Anesthesia and Memory: Awake but Not Recording

Introduction

In 1993, a small team at Addenbrooke's Hospital in Cambridge ran an experiment that should be far more famous than it is. They gave volunteers a benzodiazepine called midazolam, showed them pictures and words, and later tested what they remembered. Nothing unusual so far. The clever part was the timing. Some volunteers got the drug before they studied the material. Others got it before the memory test. And a few got it at both points.

The result was clean and slightly unsettling. Amnesia appeared only in the people who received the drug before studying. Giving it before the test changed nothing. Giving it before the test did not even partially rescue the memories lost by giving it before study [1]. The drug had not erased anything. It had never let the memories form in the first place.

That is the story of this article. Not the familiar worry about waking up foggy after surgery, though we will get to that, but something stranger and more specific. Certain drugs can switch off the brain's ability to write new episodic memories while leaving the person awake, responsive, conversational, capable of arithmetic and able to follow instructions. Awareness stays on. Recording stops.

A scope note before we go further, because the claim in this article is narrow and precision matters here. Everything about the clean separation of awareness from encoding refers to benzodiazepines, mainly midazolam, at low sedative doses. It is not a claim about general anesthesia at surgical depth, where consciousness itself is suppressed and the question of what is being blocked becomes much harder to answer. It is also not clinical guidance. This article explains mechanism and reports what studies found. It gives no advice to anyone about their own surgery, their own medication or their own recovery, and nothing here should be read that way. Those conversations belong with a doctor.

With that said, the science underneath is remarkable. More than 300 million major surgical procedures happen worldwide each year [2], which means the pharmacology of forgetting is being deployed on an industrial scale every single day. And the drugs used to do it have turned into one of the best research tools memory science ever got, because they let researchers interrupt a single stage of memory formation, on purpose, reversibly, in a healthy human volunteer who can then tell you exactly what happened.

The experiment that found the door

To see why the Cambridge result matters, you have to accept a claim that sounds obvious but is not: memory is not one thing. It is a sequence.

Something happens. Your senses register it. Attention selects a fraction of it. That fraction gets converted into a neural representation, which is called encoding. The representation then goes through a slow stabilisation process called consolidation, which continues for hours and in some accounts for years. Finally, later, a cue arrives and the stored representation is reactivated, which is retrieval.

Break any of those links and the person cannot report the event. But the reason they cannot report it is completely different in each case, and from the outside the failures look identical. This is why memory research has always struggled to say where a given deficit lives. If someone cannot recall yesterday's lunch, was the lunch never encoded, never consolidated, or simply not accessible right now? Almost nothing in ordinary life distinguishes these.

Midazolam distinguishes them, because you can decide when to give it.

Polster and colleagues used a recognition memory task, which is the standard laboratory measure of episodic memory. Their design crossed drug timing with test timing, and the outcome fell out with unusual clarity. Amnesia followed pre-study administration. Post-study, pre-test administration produced no amnesia, and, critically, no rescue either [1]. If the drug had blocked retrieval, giving it only at test should have caused amnesia. It did not. If the drug had blocked retrieval, then removing it at test should have recovered the lost material. It did not do that either.

There is a second layer to the same study that gets quoted less often. Alongside the recognition test, the researchers ran an implicit task in which degraded pictures and words become easier to identify if you have seen them before, whether or not you consciously remember seeing them. That task was relatively unimpaired. So the same drug, in the same session, in the same person, abolished conscious recognition while leaving unconscious perceptual traces of the very same items [1]. Something was going in. It just was not going into the system that produces a feeling of remembering.

The finding replicated in children. A study of 26 young cancer patients under midazolam sedation, with a mean age of 12.5 years, found the same dissociation: visual recognition memory was impaired, visual perceptual facilitation was not.

And the encoding localisation replicated too, in a different form. When 40 children aged four to ten were randomised, double blind, to 0.2 mg per kilogram of intranasal midazolam or placebo, the drug reduced anterograde recall and recognition, meaning memory for things shown after the dose, but did not touch retrograde memory for things shown before it [3]. Whatever was already in the brain stayed in the brain. Only the writing head stopped working.

If the three stage model of memory is unfamiliar, it is worth reading about how retrieval cues interact with the original learning episode, because the same architecture explains why the encoding specificity principle makes some memories feel unreachable without being gone.

Three stages, one target

Here is the sequence, with the point of drug action marked. Everything upstream of the marked node keeps running normally at sedative doses, which is exactly why the person appears fine.

