycephalum is a single cell containing many nuclei, spread into a fan of interconnected tubes that can cover a square meter. It solves mazes by connecting food sources with the shortest viable path. Given oat flakes arranged like the cities around Tokyo, it produces a network with efficiency and fault tolerance comparable to the actual rail system, which is a result that has been reproduced enough times to stop being a novelty. Worth noting what that result does and does not show: the mold is not planning a rail network, it is running a local rule about reinforcing tubes that carry flow and pruning those that do not, and the global efficiency falls out of the local rule. Impressive optimization does not imply an optimizer.
The question that stayed open for two decades was where the information sat. There was no obvious candidate. Then a 2021 study proposed a mechanism that is elegant precisely because it is so physical: the memory is in the diameters of the tubes.
The proposal works like this. When part of the network contacts food, a softening agent is released and propagates through the tubes. Tubes carrying more of the flow soften more and dilate; tubes carrying less contract. Since flow rate depends on diameter and diameter now depends on past flow, the network’s architecture becomes a record of where nutrients have been encountered, with the hierarchy of tube thicknesses encoding both the location and something like the significance of past food sources. When the organism later needs to decide which way to grow, the pre-existing thickness distribution biases the outcome. The memory is not stored in the cell. The memory is the shape of the cell.
Physarum does other things that look like more than plumbing, and they deserve a mention with the appropriate hedging attached. It anticipates periodic events, slowing its movement in advance of a temperature drop it has experienced at regular intervals, and it continues doing so for a while after the drops stop, which implies an internal oscillator being entrained rather than a simple reaction. It also makes choices that violate rational-choice axioms in the same directions animals do, showing context-dependent preferences where adding an inferior third option changes the ranking of the first two. Whether that reflects anything worth calling decision-making or is a straightforward consequence of how competing chemical gradients resolve is unsettled, and the second explanation has not been ruled out.
That is a genuinely different architecture from anything neural, and it comes with different properties. It is slow to write and slow to erase. It is inseparable from the body, since the body is the storage medium. It cannot be read out independently of acting on it. And it degrades gracefully, because a partially destroyed network still carries the thickness distribution in whatever remains.
The comparison worth making is to a river system rather than to a computer. A watershed’s channel geometry is a record of past flow, written by the flow itself, and it determines where future water will go. Nobody would call a river network intelligent, and the mechanism the slime mold appears to be running is closer to that than to anything in a textbook on synaptic plasticity. What makes it memory rather than mere erosion is that the organism can read the pattern and act on it, and that the writing is coupled to something the organism cares about rather than to whatever happens to flow downhill.
It is also contested, which is the appropriate state for a three-year-old mechanism. A published critique in the same journal argued that the observations are better explained by ordinary chemotactic and hydrodynamic responses without invoking memory at all, and that the word is doing rhetorical work the data do not support. That objection has not been resolved and should travel with the finding.
Habituation, scheduled forgetting, and memory by fusion
The behavioral work on Physarum is older than the mechanistic work and less disputed.
Slime molds habituate to repellents. Presented with quinine or caffeine in a concentration that is aversive but not lethal, they initially avoid crossing and eventually cross at full speed. The habituation is specific: a mold habituated to quinine still avoids caffeine, which rules out general desensitization and satisfies the stimulus-specificity criterion. Recovery occurs after a rest period of a couple of days, and the timing is consistent enough to describe as a forgetting curve. Work on the mechanisms underlying memory formation and preservation in slime moulds has pointed toward the absorbed substance itself acting as part of the trace, with the organism taking up the repellent and its internal concentration serving as the state variable.
The transfer result is the one that has no analogue in neural systems. Fuse a habituated mold with a naive one and the naive one behaves as though habituated. Fuse one habituated mold with several naive ones and the behavior still transfers, up to a ratio. This is memory moving between organisms by physical merger, which is possible only because the storage substrate is a diffusible chemical state rather than a wiring pattern, and it is a decent illustration of what a substrate choice buys you. No vertebrate can hand another vertebrate a memory by touching it. A slime mold can, because its memory is made of a thing that mixes. That is the clearest single demonstration in this whole subject that the properties of a memory system are properties of its material, not of its owner’s sophistication.
The cost side is equally instructive. A memory made of concentration cannot be selectively erased, cannot be indexed, cannot store two unrelated facts without them interacting, and cannot be recalled without being acted upon. Capacity is roughly one thing at a time, and the retention interval is measured in days.
