Memory prosthesis: revision 1 to 2

Revision 1

Revision 2

categories: ["cybernetics", "minds"]

tags: ["memory", "hippocampus", "implants", "cognitive enhancement", "closed-loop", "neuroethics"]

summary: "An implanted device intended to restore or improve memory formation by supplying the pattern of neural activity a damaged or underperforming hippocampus fails to generate."

summary: "An implanted device intended to restore or improve memory formation by supplying the pattern of neural activity a damaged or underperforming hippocampus fails to generate."

updated: "2026-07-28"

updated: "2026-10-29"

humanEvidence: "Within-session experiments in small numbers of epilepsy patients carrying temporary depth electrodes report better recall on laboratory tasks; no person has ever had a chronic memory implant."

humanEvidence: "Within-session experiments in small numbers of epilepsy patients carrying temporary depth electrodes report better recall on laboratory tasks; no person has ever had a chronic memory implant."

access: "Not obtainable on any terms: the work exists only inside research protocols that use electrodes implanted for epilepsy surgery, and no memory prosthesis is approved or sold anywhere."

access: "Not obtainable on any terms: the work exists only inside research protocols that use electrodes implanted for epilepsy surgery, and no memory prosthesis is approved or sold anywhere."

reversibility: "difficult"

}

```

**Memory prostheses** are devices intended to improve or restore the formation of new memories by

**Memory prostheses** are devices intended to improve or restore the formation of new memories. They

recording activity in the hippocampal circuit, computing what the healthy circuit would have

record activity in the hippocampal circuit, compute what the healthy circuit would have produced,

produced, and delivering that pattern back as electrical stimulation. The concept is the only

and deliver that pattern back as electrical stimulation. The concept is the only serious attempt at

serious attempt at a *cognitive* neuroprosthesis, as distinct from the sensory and motor devices

a *cognitive* neuroprosthesis, as distinct from the sensory and motor devices that dominate

that dominate [[neuroprosthetics]]. As of 2026 it exists as a set of within-session experiments in

[[neuroprosthetics]]. As of 2026 it exists as a set of within-session experiments in patients who

patients who already have electrodes implanted for epilepsy surgery, with reported effects on

already have electrodes implanted for epilepsy surgery, with reported effects on laboratory memory

laboratory memory tasks and no chronic device in anyone.

tasks and no chronic device in anyone.

## The idea

Sensory and motor prostheses substitute for a transmission path. A [[cochlear-implant]] replaces a

transducer; a motor [[brain-computer-interface]] carries a command around a broken spinal cord.

transducer, and a motor [[brain-computer-interface]] carries a command around a broken spinal cord.

Neither has to reproduce a computation.

Memory is different. The hippocampus does not relay signals; it transforms them, converting patterns

Memory is different. The hippocampus transforms the signals it receives instead of relaying them,

of cortical input into the sparse, distributed codes that support later recall. A prosthesis for

converting patterns of cortical input into the sparse, distributed codes that support later recall.

memory must therefore reproduce a transformation, not a channel. Theodore Berger's group posed the

A memory prosthesis must therefore reproduce a transformation rather than a channel. Theodore

problem in exactly those terms: model the input-output function of the CA3-to-CA1 projection well

Berger's group posed the problem in exactly those terms: model the input-output function of the

enough to predict, from observed CA3 firing, what CA1 should do, then stimulate CA1 accordingly when

CA3-to-CA1 projection well enough to predict, from observed CA3 firing, what CA1 should do, then

the natural pathway fails.[^berger2011]

stimulate CA1 accordingly when the natural pathway fails.[^berger2011]

The model used is a nonlinear multi-input multi-output system fitted to simultaneously recorded

The model is a nonlinear multi-input multi-output system fitted to simultaneously recorded spike

spike trains, with no attempt to represent the underlying biophysics. It treats the hippocampus as a

trains, with no attempt to represent the underlying biophysics. It treats the hippocampus as a black

black box with measurable dynamics, which is both the approach's practical strength and the reason

box with measurable dynamics. That is the approach's practical strength, and it is also the reason

its generality is uncertain.

## Development history

