Memory Engineering

Memory engineering is the deliberate alteration of what an organism remembers: writing a memory in, taking one out, changing its strength or its emotional sign, or supplying a prosthesis that does the encoding a damaged hippocampus cannot. The subject has a clean two-track structure, and merging the tracks is the commonest error made about it. Track one is rodent optogenetics. Since 2012 a small group of laboratories has labelled the neurons active during a specific experience, reactivated them with light, and produced the behaviour associated with that experience in an animal that is not in the original situation. Those demonstrations are spectacular and, within their terms, unambiguous. They are also almost entirely about contextual fear conditioning in the dentate gyrus, because freezing is the easiest behavioural read-out in neuroscience. Track two is human electrical stimulation. Here the interventions are depth electrodes placed in people undergoing intracranial monitoring for drug-resistant epilepsy, the outcome measures are percentage changes in the probability of recalling a word or an image, the effects are small, and two competent groups have obtained opposite signs from the same anatomical target.

Nothing in the human literature approaches the rodent result. There is no human experiment in which an ensemble is activated and a subject reports remembering a particular thing. The gap between the tracks is not a matter of scale or funding; it is a single missing capability, which this brief calls addressing — the ability to say which cells or which pattern carry a given memory in a given person. Every rodent triumph depends on labelling a population by its activity during a specific experience, using a viral, immediate-early-gene-driven pipeline with no human analogue. Every human intervention stimulates an anatomical target chosen by geometry. The brief is organised around that gap, around a second gap between what the field’s founding papers say and what they are cited for, and around the distinction that most of the popular coverage collapses: the difference between destroying a memory and blocking the route by which it is normally retrieved.

Established The founding engram demonstration is a sufficiency result and it is airtight within its own terms. Liu and colleagues in Susumu Tonegawa’s laboratory labelled dentate gyrus neurons active during contextual fear conditioning with channelrhodopsin-2 and reactivated them optically in a different context; the abstract states that optogenetic reactivation of hippocampal neurons activated during fear conditioning is “sufficient to induce freezing behaviour” (Nature 484:381, 2012). Established The controls are why the paper mattered: freezing “was not detected in non-fear-conditioned mice expressing ChR2 in a similar proportion of cells, nor in fear-conditioned mice with cells labelled by enhanced yellow fluorescent protein instead of ChR2,” and activating cells labelled in a context not associated with fear evoked no freezing in animals that had been conditioned elsewhere. Established What follows is that a sparse hippocampal ensemble is sufficient to drive the behavioural output of a contextual fear memory. Frontier What does not follow is that the ensemble contains the memory’s content, that the animal underwent anything resembling recollection, or that the finding generalises to non-aversive, non-contextual, content-rich memory.

Established The three follow-up results define the ceiling of memory writing in any organism, and it is low. Ramirez and colleagues labelled dentate gyrus or CA1 cells active in one context, reactivated them during foot shock in a second, and produced increased freezing in the first context, where no shock was ever delivered (Science 341:387, 2013). Established Redondo and colleagues drove fear and reward traces into opposite-valence conditioning: the dentate engram displayed the plasticity that allowed fear to become rewarding and the reverse, and the basolateral amygdala engram did not (Nature 513:426, 2014). Established Nabavi and colleagues conditioned an animal to associate a foot shock with optogenetic stimulation of auditory inputs to the amygdala, then reported that “subsequent optogenetic delivery of LTD conditioning to the auditory input inactivates memory of the shock. Then subsequent optogenetic delivery of LTP conditioning to the auditory input reactivates memory of the shock” (Nature 511:348, 2014). Established Vetere and colleagues then replaced the conditioned stimulus with optogenetic stimulation of a single olfactory glomerulus and the unconditioned stimulus with stimulation of ventral tegmental inputs, and wrote that “we created a fully artificial memory in mice” (Nature Neuroscience 22:933, 2019). Frontier Take the four together and the honest summary is that the field can install a one-bit association between a sensory channel and a valence, and can reverse the sign of an existing one. It cannot write an episode, and nobody has proposed how.

Established The most consequential result in the topic is the one that says the memory was still there. Ryan, Roy, Pignatelli, Arons and Tonegawa reported increased synaptic strength and dendritic spine density specifically in consolidated engram cells; under protein-synthesis-inhibitor-induced amnesia those properties are lacking, and yet direct optogenetic activation of the engram cells still retrieves the memory, correlating with retained engram-cell-specific connectivity (Science 348:1007, 2015). Frontier Their proposal is that a specific pattern of connectivity is the storage substrate and that strengthened synapses contribute to retrieval rather than to storage. Established Roy and colleagues made the same distinction in a disease model, writing that “direct optogenetic activation of hippocampal memory engram cells results in memory retrieval despite the fact that these mice are amnesic in long-term memory tests when natural recall cues are used, revealing a retrieval, rather than a storage impairment” in early Alzheimer’s models (Nature 531:508, 2016). Established In the same paper, optogenetic induction of long-term potentiation at perforant-path synapses restored both spine density and long-term memory, and ablating the restored engram cells abolished the rescue. Frontier The generalisation these two papers license is the spine of this brief: an animal that fails a memory test has not necessarily lost the memory. Frontier Note the tense in the field’s own synthesis, which concludes that its findings “are beginning to define an engram as the basic unit of memory” more than a century after Semon coined the term (Science 367, 2020).