That constraint is worth taking seriously rather than treating as a limitation to be apologized for. An organism whose entire behavioral repertoire is grow toward good things and away from bad things does not need to store two unrelated facts. Matching storage capacity to behavioral requirements is what an efficient system does, and a slime mold carrying a hippocampus would be paying for an instrument it has no use for. Memory without a brain is not impoverished memory. It is memory sized to the problem.
Memory kept outside the body entirely
The strangest result in the Physarum literature is that a substantial part of its spatial memory is not inside it.
As it moves, the organism leaves a mat of extracellular slime behind. Given a choice, it avoids areas already covered in its own slime and preferentially explores fresh substrate. That single rule turns the environment into a record of where the organism has already been, and it is enough to solve a class of navigation problems that would otherwise require internal spatial memory. Deprived of the ability to detect its own trails, molds perform substantially worse in mazes with dead ends, because they re-enter the same blind alleys repeatedly.
This is memory with the storage medium located outside the organism, and it is not rare once you look for it. Ant pheromone trails are the textbook case: no individual ant holds the route, the route is written in evaporating chemical on the ground, and the colony’s collective decision emerges from many ants reading and reinforcing a shared external substrate. Termite mound construction runs on the same principle, with each deposit of material changing the stimulus field that determines where the next deposit goes.
The general term is stigmergy, coordination through modification of a shared environment, and it dissolves a distinction people treat as obvious. A memory in a nervous system and a memory in a pheromone gradient differ in durability, precision, and privacy. They do not differ in kind. Both are persistent state changes caused by experience that bias later behavior, and the colonies whose collective foraging decisions emerge from exactly this kind of distributed record are running an information system whose storage is partly social and partly environmental rather than wholly inside any skull.
Externalized memory also has a property none of the internal systems have, which is that it survives the death of the individual holding it. A pheromone trail outlasts the ant that laid it. A worn path outlasts the animal that wore it. Human writing is the extreme case of the same trick and the reason it changed everything, and the cooperative hunters whose territories are marked and re-marked across generations are using a low-bandwidth version that persists past any individual’s lifespan. Whenever a system offloads memory to the environment it trades privacy and precision for durability and shared access, and that trade shows up identically in an ant colony and in a library. The bowerbird that builds and repeatedly adjusts a decorated structure is doing a version of it too, since the bower holds a record of the builder’s accumulated effort that neither the bird nor its audience has to remember internally.
Memory without a brain, but with neurons
Between the neuron-free organisms and the brained ones sits an informative middle case, and it produced one of the more consequential results of the past few years.
Cnidarians, which include jellyfish, corals, and sea anemones, have neurons organized into diffuse nerve nets with no central processing organ. They are the sister group to everything bilaterally symmetrical, which makes them the closest thing available to a window on what nervous systems were like before centralization.
In 2023 a team demonstrated operant conditioning in the Caribbean box jellyfish. The animals hunt copepods among mangrove prop roots and must avoid colliding with the roots, which they judge visually using their rhopalia, clusters containing image-forming eyes distributed around the bell. The experimenters placed jellyfish in tanks with painted stripes simulating roots and manipulated contrast. At low contrast the animals misjudged distance and bumped into the walls. Within a few minutes and a handful of collisions, they increased their average turning distance by roughly half and reduced contacts substantially. Combining a visual cue with the mechanical stimulus of a collision produced the association; neither alone did.
That is associative learning in an animal with about a thousand neurons per rhopalium and no brain. The same year, associative learning was reported in the sea anemone Nematostella vectensis, which is sessile and has an even more diffuse nervous system.
The implication is chronological. If cnidarians and bilaterians both do associative learning, then either the capacity arose independently twice, or it was present in their last common ancestor, which lived something over six hundred million years ago in an animal with a nerve net and nothing resembling a brain. Learning is older than centralization by a wide margin, and centralization was a later optimization on a capacity that already existed.
The jellyfish result also clarifies what a brain is for by showing what you can do without one. The box jellyfish learned a specific visual-mechanical association in minutes, retained it, and applied it to navigation. What it cannot do is transfer that learning to a different context, form associations between arbitrary stimuli with no natural relationship, or hold more than a small number of such associations at once. The rhopalia appear to handle their own learning locally rather than pooling it, which means the animal may be running several small independent memories rather than one shared one. Distributed, local, fast, and narrow is a coherent design, and it is what memory without a brain looks like once neurons are available but centralization is not. The animals whose nervous systems distribute most of their neurons away from any central organ sit at the other end of the same design axis, with enough centralization to coordinate and enough distribution to keep the local work local.