```timeline

[

  { "year": "2011", "title": "Rodent proof of concept", "text": "Berger and colleagues fit a nonlinear model of hippocampal CA3-to-CA1 transmission in rats performing a delayed-nonmatch-to-sample task. Model-driven stimulation restores performance after the pathway is pharmacologically blocked, and improves it in intact animals." },

  { "year": "2011", "title": "Rodent proof of concept", "text": "Berger and colleagues fit a nonlinear model of hippocampal CA3-to-CA1 transmission in rats performing a delayed-nonmatch-to-sample task. Model-driven stimulation restores performance after the pathway is pharmacologically blocked, and improves it in intact animals." },

  { "year": "2012–2013", "title": "Non-human primates", "text": "Related work in macaques shows that stimulation patterned on the firing observed during successful encoding improves later performance on a delayed match-to-sample task." },

  { "year": "2012-2013", "title": "Non-human primates", "text": "Related work in macaques shows that stimulation patterned on the firing observed during successful encoding improves later performance on a delayed match-to-sample task." },

  { "year": "2014", "title": "DARPA funds the field", "text": "The Restoring Active Memory programme funds two teams to develop closed-loop memory devices, framed around traumatic brain injury in service members." },

  { "year": "2014", "title": "DARPA funds the field", "text": "The Restoring Active Memory programme funds two teams to develop closed-loop memory devices, framed around traumatic brain injury in service members." },

  { "year": "2016–2017", "title": "Stimulation can also impair", "text": "Direct stimulation of the human hippocampus and entorhinal region during encoding is found to worsen recall in several studies, contradicting earlier reports and establishing that timing and location matter more than dose." },

  { "year": "2016-2017", "title": "Stimulation can also impair", "text": "Direct stimulation of the human hippocampus and entorhinal region during encoding is found to worsen recall in several studies, contradicting earlier reports and establishing that timing and location matter more than dose." },

  { "year": "2018", "title": "Human results published", "text": "Two groups report memory improvements in epilepsy patients: one using a hippocampal encoding model, the other using a decoded brain-state trigger for lateral temporal cortex stimulation." },

  { "year": "2018", "title": "Human results published", "text": "Two groups report memory improvements in epilepsy patients: one using a hippocampal encoding model, the other using a decoded brain-state trigger for lateral temporal cortex stimulation." },

  { "year": "2020s", "title": "No chronic device", "text": "Work continues in acute monitoring settings and in modelling; no implantable memory prosthesis has entered a clinical trial as a permanent device." }

  { "year": "2020s", "title": "No chronic device", "text": "Work continues in acute monitoring settings and in modelling; no implantable memory prosthesis has entered a clinical trial as a permanent device." }

]

Two experimental designs account for most of what is known.

The **encoding-model approach** records multi-unit activity from hippocampal depth electrodes — placed

The **encoding-model approach** records multi-unit activity from hippocampal depth electrodes while

for seizure localization, the same clinical opportunity that supports much of the work described

a patient performs a visual memory task. The electrodes are placed for seizure localization, the

under [[ecog-interfaces]] — while a patient performs a visual memory task, fits a patient-specific

same clinical opportunity that supports much of the work described under [[ecog-interfaces]]. The

model relating activity during successful encoding to activity during failures, and then delivers

team fits a patient-specific model relating activity during successful encoding to activity during

stimulation matching the successful pattern. In a study of a small number of epilepsy patients, this produced improvements of

failures, then delivers stimulation that matches the successful pattern. In a study of a small

roughly a third over each participant's own unstimulated baseline on short-term and delayed

number of epilepsy patients, this produced improvements of roughly a third over each participant's

recall.[^hampson2018] The effect was measured within the same testing sessions, in patients on

own unstimulated baseline on short-term and delayed recall.[^hampson2018] The effect was measured

anti-seizure medication with abnormal hippocampi, using tasks designed for the experiment.

within the same testing sessions, in patients on anti-seizure medication with abnormal hippocampi,

using tasks designed for the experiment.