Established The human route to modifying memory runs through reconsolidation, and its founding result is a behavioural amnesia, not a demonstrated deletion. Nader, Schafe and LeDoux reported that “consolidated fear memories, when reactivated during retrieval, return to a labile state in which infusion of anisomycin shortly after memory reactivation produces amnesia on later tests,” that the same treatment without reactivation left memory intact, and that delaying the infusion six hours produced no amnesia (Nature 406:722, 2000). Established The human translation used propranolol rather than a protein-synthesis inhibitor: Kindt, Soeter and Vervliet wrote that oral beta-adrenergic blockade before memory reactivation “erased the behavioral expression of the fear memory 24 h later and prevented the return of fear” (Nature Neuroscience 12:256, 2009). Established Read the qualifier: the outcome measure was the startle response, and the title says “erasing human fear responses.” The paper is careful and its citations are not. Established The clinical trial is real and modest. Brunet and colleagues randomised 60 adults with long-standing post-traumatic stress disorder to propranolol or placebo 90 minutes before brief weekly trauma-reactivation sessions for six weeks; the estimated post-treatment CAPS group difference was 11.50, and symptom checklist scores fell 2.43 points per week on propranolol against 0.55 on placebo (American Journal of Psychiatry 175:427, 2018). Established The number the enthusiast summaries omit is the placebo arm’s within-group effect size of 1.25 against propranolol’s 1.76: six weekly sessions of guided trauma reactivation is itself an active treatment, and the between-group increment is smaller than the headline. Frontier The senior author of the 2009 paper later co-wrote the field’s critical review, which finds “a wealth of observations consistent with reconsolidation” but concludes that alternative explanations mean the mechanism cannot be inferred from current evidence, and that reconsolidation “remains a viable but hotly contested explanation for some observed changes in memory expression” (Psychological Bulletin 144:797, 2018).

Frontier The memory prosthesis is the topic’s headline and the number attached to it should not be quoted as a result. The lineage begins with Berger, Hampson, Song and Deadwyler, who fitted a nonlinear multi-input/multi-output model to the CA3-to-CA1 relation in rats and showed that model-derived stimulation improved performance and recovered it under pharmacological blockade of synaptic transmission (Journal of Neural Engineering 8:046017, 2011). Frontier The 2018 human paper from the same collaboration is where the widely circulated figures come from: a reported 37 per cent facilitation of short-term and working memory from model-derived CA1 stimulation, and a 35 per cent improvement in retention on a delayed-recognition task (Journal of Neural Engineering 15:036014, 2018). Frontier This brief does not print those as an established outcome, for a specific and checkable reason: the abstract states no participant count, the full text sits behind a blocked publisher, and the research pack behind this brief could not establish how many people were studied. Established What can be reported is what the same collaboration found when it extended the work. Roeder and colleagues, in 14 adult subjects, using a memory-decoding model to target specific stimulus categories, report that “across all subjects, the stimulated trials exhibited significant changes in performance in 22.4% of patient and category combinations,” that those changes were a mixture of increases and decreases “with increases in performance occurring at almost 2 to 1 relative to decreases,” and that in patients with impaired memory receiving bilateral stimulation the figure rose above 37.9 per cent of combinations at better than 4 to 1 (Frontiers in Computational Neuroscience 18:1263311, 2024). Established Roughly one significant stimulation effect in three made memory worse. Established Gains were concentrated almost entirely in already-impaired patients receiving bilateral stimulation; the session was 150 delayed-match-to-sample trials with recognition tested 15 to 20 minutes later, on at most 16 independent CA1 channels. Frontier No group outside this collaboration has replicated any of it in the fourteen years since the rat paper.

Established Human memory stimulation produces real, small effects whose sign is not reliable, and the field should be reported as a declared tie rather than a trend. Suthana and colleagues stimulated below afterdischarge threshold in seven patients learning destinations in virtual environments: entorhinal stimulation during learning enhanced later memory for landmark locations — shorter routes, faster arrival — direct hippocampal stimulation did not, and entorhinal stimulation reset the phase of the hippocampal theta rhythm (New England Journal of Medicine 366:502, 2012). Established A larger multi-site study by a different group is titled “Direct Electrical Stimulation of the Human Entorhinal Region and Hippocampus Impairs Memory” (Neuron 92:983, 2016); the title is the finding, and the numbers behind it could not be obtained for this brief. Established The best-characterised positive result is closed-loop: Ezzyat and colleagues decoded encoding state from direct recordings in 25 epilepsy patients and triggered stimulation of lateral temporal cortex during poor encoding, reporting that “lateral temporal cortex stimulation increased the relative probability of item recall by 15% (t(3828) = 2.31, P = 0.02),” with an odds ratio of 1.18 at P = 0.04 (Nature Communications 9, 2018). Established Five years later, a non-invasive protocol using two interfering kilohertz fields to place an envelope at the hippocampus increased the odds of correct recall by 12 per cent, 95 per cent confidence interval 0.02 to 0.21, in 21 healthy participants with no surgery (Nature Neuroscience 26:1994, 2023). Frontier Two very different technologies, one requiring intracranial electrodes and one requiring scalp electrodes, produce memory effects of indistinguishable magnitude. Either the invasive route is being wasted, or both are modulating the same small, non-specific encoding-state variable.