Notice the branch. Working memory feeds two different places. One path leads towards permanent storage. The other leads straight out to behaviour, which is why a sedated patient can answer a question sensibly while forming no record of having done so. Cut the encoding link and the behavioural branch is untouched. The person is still there. The transcript just stops being written.

What survives when memory does not

If the encoding claim is right, then a whole set of cognitive abilities should stay intact under a dose that produces dense amnesia. They do, and the list is longer than most people expect.

Short term and working memory hold up. Under a dose sufficient to abolish later episodic recall, participants can still keep material in mind across short intervals and manipulate it [4]. That matters conceptually, because it kills the simplest possible explanation. If the drug worked by making people too sleepy to pay attention, working memory would collapse first. It does not.

Semantic retrieval holds up as well. Hirshman, Passannante and Arndt asked people under midazolam to generate members of a semantic category. Performance was largely spared, even though those same category members would later be unrecallable on an episodic test [5]. So the volunteer can pull the word "sparrow" out of their existing knowledge of birds, and twenty minutes later have no memory of having done so. The knowledge base is fine. The diary is not being written.

Then there is the flumazenil evidence, which is the cleanest argument that amnesia is not just a side effect of drowsiness.

Flumazenil is a benzodiazepine antagonist. It knocks midazolam off the receptor. If amnesia were simply what sedation looks like from the inside, then reversing sedation should reverse amnesia in lockstep. It does not. In a double blind multicentre study of 240 patients given flumazenil, at a mean dose of 0.7 mg intravenously, against 114 given placebo, complete reversal of sedation occurred in 80 percent of the flumazenil group compared with 30 percent on placebo at five minutes, and psychomotor performance normalised in 80 percent against 28 percent. Amnesia reversal lagged behind at roughly 70 percent [6].

The mirror image is more striking still. When volunteers were pretreated with flumazenil before receiving diazepam, the sedative and attentional effects of the benzodiazepine were blocked, and yet a marked impairment of episodic memory still appeared. Two effects, two dose response relationships, one drug.

This is the same logic that makes recognition and recall dissociate in ordinary memory experiments. Two behaviours that look like they must be measuring the same underlying thing turn out, under the right manipulation, to come apart completely.

The two curves that refuse to overlap

The Cambridge study established where the drug acts. A programme of work at Memorial Sloan Kettering, led by Robert Veselis, established something arguably more useful for clinical thinking: how much drug it takes.

The 1997 study is the anchor. Sixty seven healthy volunteers were randomised, double blind, to intravenous midazolam, propofol, thiopental, fentanyl with ondansetron, ondansetron alone or placebo. Each received three increasing and then two decreasing target concentrations through computer controlled infusion. The team then normalised drug concentrations to equal sedative effect using pharmacodynamic modelling, so that every drug could be compared at the point where it produced the same amount of sleepiness [7].

At that equalised point, the drugs behaved completely differently. Propofol impaired memory as much as midazolam. Thiopental produced only mild memory effects. Fentanyl produced none at all. Same sedation, radically different amnesia. The accompanying editorial in the same issue made the implication explicit: memory effects have their own specificity and are not a byproduct of the sedative effect [8].

Four years later the same group added an electrophysiological layer. In 65 volunteers held at five stable target concentrations, the N2 latency of the auditory event related potential tracked reaction time across every group, with a prediction probability between 0.58 and 0.71, while a different component, the N2P3 amplitude, was the better predictor of memory performance for midazolam at 0.63, propofol at 0.62 and thiopental at 0.66 [9]. Two brain signals, two behavioural outcomes, moving separately.

A 2004 study pushed further into the shape of the deficit. Testing recognition 225 minutes after infusion, the group found that propofol and midazolam interfered with the primacy effect, the advantage for items at the start of a list, while leaving the recency effect alone. Thiopental barely touched the serial position curve at all. Propofol and midazolam also reduced the event related potential difference between new and just recognised items, a marker associated with familiarity [10]. Primacy is generally taken to reflect the extent to which early items got rehearsed into longer term storage. Recency reflects what is still sitting in the short term buffer. So the drugs ate the part of the curve that depends on transfer to durable storage and left the buffer intact. Which is, again, an encoding story.