The **state-triggered approach** does not model content at all. A classifier trained on the

patient's own neural activity detects moments of poor encoding — states in which a word about to be

patient's own neural activity detects moments of poor encoding, meaning states in which a word about

presented is unlikely to be recalled — and delivers stimulation to lateral temporal cortex only

to be presented is unlikely to be recalled, and stimulation is delivered to lateral temporal cortex

during those moments. Recall improved by roughly fifteen percent relative to unstimulated

only during those moments. Recall improved by roughly fifteen percent relative to unstimulated

trials.[^ezzyat2018] Crucially, the same stimulation delivered during good encoding states impaired

trials.[^ezzyat2018] The same stimulation delivered during good encoding states impaired

performance, which is why open-loop stimulation had produced contradictory results for years.

performance, which explains why open-loop stimulation had produced contradictory results for years.

That contradiction is itself an important result. Direct stimulation of the human hippocampus and

That contradiction is an important result in its own right. Several careful studies have shown that

entorhinal region during learning has been shown to *degrade* subsequent recall in several careful

direct stimulation of the human hippocampus and entorhinal region during learning *degrades*

studies.[^jacobs2016] An earlier report of enhancement from entorhinal stimulation has not

subsequent recall.[^jacobs2016] An earlier report of enhancement from entorhinal stimulation has not

replicated consistently. Any claim that stimulating the memory system improves memory has to specify

where, when, and in what state, and the size of the literature that failed to do so is a reason for

caution about the field's headline numbers.

## Why episodic memory resists this approach

Motor decoding succeeded because the problem is well posed: an intended movement is continuous,

Motor decoding succeeded because the problem is well posed. An intended movement is continuous,

observable, low-dimensional, and repeatable, so a decoder can be trained on thousands of labelled

observable, low-dimensional and repeatable, so a decoder can be trained on thousands of labelled

attempts. Episodic memory has none of those properties.

**There is no ground truth signal to train against.** A memory's content is not observable from

**There is no ground truth signal to train against.** A memory's content cannot be observed from

outside except through a later report, which arrives minutes or days after the activity that must be

modelled. Feedback is sparse, delayed, and binary.

modelled. Feedback is sparse, delayed and binary.

**The code is idiosyncratic and drifting.** Hippocampal representations differ between individuals,

between items, and between days in the same individual — a phenomenon documented as representational

between items, and between days in the same individual, a phenomenon documented as representational

drift. A model fitted on Monday may not describe Tuesday's circuit, and no anatomical map of the

drift. A model fitted on Monday may not describe Tuesday's circuit. No anatomical map of the kind

kind produced by [[connectomics]] specifies the code, because the same wiring supports different

produced by [[connectomics]] specifies the code, because the same wiring supports different

representations at different times.

**Encoding, consolidation, and retrieval are separate problems.** A device that improves encoding

**Encoding, consolidation and retrieval are separate problems.** A device that improves encoding

does nothing for a memory already formed but inaccessible; consolidation unfolds over hours to years

does nothing for a memory already formed but inaccessible. Consolidation unfolds over hours to years

and involves systems-level transfer to neocortex during sleep, which no implanted device addresses.

**Content and process are entangled.** A motor prosthesis needs to know *how fast and in what

direction*. A memory prosthesis that restores a specific experience would need to supply the

direction*. A memory prosthesis that restored a specific experience would need to supply the pattern

specific pattern corresponding to that experience, which requires knowing what the person

corresponding to that experience, which requires knowing what the person experienced. The existing

experienced. The existing devices sidestep this entirely: they improve the *probability* that

devices sidestep this entirely: they raise the *probability* that encoding succeeds without touching

encoding succeeds without touching *what* is encoded. That is a real and useful effect, and it is

*what* is encoded. That is a real and useful effect, and it is much less than the term "memory

much less than the term "memory prosthesis" implies.

prosthesis" implies.

These are the same limits that constrain [[neural-decoding]] as it moves from movement to internal

The same limits constrain [[neural-decoding]] as it moves from movement to internal states. They

states, and they bear on the assumption in [[whole-brain-emulation]] and

also bear on the assumption, in [[whole-brain-emulation]] and [[mind-uploading]], that memory

[[mind-uploading]] that memory content is straightforwardly readable from neural structure. The

content can be read straightforwardly from neural structure. The manipulations that come closest to

manipulations that come closest to writing specific content are the optogenetic engram experiments

writing specific content are the optogenetic engram experiments described under [[optogenetics]],

described under [[optogenetics]], which require genetic access to individual cells and have been

which require genetic access to individual cells and have been performed only in mice.

performed only in mice.