Established The best-evidenced memory intervention in this entire literature involves no device, no drug and no surgery. A multilevel meta-analysis of targeted memory reactivation — replaying a learned cue quietly during sleep — covering 91 experiments, 212 effect sizes and 2,004 participants returns an overall Hedges’ g of 0.29, 95 per cent confidence interval 0.21 to 0.38, with g = 0.32 in stage NREM 2 and 0.27 in slow-wave sleep, and no effect during REM sleep or wakefulness (Psychological Bulletin 146:218, 2020). Established Put that beside a 15 per cent relative recall change at P = 0.02 in 25 people with implanted electrodes and the ordering is uncomfortable but plain: the cheap, non-invasive, well-replicated intervention has by a wide margin the better evidence base. Established And there is a structural feature of the human record that almost never gets stated. Every human intracranial memory result cited above — Suthana, Jacobs, Ezzyat, Hampson, Roeder — was obtained in patients undergoing monitoring for drug-resistant epilepsy, on antiepileptic medication, with electrodes placed for clinical rather than scientific reasons, in hippocampi that are by definition abnormal. Frontier It is the only way the work can be done, and it is nonetheless a permanent confound on what every number in the field means — one that better electrodes will not remove.

Frontier The prosthesis got more specific and smaller at the same time, which is what real progress usually looks like. The 2024 follow-up is the first demonstration in this line of stimulation patterns derived from a memory-decoding model aimed at particular stimulus categories rather than at a general encoding boost. It is a genuine advance in specificity. It is also 14 subjects, one collaboration, effects in 22.4 per cent of patient-and-category combinations, and a third of the significant changes going the wrong way. Frontier The single most important open item is not scientific but sociological: fourteen years without an independent replication of the field’s flagship result. A second centre running a pre-registered version of the same protocol is the cheapest, most overdue and most informative experiment available in this topic, and nobody has run it.

Frontier There is now a mechanistic hypothesis that could reconcile the contradictory human results, and it is a preprint. Characterisation of hippocampo-cortical stimulation responses in seven epilepsy patients reports that single neurons respond to local electrical stimulation with short-lived excitation followed by sustained inhibition, that neocortex-to-hippocampus input takes roughly 100 milliseconds, and that hippocampal output is gated by theta phase (bioRxiv 2024.11.28.625915). Speculative If the net effect of a stimulation pulse is excitation followed by sustained inhibition, then “stimulating to enhance memory” may often be net inhibitory, which would explain both the impairment finding of 2016 and the decreases in 2024. That is the most promising available reconciliation of the field’s central contradiction and it is a hypothesis rather than a result. Frontier A related human result bears on what a prosthesis would have to respect: co-occurring ripple oscillations recorded on four 96-channel arrays in three patients coordinate neuronal firing between cortical sites up to 16 millimetres apart (PNAS 121:e2312204121, 2024), which is the best current data on the binding mechanism any memory-writing device would be interfering with.

Frontier Non-invasive deep stimulation of the human hippocampus is now real, and it changes the cost structure of the whole field if it holds. Temporal-interference stimulation delivers two kilohertz-range fields whose difference frequency creates an envelope at depth; the field focusing was validated by modelling and by measurement in a human cadaver, then tested behaviourally in 21 people. Frontier One group, one protocol, a small effect, and a field-steering claim validated in a cadaver rather than a living brain. Speculative If it replicates, the argument for craniotomy in a healthy or mildly impaired person collapses, because the invasive route currently buys no additional effect size.

Speculative Direct memory transfer has exactly one modern experimental result and it is in a sea slug. RNA extracted from the central nervous systems of Aplysia given long-term sensitization training induced sensitization when injected into untrained animals; the induced sensitization required DNA methylation, and RNA from trained animals increased excitability in dissociated sensory neurons but not in motor neurons, leading the authors to conclude that RNA is sufficient to generate an engram for long-term sensitization (eNeuro 5(3), 2018). Speculative One laboratory, one species, and decisively: long-term sensitization is a non-associative memory. It is a change in reflex gain, not a memory of anything. Handwave The result deliberately revives a tradition — planarian transfer experiments in the 1960s, the scotophobin programme in the 1970s — that the field abandoned after replication failures, and the objection that killed that tradition has never been answered: two nervous systems that never shared an experience have no reason to share an addressing scheme, so there is nothing for transferred material to land on. It is honestly reported, unreplicated, and about the simplest form of memory there is.

Frontier Engram manipulation has become an assay rather than a therapy, and that is its most useful current role. The Alzheimer’s-model work is the template: use optogenetic access to the engram to distinguish a storage deficit from a retrieval deficit, a distinction that no behavioural test can make on its own. Frontier Influential, mouse-only, and with no human delivery route for its therapeutic implication, which is to restore spine density in a specific cell population. Frontier The reconsolidation clinic occupies a similarly odd position: one positive randomised trial, and the field’s own critical review saying the mechanism cannot be inferred. That is a treatment that may work through a mechanism that may not exist, which is a legitimate scientific state to be in and should be described that way rather than resolved by preference. Frontier Finally, the DARPA-era human memory programme has effectively stopped and its peer-reviewed outputs are public, but the programme documents themselves — budgets, dates, performers, stated goals — could not be retrieved for this brief, and no figure from that source appears here.