The immune system is a memory system
Here is the largest and best-characterized non-neural memory system on Earth, and it is routinely left out of discussions of memory because it lives in a different department.
Adaptive immunity works by clonal selection. The body maintains an enormous repertoire of lymphocytes with randomly generated receptors. When a pathogen appears, the few cells whose receptors happen to bind it proliferate massively, and a subset persists afterward as memory cells. On re-exposure the response is faster, larger, and higher-affinity. Antibody responses to some pathogens persist for decades, and in the case of certain infections, essentially for life, which is a retention interval no neural memory reliably matches.
Run that against the specification. Internal state changes with experience: the clonal composition of the lymphocyte population is permanently different. The change persists: memory cells survive for years. It alters future behavior: the secondary response differs dramatically from the primary.
The storage substrate here is population structure. The information is not in any one cell; it is in which cells exist and in what numbers. The immune system also solves a problem neural memory never has to face, which is that it must remember things it has never encountered. The randomly generated receptor repertoire exists before any infection, which means the system is pre-loaded with candidate responses to pathogens that do not yet exist. Nothing about memory without a brain requires the memory to be written after the fact; here the writing consists of selecting from possibilities already present, which is a mechanism with no neural analogue at all.
That gives it properties no neural memory has. It is enormously specific, distinguishing molecular differences a neural system could never resolve. It is content-addressable in the most literal way, since retrieval happens by the antigen physically binding its match. It is distributed with no addressing problem. And through affinity maturation, in which memory cells undergo further mutation and selection, the stored representation actually improves after storage, which is not something a synapse does.
Innate immunity was long assumed to have no memory at all, and that turned out to be wrong too. Trained immunity describes durable functional reprogramming of innate cells and their bone marrow progenitors after certain exposures, mediated by epigenetic and metabolic changes rather than by receptor rearrangement, producing altered responses to unrelated pathogens months later. That is memory stored in chromatin state.
The immune comparison is the one that most usefully disciplines thinking about memory without a brain, because nobody disputes it. There is no argument in immunology about whether immunological memory is real memory, no philosophical hand-wringing about whether a B cell truly remembers. The field simply defined its terms operationally and got on with it. That the identical operational definition applied to a slime mold generates decades of argument says more about which organisms people are prepared to grant the word to than about any difference in the underlying phenomenon.
CRISPR is the most literal memory in biology
Bacteria and archaea maintain, in their own genomes, an ordered archive of the sequences of viruses that have previously attacked them.
The CRISPR system captures short fragments of invading viral DNA and inserts them into an array in the host chromosome, separated by repeats. New spacers are typically added at one end, which means the array preserves chronology: the order of entries reflects the order of infections. Transcripts of these spacers guide nucleases to matching sequences, so a subsequent infection by the same virus is recognized and cut.
This is a written record, in a durable medium, of specific past events, indexed in temporal sequence, used to guide future action. It is heritable, passing to daughter cells, which means an individual bacterium can be born already remembering an attack that happened to an ancestor.
The reason this matters beyond the genome-editing applications everyone knows about is what it demonstrates about substrate. DNA is an information storage medium with capacity, stability, and copy fidelity that no neural tissue approaches. What it lacks is speed: writing a spacer takes an infection event, and reading it out takes transcription and translation. Neural memory is fast and lossy. Genomic memory is slow and exact. Neither is better; they are solving different problems with different physics, which is the pattern this whole subject keeps producing.
There is a second lesson in the CRISPR case about what makes a record useful. The array is ordered, which means it carries not just what happened but roughly when, relative to everything else. Temporal structure is expensive to maintain in most substrates and it is what allows a record to support inference rather than mere recognition. Most examples of memory without a brain store recognition only, which is why the ordered CRISPR array stands out. Neural systems achieve it through sequence replay and through cells that encode elapsed time. A bacterium achieves it by appending to one end of a list. Both are solving the problem that raw associations without order are much less informative than ordered ones.
Plants, and the memory of a winter
Plants have no neurons, and they have several genuine memory systems that are well characterized at the molecular level.