Now for the honest complication, because this article would be misleading without it. The same research group later argued against the simplest version of the encoding account. A 2008 paper from Veselis and colleagues, working with low dose propofol, concluded that the amnesia they observed was not purely a failure of encoding, and that processes beyond the initial registration of the material were involved [11]. That result does not overturn the midazolam timing experiments. It does say that different drugs, at different doses, may be doing subtly different things, and that "it blocks encoding" is a clean summary rather than a complete one. Anyone who tells you this field is settled has not read enough of it.

A short history of the amnesia problem

The idea that a drug could remove memory without removing the person is younger than surgical anesthesia itself, and the tools to test it are younger still.

Two entries on that list are doing quiet work. The 1957 case of Henry Molaison arrives a century after ether and gives the field its first clear picture of what a permanently blocked encoding stage looks like. The 1970 interview method arrives thirteen years after that and gives the field a way to ask patients what they experienced without leading them. Everything after depends on both.

The receptor that stays switched on

Behavioural experiments tell you what a drug does. They do not tell you how. For that, the work of Beverley Orser's group at the University of Toronto is the most complete story available, and it comes with a caveat that has to be stated up front rather than buried at the end.

All of what follows is rodent evidence. It has not been established in humans. The chain from receptor to behaviour is elegant and internally consistent in mice, and it is entirely reasonable to describe it as the leading mechanistic hypothesis, but nobody has demonstrated it in a human brain, and the article would be dishonest to imply otherwise.

With that stated, here is the chain.

Benzodiazepines and drugs like etomidate act on the GABA-A receptor, the brain's main inhibitory receptor. GABA-A receptors are not uniform. They are assembled from different subunits, and the particular subunit mix determines where a receptor sits and what it does. The alpha-5 subunit is unusual in that it is heavily concentrated in the hippocampus, the structure most closely tied to forming new episodic memories, and it generates a persistent background inhibition called a tonic current rather than the brief synaptic pulses most GABA-A receptors produce.

In 2006, Orser's group showed that etomidate sharply increased that alpha-5 tonic current in hippocampal pyramidal neurons. They then showed that etomidate reduced long term potentiation, the synaptic strengthening widely regarded as the cellular substrate of learning, in the CA1 region of wild type mice, but not in mice genetically lacking the alpha-5 subunit. Those alpha-5 null mice were also protected behaviourally: etomidate impaired hippocampus dependent learning in normal mice and failed to impair it in the knockouts. And here is the part that matters most for this article's argument. The sedative and hypnotic effects of etomidate, measured by rotarod performance, loss of righting reflex and spontaneous motor activity, were identical in both groups. The genetic manipulation removed the amnesia and left the sedation completely intact [44].

A follow up in 2009 tied the same knot pharmacologically rather than genetically, using an alpha-5 preferring compound to modulate synaptic plasticity and memory blockade in hippocampal slices and in behaving animals [12].

Then came the finding that genuinely surprised people. Everyone had assumed that once an anesthetic clears the body, receptor function returns to baseline. In 2014, publishing in the Journal of Clinical Investigation, the group reported that a single in vivo dose of etomidate produced a sustained increase in alpha-5 receptor function that long outlasted the drug. Cell surface expression of alpha-5 receptors in the hippocampus rose to 128 percent of control at 24 hours and 130 percent at one week, returning to baseline only by two weeks. Total expression never changed, so this was trafficking to the membrane rather than new synthesis. The increased tonic current persisted at 72 hours and at one week and was gone at two weeks [13].

The causal test came next. Treating the animals with L-655,708, a compound that preferentially dials down alpha-5 receptors, at 0.5 mg per kilogram thirty minutes before testing, reversed the memory deficit [13]. The receptor was not correlated with the deficit. It was producing it.

One more detail from that study deserves attention, because it sets up a comparison we will return to. Dexmedetomidine, a sedative that works through an entirely different receptor system, served as the active control. It sedated the animals. It did not impair the memory task, and it did not increase the alpha-5 tonic current. Sedation without the receptor change, and therefore sedation without the amnesia.

More recent work has refined rather than confirmed the original picture. Selectively removing alpha-5 receptors from interneurons, rather than from the pyramidal cells originally implicated, prevents etomidate from blocking hippocampus dependent memory, which suggests the circuit is more indirect than the first model assumed. That is how mechanistic neuroscience usually goes, and it is a reason to hold the whole account loosely.

If the hippocampus and its role in selecting what gets kept is new territory, the wider story of how the hippocampus decides what to remember gives the necessary background for why a drug acting there produces such a specific deficit.