Established The workback target is a prosthesis that restores one specific, identified memory in a human, and the chain to it has nine links. Frontier Link 1, addressing. A human-compatible way to identify the cells or the patterns carrying a specific memory. Partially approached by the memory-decoding-model line, never validated against a held-out item set. Frontier Link 2, specificity. Show that stimulating a pattern derived from one item improves that item and not others. Never run. Frontier Link 3, replication. A second group, a different centre, a pre-registered protocol. None in fourteen years. Frontier Link 4, the sign problem. A parameter rule that predicts whether entorhinal stimulation will enhance or impair. The contradiction has stood since 2016. Speculative Link 5, a real-time read-out of whether stimulation is helping this patient today, which means extending a closed-loop classifier from encoding state to the sign of the stimulation effect. Not attempted. Frontier Link 6, population. Any of the above outside epilepsy monitoring. Speculative Link 7, durability. Does a facilitated memory survive 90 days? Frontier Link 8, a chronic implant that could deliver it. Speculative Link 9, a regulatory pathway, for which there is no predicate device and no indication.

Frontier Link 1 binds. Everything downstream is engineering around a scientific hole: for a given human memory, nobody can presently say which pattern carries it. Better electrodes, more channels and smarter classifiers all improve the delivery of a signal whose correct address is unknown. Established Link 6 is the structural one and it is rarely stated aloud. The entire human evidence base comes from people whose hippocampal circuits are epileptogenic enough to warrant intracranial monitoring. There is no ethical route to implanting electrodes in a healthy brain for this purpose, so the confound is not a funding problem or a recruitment problem; it is a permanent feature of the field unless a non-invasive method carries the whole programme. Frontier Link 3 is the cheapest and the most overdue, and its absence is the single most important caveat on every number in this topic. Established Link 7 is barely on anybody’s list and should be near the top: the longest post-stimulation interval anywhere in the human memory-engineering literature reached for this brief is 15 to 20 minutes. A memory prosthesis that improves recognition twenty minutes later has not yet been shown to do anything a patient would notice. Frontier Note what is not on this list: channel count, computational power and model sophistication. All three are improving and none of them is the constraint. A field that keeps building better instruments for an unsolved addressing problem will keep producing small, sign-unstable effects and calling them progress.

Established This brief depends first on the device layer, which it does not itself study. Any memory prosthesis is a chronic intracranial implant, so the yield, longevity and failure-mode literature carried by Neural Interfaces and the system-level results in Human-Machine Symbiosis are load-bearing here. Frontier The relevant new datum is the first published long-run record of intracortical arrays in humans: 20 arrays across the first 14 BrainGate participants, 2,319 recording sessions, implants up to 7.6 years, with 11 of 14 arrays providing meaningful decoding throughout enrollment. Frontier The same work reports that decoding signal-to-noise ratio increases logarithmically with electrode count, which is simultaneously a scaling pathway and the mathematical form of diminishing returns.

Established It depends second on human single-neuron methods. Over 200 well-isolated cortical single units have been recorded simultaneously in humans intraoperatively with high-density probes (Nature Neuroscience 25:252, 2022), which is the nearest existing approach to identifying engram membership from extracellular recording — acutely, in cortex, for minutes to hours. Speculative It depends third on a capability nobody has proposed: a human-compatible, activity-dependent labelling method. Optogenetic engram tagging requires immediate-early-gene-driven transgene expression delivered by viral vectors and activated through an implanted optical fibre. There is no plausible human version of that pipeline, and this is the hardest technical barrier in the subject. Frontier It depends fourth on the physical limits of neural recording, which bound how much of an engram could ever be read at all, and which are carried in the symbiosis brief rather than re-derived here. Frontier It depends fifth on a result nobody is producing: a characterisation of what a stimulation pulse actually does to a hippocampal circuit over the following few hundred milliseconds, without which the field cannot predict the sign of its own intervention. Speculative And it depends, less obviously, on the sleep sciences. If the highest-value clinical target turns out to be retrieval rather than encoding, then the consolidation and cueing literature owns the mechanism and the implant community owns only the delivery hardware, which would reverse the current allocation of attention and money in this field.

Frontier The decisive experiment for the prosthesis has never been run and it is not expensive. Derive a stimulation pattern from the model of item A, deliver it while the subject encodes item B, and test both. Pass condition: a double dissociation in which recall of A improves and recall of B does not. Frontier The 2024 category-specific design is the closest approach in the literature and it does not do this, because category-level targeting cannot distinguish content addressing from a category-tuned encoding boost. Established Until that dissociation exists, every human memory-enhancement result remains consistent with the modulation of a general encoding-state variable rather than of a memory.