Vernalization is the cleanest. Many plants must experience prolonged cold before they will flower, which prevents them flowering in a warm autumn spell and then being killed. In the model plant Arabidopsis, a gene called FLC represses flowering, and extended cold progressively silences it through the accumulation of repressive histone modifications at the locus. The silencing persists through subsequent cell divisions and through the return of warm weather, so the plant behaves in spring according to how cold the winter was. It is reset in the next generation, which is itself a designed feature rather than a limitation, since a plant that inherited its parent’s winter would flower on the wrong schedule. That is a durable, quantitative, environmentally acquired record held in chromatin, and it satisfies every criterion in the specification.
Vernalization also has a quantitative property worth flagging, which is that the plant is not registering cold as a binary. The degree of silencing scales with the duration of the cold period, so the plant emerges from winter with something closer to a measurement than a flag, and it flowers accordingly. A memory without a brain that stores a magnitude rather than a fact is doing more than most people assume such systems can.
Defense priming is the second. A plant attacked by a pathogen or herbivore in one part of the body mounts faster and stronger defenses everywhere afterward, sometimes for weeks, through a combination of chromatin changes and accumulated signaling intermediates that sit primed but inactive until needed.
The Venus flytrap does something closer to counting than to memory but belongs in the same conversation. Trigger hairs inside the trap generate action potentials, propagating electrical signals in tissue with no neurons in it, and a single touch does nothing. Two within roughly twenty seconds close the trap. Further stimulation from a struggling insect drives secretion of digestive enzymes, with the amount scaling with the number of triggers. The mechanism appears to run on accumulated cytosolic calcium that decays between events, so the plant is integrating a signal over time against a threshold and a leak rate, which is functionally a short-term memory with a defined decay constant and no neurons involved. The trap is running an integrator with a leak, which is the same computation a neuron performs at its membrane and which the plant implements with calcium and time rather than with sodium and voltage.
Bioelectricity, and memory that survives losing your head
Planarian flatworms can be cut in pieces and each piece will regenerate a complete animal, including a new brain.
In experiments that deserve their reputation, planarians were trained on a task, decapitated, allowed to regenerate a new head over about two weeks, and then retested. They showed savings, relearning faster than untrained controls that had been through the same decapitation. Something about the training survived the removal of the organ that had done the learning.
The candidate explanation involves bioelectric state. Cells maintain membrane voltages and are electrically coupled through gap junctions, producing tissue-level voltage patterns that carry positional and patterning information. Manipulating those patterns in planarians can produce animals that regenerate two heads, and the altered pattern can persist through subsequent rounds of cutting even though the genome is unchanged, which is an inheritable anatomical memory held in a physiological rather than genetic medium.
Appropriate caution applies. The savings effect is real but modest and has a difficult history, since this exact organism was at the center of the field’s most notorious debacle. The bioelectric account is a hypothesis with substantial supporting work on patterning and thinner direct evidence connecting it specifically to behavioral memory. The honest position is that something persists, that the most plausible substrate is not synaptic, and that the mechanism is not established.
What makes the planarian case worth including despite the uncertainty is the question it forces. If a memory can survive the destruction and rebuilding of the organ that formed it, then the memory was never only in that organ, which means either the trace is distributed through tissue that was not removed, or the regenerating brain is being rebuilt according to a template that itself carries information. Both possibilities are strange, and both are testable, and neither is what anyone would predict from a purely synaptic account.
The claims that do not hold up
An audit, because this subject has a worse track record than most.
The planarian cannibalism experiments are the field’s cautionary tale. In the 1960s, researchers reported that untrained planarians fed the ground-up bodies of trained ones acquired the training, and the finding launched a search for a chemical memory molecule. It did not replicate reliably, the effects were plausibly attributable to slime trails and handling artifacts, and the episode set back serious work on non-neural memory by a generation. The specific damage is worth naming: for decades afterward, proposing that memory might exist outside a nervous system was a reputational risk, which meant the genuinely solid cases in immunology and plant biology were studied by people who never described their work in those terms and never talked to each other. A modern RNA-transfer result in sea slugs revived the idea in a more careful form, showing transferred sensitization, and it remains interesting and unconfirmed as a general mechanism.
Single cells learn is a claim the field has itself pushed back on. A prominent reassessment argued that much of the cited evidence for learning in single cells fails on methodological grounds, that many reported effects are consistent with simple adaptation or with experimental artifacts, and that the field has been insufficiently rigorous about controls. That critique is from within the community rather than from outside it, and it applies with real force to the more excitable end of the literature.
The slime mold memory mechanism is contested, as noted, by a published objection arguing the observations need no memory to explain them.