The rhythms question, and what it cannot yet answer

Zoom out from single receptors and you reach brain rhythms. A 2024 review in Frontiers in Neuroscience gathered what is known about how anesthetics reshape the oscillations that memory formation appears to depend on [14].

The rhythms in question are familiar to anyone who has read about sleep and learning. Theta and gamma oscillations dominate during active encoding, and the coupling between them, where fast gamma bursts ride on the phase of slower theta waves, is thought to organise the timing of neural firing precisely enough for synaptic strengthening to occur. During slow wave sleep, a different triple coupling matters: slow oscillations, sleep spindles and hippocampal sharp wave ripples nest inside one another with increasing temporal precision, and this arrangement appears to coordinate the reactivation and redistribution of memory traces between hippocampus and cortex.

Anesthetics disturb all of these. They alter theta and gamma power, they change spindle and slow oscillation structure, they interfere with spike timing dependent plasticity, and they disrupt the cross frequency coupling that binds the rhythms together. The effects show up even at concentrations well below those needed for surgical anesthesia.

There is a striking human demonstration of the subanesthetic point. Alkire and colleagues gave volunteers very low concentrations of sevoflurane while showing them emotional and neutral images. Normally, emotionally arousing material gets a memory advantage. That advantage was present at placebo, at 0.1 percent and at 0.2 percent sevoflurane. At 0.25 percent it disappeared, and the imaging pointed to suppressed connectivity between amygdala and hippocampus [15]. A concentration far too low to make anyone unconscious was enough to remove a specific memory enhancement.

Now the necessary restraint. It would be easy to write a sentence claiming that disrupted oscillations cause anesthesia induced amnesia. The review's own authors decline to. They state directly that the causal relationship between EEG rhythms and memory impairment remains to be fully elucidated, and that the consequences of disrupted coupling for amnesia require further investigation [14]. Rhythm changes and memory failure occur together under these drugs. Which one drives which, or whether both follow from something upstream, is not established. That is the current state of the field and it should not be rounded up.

Anyone interested in the sleep side of this coupling story will find the relevant background in the account of what the sleeping brain does with the day's memories, since it is the same slow oscillation and spindle machinery that anesthetics disturb.

The hand that moves but does not remember

Everything so far has been laboratory work on volunteers. The operating theatre supplies a harder test, and the method is close to unbelievable when you first hear it.

Most general anesthesia involves neuromuscular blocking drugs, which paralyse the patient. Paralysis is useful surgically and catastrophic experimentally, because a patient who happens to be conscious has no way to say so. The isolated forearm technique solves this. A cuff is inflated on one arm before the paralysing drug is given, so the blood carrying the drug never reaches that limb. The arm stays movable. The anesthetist can then ask the patient to squeeze their hand.

In 2017, an international multicentre study applied this at six sites in 260 adult patients, testing immediately after tracheal intubation. Twelve patients, 4.6 percent, responded to command. Five of those twelve, when asked a second question, signalled that they were in pain. The responders were younger than the non responders, at 39 plus or minus 17 years against 51 plus or minus 16, and showed sympathetic activation far more often, at 50 percent against 2.4 percent [16].

Then the finding that anchors this entire article. Of the 253 patients who were questioned afterwards, none had explicit recall of any of it [16].

Read that again slowly. One in twenty patients was conscious enough to understand a spoken instruction and execute a voluntary motor response, some of them while experiencing pain, and essentially none of them remembered afterwards. Not a fuzzy memory. No memory.

A meta analysis of the technique across 22 studies and 1131 patients found responsiveness in 34.8 percent overall, less common during induction at 19.7 percent with a confidence interval of 17.5 to 22.1, and more common during maintenance at 31.2 percent with a confidence interval of 27.8 to 34.8. Almost none of the responders reported postoperative recall [17]. A later multicentre study in younger adults found responses in 37 of 338 subjects, or 11 percent, with women about twice as likely to respond as men.

This is the empirical heart of the matter. Under real clinical conditions, awareness and memory come apart on a routine basis. Not as an exotic laboratory artefact. As a normal feature of how these drugs work.

It is also why the philosophical framing offered by Sanders, Tononi, Laureys and Sleigh carries weight. They distinguish three states rather than two: connected consciousness, in which the person is conscious and experiencing the environment, disconnected consciousness, in which they are conscious but not experiencing the environment, as in dreaming, and genuine unconsciousness. Their conclusion is deliberately uncomfortable, that the apparent absence of subjective experience under anesthesia may sometimes be an illusion produced by amnesia [18]. We infer that nothing happened because nobody can report it. That inference is not safe.