Frontier Second: the replication. An independent group, a different centre, the same multi-input/multi-output and memory-decoding protocol, pre-registered, reporting an effect size with a confidence interval rather than a percentage. Frontier Third: resolve the sign. A within-subject comparison of entorhinal stimulation under the 2012 and 2016 parameter sets against a common endpoint, with the pass condition being a rule that predicts enhancement versus impairment from stimulation parameters and phase. Speculative The excitation-then-sustained-inhibition finding gives that experiment a specific hypothesis to test, namely that the sign depends on the timing of the pulse relative to theta phase and on the duration of the post-stimulation inhibitory window. Frontier It is worth noting how cheap this experiment is relative to the rest of the list: it needs no new hardware, no new model and no new population, only two parameter sets and a common outcome measure in one monitoring unit.

Speculative Fourth: a sign-predicting closed loop. Extend the encoding-state classifier so that it predicts prospectively whether the next stimulation will help or hurt this patient in this session, then withhold stimulation when the prediction is negative. This is the technical version of what is usually raised as an ethical objection, and it is the more actionable version. Speculative Fifth: durability. Recall advantage at 90 days for a facilitated item. Handwave Sixth, and the one that would settle the erasure question: a human analogue of the engram-retention experiment — a retrieval route that bypasses the blocked one, applied after a reconsolidation-blocking drug, to test whether the memory is gone or merely unreachable. No such route exists in humans and it may not be constructible, which is precisely why the erasure claim is unfalsifiable in humans at present and should be reported as such.

Established The prosthesis hardware is far behind the ambition, and the binding number is channel count in the right place. The 2024 human study had at most 16 independent CA1 channels available for stimulation, drawn from 8 to 10 macro-style depth probes plus one to four macro-micro probes per subject. Frontier A device that addressed individual memories would need to write patterns across a hippocampal ensemble, and 16 channels is not a coarse version of that capability; it is a different kind of thing. Established The recording side is better provisioned than the writing side: the long-run human array data show mean electrode yield around a third with only a few percentage points of decline across years of enrollment, and human implants outlasting non-human-primate ones.

Frontier A memory prosthesis is a stimulating device, and stimulation is harder on tissue and on hardware than recording. Explant analysis of human arrays shows insulation degradation and encapsulation as the dominant chronic failure modes rather than the glial scar of the textbook account, and the same failure modes govern any device that would stimulate hippocampus for years. Speculative The engineering requirement nobody has specified is the closed-loop budget: a device that must classify encoding state, predict the sign of its own effect and decide within a theta cycle whether to fire has a latency budget of tens of milliseconds and a power budget set by the thermal limits of implanted electronics. Frontier No published memory device operates under both constraints. Frontier The closest existing engineering template is not a memory device at all but individualised closed-loop neuromodulation for depression, in which the biomarker is discovered in the individual patient before the device is configured — the only methodology in this cluster that could plausibly be adapted to a memory target, because it does not assume a shared code across people.

Established The adjacent technology with the strongest evidence is a speaker and a sleep schedule. Targeted memory reactivation needs no implant, has a meta-analytic effect size across 2,004 participants, and is already the best-evidenced version of memory engineering in existence; it is almost never described as such because it does not look like engineering. Frontier Non-invasive deep stimulation by temporal interference is the adjacent hardware that matters most, because it is the only route by which a hippocampal intervention could ever reach a population that does not already have electrodes in its head.

Frontier Decoding methods developed for speech and movement interfaces transfer directly to the addressing problem, and the durability figures from that literature are the only long-run human decoding-stability data anywhere: a speech neuroprosthesis sustaining 97.5 per cent accuracy over 8.4 months in one participant sets the bar for what a fitted per-patient model can be expected to hold. Frontier The demonstration that inner speech is decodable from motor cortex in four participants, published together with a “mental password” gating mitigation, is the nearest thing the field has to a memory-reading result and should be read as a warning about what becomes possible incidentally. Established Cognitive Enhancement carries the pharmacological and transcranial-stimulation comparators; Neuroplasticity Engineering carries the plasticity-window work that any durable memory intervention would exploit; Cognitive Liberty carries the rights framing. Speculative Mind Uploading and Brain Preservation depend on this subject far more than it depends on them: both assume that an engram is readable at whatever resolution carries its content, which is exactly the assumption this brief cannot support, and the engram-retention result sharpens the problem rather than easing it — if storage lives in a connectivity pattern rather than in synaptic weights, then a preservation method optimised to fix synaptic strengths may be optimising the wrong quantity. Frontier Closer to the clinic, adaptive deep-brain stimulation is the adjacent product line that proves the closed-loop control architecture a memory device would need, in a different indication and with a different biomarker; the memory field inherits its hardware and none of its endpoints. Speculative And the drug side is adjacent in an underappreciated way: propranolol is a generic, oral, cheap beta-blocker, so if the reconsolidation clinic is real it is the only branch of memory engineering with a plausible route to population scale.

Established There is no regulatory object called a memory prosthesis. It has no predicate device and no indication; the nearest approved neighbour is responsive neurostimulation for epilepsy, which is a seizure device that happens to sit in the same tissue. Speculative A first approval would have to define an endpoint, and the field has not agreed one: percentage change in recognition twenty minutes later is a laboratory measure, not a clinical benefit, and no consensus outcome instrument for episodic memory prosthetics exists. Frontier That absence is not neutral. In the absence of an agreed endpoint, a sponsor picks the measure on which its device performed, which is precisely how a 37 per cent figure with no stated denominator enters circulation and stays there.