The wood-wide web is the most oversold claim in plant science. Mycorrhizal fungi do connect plants, resource transfer does occur, and the leap from there to forests as cooperative information networks with mother trees deliberately nurturing offspring is unsupported. Reviews by researchers in the field have found that the most-repeated claims are not backed by the cited studies, that field evidence for adaptive resource transfer between trees is weak, and that the popular version substantially misrepresents what has been shown.
Plant neurobiology as a discipline overreached, and the pushback was severe enough to be published as an open letter from plant scientists. Specific results have failed to replicate, including a widely covered demonstration of associative conditioning in pea plants that a subsequent attempt could not reproduce.
Cellular memory in transplant recipients, the idea that organ recipients acquire donor preferences and personality traits, has no mechanism and no controlled evidence, and it recurs because the anecdotes are compelling.
Water memory, the claim underlying homeopathy that water retains an imprint of substances once dissolved in it, fails on physics rather than on biology. Hydrogen bond networks in liquid water reorganize on picosecond timescales, which forecloses any structural trace persisting long enough to matter. It belongs in this audit because it is the reductio of the whole subject: memory does require a substrate that can hold a state, and not everything can.
What brains actually added
If memory is this widespread, the interesting question inverts. It is not why so many things remember. It is what a nervous system was for.
Line up the substrates and the tradeoffs are legible. Tube diameters are durable and slow and inseparable from the body. Chemical concentrations are fast to write, uncopyable, and hold roughly one thing. External trails have unlimited capacity and no privacy and decay on the environment’s schedule. Chromatin marks are stable across cell divisions and take hours to write. Clonal populations are astonishingly specific and take days to mount. Genomic spacers are exact, heritable, and require an infection to write.
Against that list, what neural memory offers is a specific combination that none of the others achieves. Writing takes milliseconds. Capacity is enormous and the items do not have to interact. Arbitrary associations can be formed between things with no physical or chemical relationship to each other, which is the capability that unlocks essentially everything a large animal does. And crucially, the memory can be read without being acted on, which is what makes planning possible: a system that can consult a record without committing to a behavior can evaluate options.
That last property is the real invention. A slime mold cannot consider a route without growing down it. The rodent that sweeps its place cell activity down one maze arm and then the other before choosing is doing something no chemical gradient can do, and the ravens that select a tool for a job seventeen hours away are exercising exactly that decoupling of retrieval from action. It is also, not coincidentally, the property that makes deception possible, since an animal that can consult what another animal knows without acting on it can act on something else instead.
Brains did not invent memory. They industrialized it, and the specific gains were speed, capacity, arbitrariness, and the ability to look something up without doing anything about it.
That list is also a decent specification for anyone trying to build memory deliberately. The engineering efforts to read and write to nervous systems are attempting to interface with a substrate optimized for exactly those four properties, and the difficulty they encounter is a direct consequence of what makes the substrate good: fast, distributed, arbitrary associations are hard to address from outside precisely because nothing about them is laid out in a fixed physical order. A slime mold’s memory, by contrast, could be read with a ruler. The systems being built to store and retrieve information without any biological substrate at all face the mirror-image problem, having near-perfect addressability and no obvious way to decide what is worth keeping.
Which reframes almost every argument in comparative cognition. The great apes maintaining tool traditions across generations, the parrots solving problems nobody set for them, the elephants carrying decades of water-source knowledge, the cetaceans maintaining vocal traditions, the macaques whose innovations spread through a troop and the birds whose song dialects encode where they were raised are all running variations on a capability that predates neurons entirely. What varies is the substrate and its tradeoffs, not the presence or absence of something magical.
The 24-lecture Neurozoology course works the tree of life on that basis throughout, alongside the first edition’s survey of nervous systems, the study of how knowledge moves between animals, and the working animals whose capacities were discovered by the people depending on them. The long-distance migrants whose routes must be learned and can be lost and the fish populations whose migratory knowledge disappeared with the individuals holding it are reminders that a memory system’s most important property is often just whether the substrate survives. A tube network can be cut in half and still work. A population’s accumulated route knowledge cannot, and neither can a song tradition once the birds carrying it are gone.
The boundary of who remembers, in other words, has never been drawn by nature. It has been drawn by which organisms somebody thought to test, using criteria borrowed from the one lineage that happens to do it fastest.
A cell the size of a dinner plate remembers where the food was by being a different shape than it used to be. That is not a lesser version of what a brain does. It is the same problem, solved in the only material available, several hundred million years before anything had a head to keep it in.