What the monitors could and could not fix

Given all this, the obvious response is to build a device that measures depth of anesthesia. One exists. The bispectral index, or BIS, processes the EEG into a single number, and the target range is 40 to 60.

The first big trial looked excellent. B-Aware randomised 2463 high risk adults to BIS guided anesthesia or routine care and reported an 82 percent reduction in the risk of awareness, with a confidence interval running from 17 to 98 percent, and a number needed to treat of 138 [19]. The trial also gave the standard background figure, that awareness affects roughly 0.1 to 0.2 percent of surgical patients.

Then the comparison changed. B-Unaware asked a sharper question: is BIS better than simply monitoring the concentration of anesthetic gas the patient is breathing out, which is cheap and already routine? Across 967 BIS patients and 974 gas guided patients, there were two definite cases of awareness in each group. The absolute difference was zero, with a confidence interval of minus 0.56 to plus 0.57 percent [20]. BAG-RECALL, run in a high risk population, went further and found fewer awareness episodes in the gas guided group than in the BIS group [21].

So the technology works against doing nothing and does not beat the simple alternative.

The measurement problem runs deeper than any monitor, though, and the UK's fifth National Audit Project exposed it beautifully. NAP5 collected reports of accidental awareness during general anesthesia across the country. Based on patients spontaneously coming forward, the incidence was about 1 in 19,600, with a confidence interval of 1 in 16,700 to 1 in 23,450. The risk was concentrated sharply. With neuromuscular block it was roughly 1 in 8,200. Without it, roughly 1 in 135,900. In caesarean section, roughly 1 in 670. Two thirds of episodes occurred during induction or emergence rather than during surgery itself [22].

Now compare that with what happens when you ask instead of waiting to be told. The structured interview introduced by Brice in 1970 puts a fixed set of questions to every patient, and studies using it have consistently returned incidence figures around one to two per thousand.

One in 19,600 against one to two per thousand. Roughly a hundredfold difference, produced by nothing except the method of asking. Both numbers are real. They answer different questions. Spontaneous reporting measures how often awareness bothers a patient enough to volunteer it. Structured interviewing measures how often a memory can be elicited at all. The gap between them is a lesson about measurement that extends well beyond anesthesia.

The memory you cannot report but still have

If explicit memory can be switched off while awareness stays on, the next question is whether anything gets recorded at all. The answer appears to be yes, sometimes, in a form the patient cannot access voluntarily.

Implicit memory is the technical term. It refers to influences of prior experience on later behaviour that occur without any conscious recollection. The classic laboratory demonstrations use word stem completion, where seeing a word makes you more likely to complete a fragment with it later, or free association priming, where exposure biases what comes to mind.

Kihlstrom and colleagues ran the emblematic study in 1990. Paired associate words were presented to 25 surgical patients after induction with thiopental, vecuronium and isoflurane. Explicit memory was flat: no free recall, and no significant cued recall or recognition. A free association task, however, showed significant priming on both immediate and delayed testing [23]. Words the patients had no memory of hearing were changing what they said.

The result is not universal, which is a point worth dwelling on. The same group later found no implicit memory under a sufficient anesthetic regimen using different drugs. Implicit memory formation under anesthesia depends heavily on what was given and how deeply.

The 2021 meta analysis quantified this across 61 studies, 3906 patients and 119 cohorts. Implicit memory events were reported in 43 of the 119 cohorts, or 36.1 percent. Three odds ratios stand out. Patients with ASA physical status III to IV, meaning more severe systemic disease, were more likely to form implicit memories than healthier patients, with an odds ratio of 3.48 and a confidence interval of 1.18 to 10.25. Deep sedation was much less likely to produce implicit memory than general anesthesia, with an odds ratio of 0.10 and a confidence interval of 0.01 to 0.76. And premedication with a benzodiazepine before general anesthesia reduced implicit memory formation, with an odds ratio of 0.35 and a confidence interval of 0.13 to 0.93 [24].

That last number closes a loop rather satisfyingly. The drug class that blocks explicit encoding also reduces the formation of implicit traces. Whatever the benzodiazepines are doing to the hippocampus is not confined to the system that produces conscious recollection.