Established The research venue is an institutional bottleneck disguised as a scientific one. Human intracranial memory work happens in epilepsy monitoring units, on patients whose electrode placement is clinically determined, during a window of days. That constrains sample sizes to the single digits and low tens, prevents randomisation to electrode location, and makes every result a secondary use of a clinical procedure. Frontier An institution that wanted to change the field’s trajectory would fund a multi-centre consortium with a common task, a shared test set and a pre-registered analysis, exactly as the movement-decoding community is now beginning to do with a cross-participant signal-to-noise metric. Frontier Nothing comparable exists for memory.

Frontier The replication gap is a funding-design problem with a known fix. Fourteen years of a single collaboration producing the field’s headline result is not a scandal; it is what happens when no funder has a line item for replicating an expensive intracranial protocol at a second centre. Speculative A modest, ring-fenced replication fund attached to any grant that produces a flagship human neuromodulation result would cost a fraction of the original programmes. Established It is also worth recording what could not be established for this brief: the programme documents of the government-funded human memory work of the 2010s — budgets, dates, performers and stated goals — were unreachable, so no figure from that source appears anywhere in this brief. A field whose institutional history is harder to obtain than its papers is one where narrative drift is cheap.

Established Start with the finding rather than the principle: the same device, the same model and the same patient can make memory worse. About a third of the significant stimulation effects in the 2024 study were decreases in memory performance, and the largest earlier multi-site study reports impairment as its headline. Frontier A memory prosthesis is necessarily also a memory disruptor, and there is no published protocol for detecting which one a given patient is receiving on a given day. Speculative That reframes the central ethical question as a technical one — build the sign-predicting read-out — and the technical framing is preferable because it is actionable, whereas the abstract version has been discussed for a decade without producing an instrument.

Established Consent in this field is structurally unusual. Every human participant is an epilepsy patient in a monitoring unit, awaiting surgery, approached for research during a clinically stressful admission, with electrodes placed for reasons unrelated to the study. The work is done well and it is the only way it can be done, but the population cannot consent to the placement, only to the additional task. Frontier The reconsolidation clinic raises a different question that the evidence has already narrowed: propranolol reduces the behavioural expression of a fear memory and reduces post-traumatic symptoms, but patients still remember the trauma, so the widely aired worry about erasing morally significant memories is not the situation the evidence describes. Speculative The live version is subtler — a treatment that reliably attenuates the emotional force of an accurate memory used in testimony, in litigation, or in a professional context where the affect is load-bearing. Handwave If content editing ever became possible in humans, the question would move from attenuation to fabrication, and nothing in the current evidence base suggests that transition is near. Frontier The mental-privacy result that matters most here was produced incidentally by a speech interface, which is the pattern to expect: memory-reading capability is likelier to arrive as a by-product of decoding research than as its stated goal.

Speculative If the addressing problem were solved, the first consequence would not be memory editing but memory verification, and that is the larger civilizational variable. A method precise enough to identify which pattern carries a given episode would, before it could write anything, make it possible to ask whether a memory was encoded at all. Testimony, clinical history and personal identity all currently rest on the fact that nobody can check. Handwave Nothing in the present evidence supports that capability, and the physical bounds on how much of an engram can be read at all are unfavourable.

Frontier The more likely civilizational path runs through retrieval rather than storage. If a large fraction of clinical memory failure is retrieval failure with intact storage — which is what the engram-retention and Alzheimer’s-model results suggest — then the engineering target changes from writing memories in better to providing better retrieval cues, and the best existing instrument for that is a sound played during slow-wave sleep. Speculative That is a cheap, distributable, non-invasive intervention with a real meta-analytic effect size, and a civilization that took memory engineering seriously would industrialise it long before it implanted anybody. Speculative The same logic applies to forgetting. If amnesia after protein-synthesis blockade is a retrieval block rather than a deletion, then a good deal of ordinary forgetting may also be loss of access rather than loss of content — the active-forgetting position — and the practical upshot is that lost memories may be recoverable in principle and that the loss is not the failure it appears to be. Frontier This brief flags that as a reading of the retrieval-versus-storage results rather than as an established finding; the dedicated active-forgetting literature was not available to it. Speculative External and technological memory is now treated in serious taxonomies as part of human memory rather than as a metaphor, and on current evidence the most consequential memory engineering of the next century will be done to the environment rather than to the hippocampus.

These horizons track the two tracks separately, because they are not converging on any schedule that can be read off the current literature.