## Restoration, enhancement, and ethics

The clinical framing is restoration: traumatic brain injury, stroke, epilepsy-related memory

impairment, and eventually Alzheimer's disease. The last is a poor fit for the current approach.

impairment, and eventually Alzheimer's disease. The last fits the current approach poorly.

Stimulation of the fornix in Alzheimer's patients, tested in a randomised trial, did not improve

outcomes overall, with a possible signal in older participants that a follow-up trial is examining.

outcomes overall; there was a possible signal in older participants, which a follow-up trial is

Degenerative disease removes the neurons a prosthesis would need to record from and stimulate, which

examining. Degenerative disease removes the neurons a prosthesis would need to record from and

is a different problem from a damaged pathway between intact structures.

stimulate, a different problem from a damaged pathway between intact structures.

Enhancement in healthy people is the more discussed and less tractable prospect. The rodent work

Enhancement in healthy people is discussed more and is less tractable. The rodent work showed

showed model-driven stimulation improving performance in intact animals, which is why the topic

model-driven stimulation improving performance in intact animals, which is why the topic appears in

appears in discussions of [[intelligence-amplification]] and

discussions of [[intelligence-amplification]] and [[human-enhancement]]. Nothing in humans supports

[[human-enhancement]]. Nothing in humans supports it: no healthy person has received such a device,

it. No healthy person has received such a device, and the risk-benefit calculation for elective

and the risk-benefit calculation for elective intracranial surgery in someone with normal memory is

intracranial surgery in someone with normal memory is not close. The realistic alternatives for

not close. The realistic competitors for memory improvement in healthy people remain unimpressive

improving memory in healthy people remain unimpressive for other reasons: the thin evidence behind

for different reasons — the thin evidence behind [[nootropics]], and the replication problems that

[[nootropics]], and the replication problems that dog cognitive claims for

dog cognitive claims for [[non-invasive-neuromodulation]].

[[non-invasive-neuromodulation]].

The ethical questions that follow are unusually sharp because memory is constitutive of the self in

The ethical questions are unusually sharp because memory is part of what constitutes the self in a

a way that hearing is not. A device that shapes what is retained shapes who a person becomes, which

way that hearing is not. A device that shapes what is retained shapes who a person becomes, which

connects directly to [[personal-identity-and-continuity]]. A device that records hippocampal

activity generates data from which content may eventually be inferred, the concern formalized under

[[mental-privacy]] and legislated under [[neurorights]]. And the possibility of selective

[[mental-privacy]] and legislated under [[neurorights]]. Selective dampening of memories is already

dampening — already pursued pharmacologically with reconsolidation blockade — raises the question of

pursued pharmacologically through reconsolidation blockade, and its possibility raises the question

whether removing a memory is a therapy or an erasure of testimony.

of whether removing a memory is a therapy or an erasure of testimony.

## Outlook

The credible near-term path is narrow: state-triggered stimulation, delivered by a device already

implanted for epilepsy, to improve encoding in patients with documented memory impairment. That is

an incremental extension of responsive neurostimulation hardware, and it does not require solving

the content problem. Chronic implants would also allow the first test of whether these effects

the content problem. Chronic implants would also allow the first test of whether these effects last

survive beyond a session, which is the single most important unknown.

beyond a session, which is the single most important unknown.

The far-term version — a device that restores lost memories or writes new ones — requires an

The far-term version, a device that restores lost memories or writes new ones, would need an

encoding model that generalizes across individuals and across content, stable long-term recordings

from thousands of hippocampal neurons, and a write channel with cell-type specificity that

from thousands of hippocampal neurons, and a write channel with a cell-type specificity that

electrical stimulation cannot provide. Each of those is a research programme rather than an

engineering task. The question that determines whether the field is on a path to anything more than

engineering task. Whether the field leads to anything more than a modest clinical adjunct depends on

a modest clinical adjunct is whether hippocampal codes are learnable at all from the outside, or

whether hippocampal codes can be learned from the outside at all, or whether they are private in the

whether they are private in the technical sense: idiosyncratic, drifting, and legible only to the

technical sense: idiosyncratic, drifting, and legible only to the brain that built them.

brain that built them.

## See also