• 10 yr: Frontier An independent replication of the hippocampal memory prosthesis is either published or its absence becomes the field’s defining embarrassment; this brief judges publication more likely than not, and judges the replicated effect likelier to resemble the 2024 numbers than the 2018 ones. Frontier Temporal-interference stimulation is either replicated by a second group or quietly drops out, and its replication is the highest-leverage single event in the topic. Speculative The sign problem gets a parameter rule, most plausibly a theta-phase one, from the closed-loop community rather than from the prosthesis community. Frontier Targeted memory reactivation reaches consumer sleep hardware long before any implant reaches a clinic, at effect sizes smaller than the laboratory meta-analysis.
• 25 yr: Speculative A memory-assist device is approved for a narrow indication in an already-impaired population, on an endpoint negotiated rather than discovered, most plausibly as an adjunct to an implanted device placed for another reason. Speculative The addressing problem is unsolved but partially routed around by per-patient fitted models that make no claim to read content. Frontier The reconsolidation clinic either becomes standard practice for post-traumatic stress or is absorbed into exposure therapy once the drug’s contribution is separated from the reactivation sessions; the second is at least as likely. Speculative Human activity-dependent labelling still has no route.
• 50 yr: Speculative Either a human-compatible addressing method exists — most plausibly a decoding method rather than a labelling one — or the topic has settled permanently into retrieval assistance, cueing and prosthetic external memory, which is the outcome this brief considers more probable. Handwave Content-level editing of a human episodic memory remains undemonstrated. Speculative If the retrieval framing is right, the dominant clinical use by then is restoring access to memories that were never lost — a different and more tractable industry than the one currently promised.
• 100 / 250+ yr: Handwave Memory backup, transfer or exogenous storage requires three things that do not exist: an engram readable at whatever resolution carries its content, a substrate that supports retrieval from what was read, and a solution to the addressing problem in both directions. Handwave The gap is quantifiable and that is the useful thing about it: the largest memory ever written into a nervous system from outside is a single association between one olfactory channel and one valence, installed by optogenetics, in a mouse. Handwave Direct transfer between two nervous systems that never shared an experience remains blocked by the objection that ended the scotophobin programme, which has never been answered rather than merely being unfashionable.

• Depends on This brief depends on the device layer it does not itself study. A memory prosthesis is a chronic intracranial stimulating implant, so everything in the human-machine symbiosis record about what implanted arrays do over years — yield, longevity, failure modes, the logarithmic relation between electrode count and decoding signal-to-noise ratio, and the single-participant character of most headline results — is inherited here rather than reproduced. The memory literature has no chronic device of its own and no participant who was implanted in order to have their memory studied. It depends second on human single-neuron and population recording, because the only plausible route to identifying engram membership in a person runs through decoding rather than through the viral, immediate-early-gene labelling pipeline that every rodent result uses and that has no human analogue. It depends third on the physical bounds on neural recording, which set a ceiling on how much of any engram could ever be read. And it depends fourth on something no discipline currently owns: an agreed clinical endpoint for episodic memory, without which no result here can become an approvable claim.
• Enables What this brief enables is narrower than its reputation and clearer for being stated exactly. It supplies the storage-versus-retrieval distinction, the most useful conceptual export in the cluster: an animal or a person who fails a memory test has not necessarily lost the memory, and that reframing changes the target of Alzheimer’s therapeutics, of amnesia research and of the reconsolidation clinic. It supplies the strongest available discipline on claims about memory reading, which mind uploading and brain preservation both depend on and neither can supply for itself: if the engram is a connectivity pattern rather than a set of synaptic weights, the readout requirement changes character. It supplies the empirical case for cueing over encoding, directly usable in education, rehabilitation and sleep technology. And it supplies, for governance, the finding that a memory device is necessarily bidirectional in effect — a sharper starting point for regulation than any principle about mental integrity.
• Adjacent Adjacent to this brief sit three literatures that share its instruments and none of its conclusions. Cognitive enhancement uses the same stimulation hardware against different endpoints and arrives at the same uncomfortable place: small effects, contested replication, and a non-invasive comparator that performs as well as the invasive one. Neuroplasticity engineering owns the plasticity windows that any durable memory intervention would have to open, and its results bound how long an engineered change could be expected to last. Cognitive liberty and the neurorights literature own the rights framing, and this brief deliberately hands them a technical fact rather than a principle: about a third of significant stimulation effects in the best-documented human memory-prosthesis study were decreases. Further out, the speech and movement interfaces are adjacent in the way that matters most, because the first genuine memory-reading capability is far likelier to arrive as a by-product of decoding research than as the announced goal of a memory programme.

Established “Blocking reconsolidation deletes the memory.” This is the topic’s headline correction, and it comes from inside the house. The entire reconsolidation-erasure literature rests on the 2000 finding that anisomycin infused into the amygdala after memory reactivation produces amnesia. In 2015, Ryan, Roy, Pignatelli, Arons and Tonegawa — the same laboratory lineage that produced the engram-manipulation results this field is famous for — showed that under exactly that protein-synthesis-inhibitor-induced amnesia, engram cells retain the memory and direct optogenetic activation retrieves it, with engram-cell-specific connectivity preserved despite lost spine density and synaptic strength (Science 348:1007). Frontier Their proposed reading is that connectivity pattern is the storage substrate and that synaptic strength serves retrieval. Established The amnesia is a retrieval block, not a deletion. Any sentence claiming that a drug erased a memory has to answer that paper, and it cannot be dismissed as an outsider’s objection: it is under-cited relative to its importance in exactly the places where erasure is claimed.

Established “Propranolol erases traumatic memories.” The flagship human paper wrote that propranolol before reactivation “erased the behavioral expression of the fear memory 24 h later and prevented the return of fear,” and titled itself “erasing human fear responses.” The outcome measure was startle. Established The qualifier is not decoration; the difference between an expression and a memory is the entire scientific question, and the citations routinely drop it. Established The randomised trial reduced post-traumatic symptoms with a between-group CAPS difference of 11.50 and a placebo within-group effect size of 1.25 against propranolol’s 1.76. Established The patients still remember the trauma.

Established “A hippocampal memory prosthesis improved human memory by 37 per cent.” The figure comes from the abstract of a 2018 paper that states no participant count, whose full text was not obtainable for this brief. The same collaboration’s 2024 follow-up in 14 subjects found significant changes in 22.4 per cent of patient-and-category combinations, of which roughly a third were decreases in memory performance, with gains concentrated almost entirely in patients who already had impaired memory and received bilateral stimulation. Frontier The 2024 paper itself describes the earlier work as a “34-37% improvement in memory overall.” Established No independent group has replicated any of it in fourteen years. Frontier This is the exact pattern to watch for: an effect quoted at its founding magnitude while its own authors’ later work reports something far more modest and considerably more interesting.

Established “Scientists implanted a false memory in a mouse.” What was demonstrated is that a hippocampal ensemble labelled in one context, optically reactivated during foot shock in a second context, causes the mouse to freeze in the first. Established That is an artificial association between an ensemble and an aversive outcome. Frontier No measurement in the experiment could detect whether the mouse remembers an event that did not happen, and none was attempted. The work is real, elegant and important; the phrase “false memory” — the authors’ own title — is doing a great deal of work.

Established “Optogenetics showed we can switch memories on and off.” It showed that optical activation of a labelled population is sufficient to drive a conditioned behavioural output, in the dentate gyrus, for contextual fear, in transgenic mice, through an implanted optical fibre. Established Almost every result in the engram literature is a fear memory, because freezing is easy to measure, and the generality to content-rich memory is assumed rather than shown. Established Note also the structure of the field’s strongest causal result: a memory inactivated by long-term depression and reactivated by long-term potentiation. Off and back on again means the depression did not destroy it.

Established “DARPA showed brain stimulation boosts memory 15 per cent.” The 15 per cent is a relative increase in the probability of item recall, at P = 0.02, with an odds ratio of 1.18, in 25 epilepsy patients, from closed-loop stimulation of lateral temporal cortex. Established A scalp-applied, non-invasive protocol produced a 12 per cent increase in the odds of correct recall in 21 healthy volunteers five years later, with no surgery at all. Frontier A reader who sees “15 per cent” without the qualifiers will imagine something an order of magnitude larger than what was measured.

Established “Deep-brain stimulation of the entorhinal cortex improves memory.” One study of seven patients found enhancement and a theta phase reset. A larger multi-site study by a different group is titled “Direct Electrical Stimulation of the Human Entorhinal Region and Hippocampus Impairs Memory.” Established Both are peer-reviewed, both used epilepsy patients, both targeted the same region, and the contradiction has been unresolved since 2016. Frontier This brief declares the tie rather than averaging it, and notes that the most promising reconciliation on offer — short-lived excitation followed by sustained inhibition — is a preprint.

Speculative “Memory has been transferred between animals by RNA.” One laboratory transferred long-term sensitization — a non-associative increase in reflex gain — in Aplysia, in 2018. Established That is not a memory of anything; it is a change in how strongly a reflex fires. Handwave The result is honestly reported, mechanistically specific and unreplicated, and it revives a tradition abandoned half a century ago whose central objection — that two nervous systems which never shared an experience share no addressing scheme — remains unanswered.

Frontier “None of the human memory stimulation work is real.” The sceptical overcorrection is also wrong. A within-list, interleaved, item-level design found a significant effect at 25 patients; entorhinal stimulation produced route-length improvements at seven; non-invasive temporal interference produced a 12 per cent odds increase at 21. Established The effects are small, the sign is not always positive and the population is always unusual, but they are measured effects with published statistics and confidence intervals, not artefacts.

Established “The best way to improve memory is a device.” The best-evidenced memory intervention in this brief is playing a learned cue during slow-wave sleep: 91 experiments, 212 effect sizes, 2,004 participants, Hedges’ g = 0.29, with no effect during wakefulness or REM. It costs nothing and requires no implant. Frontier That it is rarely counted as memory engineering says more about the field’s self-image than about the evidence.

Established “Human memory research uses normal brains.” Every human intracranial memory result in this brief was obtained in patients undergoing monitoring for drug-resistant epilepsy, on antiepileptic medication, with electrodes placed for clinical reasons. Frontier It is the only way the work can be done, and it is a structural limit on what every number in the field means. Speculative A related discipline this brief has imposed on itself: where a primary text could not be obtained — the participant count of the 2018 prosthesis paper, the numbers behind the 2016 impairment result, the programme records of the government-funded memory work of the 2010s — the absence is stated rather than filled in from memory. In a subject this saturated with confident secondary summaries, that is not a limitation of the brief; it is most of its value.