A companion meta analysis on the perception of auditory stimuli under anesthesia found implicit memory on word stem completion in one propofol and nitrous oxide day case group but not in two studies using total intravenous propofol with benzodiazepine premedication [25]. Regimen matters more than the general question of whether anesthesia permits learning.

The idea that a trace can exist without being retrievable connects directly to the broader phenomenon of memories that exist but cannot be reached, which turns out to be far more common in ordinary cognition than most people assume.

The strange gift of forgetting forward

Here is the finding that surprises even people who work in this area. Benzodiazepines make you worse at remembering things learned after the dose. They make you better at remembering things learned before it.

The effect is called retrograde facilitation, and the definitive early treatment came from Hinrichs, Ghoneim and Mewaldt in 1984. They tested three competing explanations across three experiments. Was the improvement due to differential effort? No. Was it due to some direct enhancement of consolidation? Also no. What survived was interference reduction: because the drug prevents new information from being learned, there is less new material competing with the older memories during the consolidation window, so those older memories emerge stronger [26].

Later work extended the effect across the benzodiazepine class and found it to be dose dependent. In healthy older adults, lorazepam produced dose dependent retrograde facilitation on verbal memory, with an F value of 15.358 and a p value below 0.001 [27]. It has been shown with triazolam, midazolam and diazepam, and in 2022 it was demonstrated in mice on object recognition and object location tasks, resolving an older discrepancy where rodent studies had been inconsistent.

The interference reduction account has a much wider reach than pharmacology. It is essentially the same argument used to explain why sleep benefits memory. Sleep is a period in which very little new encoding happens, which protects recently formed traces from retroactive interference. Alcohol produces retrograde enhancement for the same reason. So does a benzodiazepine. Three completely different states, one shared property: they all stop the flow of new memories, and older memories benefit.

This is the point in the article where anesthesia research stops being about anesthesia and starts being about learning in general. The reason spaced practice works, the reason a fresh night's sleep beats another hour of cramming, and the reason a sedative can paradoxically improve recall of the last thing you read before taking it are all versions of the same constraint. Consolidation is a limited resource, and new learning competes for it. The same competition sits underneath the shape of the forgetting curve itself.

Four doors into the same room

Not every amnestic drug works the same way. Comparing them shows that memory has several independent points of failure, and that the relationship between consciousness and encoding differs depending on which door you go through.

Ketamine is the most instructive contrast because it bypasses GABA entirely. It blocks NMDA receptors, and NMDA receptor activation is required to induce long term potentiation. The foundational demonstration came from Morris and colleagues in 1986, who showed that an NMDA antagonist prevented hippocampal LTP and selectively impaired spatial learning, while the inactive isomer of the same compound did nothing [28]. The human data match the encoding pattern. Krystal and colleagues gave 0.1 and 0.5 mg per kilogram intravenously over 40 minutes to 19 subjects and found that ketamine preferentially disrupted delayed word recall while sparing immediate recall [29]. Morgan and colleagues, using 0.4 and 0.8 mg per kilogram, found dose dependent impairment of episodic and working memory with perceptual priming and executive function untouched [30], and a later study targeting plasma concentrations of 50 and 100 nanograms per millilitre concluded that the damage was attributable primarily to encoding rather than retrieval [31]. A review by the same group noted that ketamine may distinctively affect semantic memory, unlike the GABAergic and anticholinergic amnestics [32].

Dexmedetomidine occupies the opposite corner. It works on alpha-2 adrenergic receptors and produces a state that resembles natural sleep, from which patients can be roused and then drift back. In seven volunteers given 50 minute infusions at 0.2 or 0.6 micrograms per kilogram per hour after a loading dose, sedation reached 30 to 60 percent and memory impairment about 50 percent [33]. So it is not memory neutral in humans at sedating doses. What makes it interesting is the mouse data, where it sedated without impairing memory and without touching the alpha-5 current [13]. Different receptor, different relationship between the two effects.

Scopolamine is the oldest model of the four. It blocks muscarinic acetylcholine receptors and has been used to mimic age related memory decline since Drachman and Leavitt's work in the 1970s. Ghoneim and Mewaldt tested it head to head against diazepam and found that scopolamine at 8 micrograms per kilogram and diazepam at 0.3 mg per kilogram both impaired the learning of new material while leaving recall of pre-drug material untouched [34]. Same behavioural signature, different neurotransmitter system entirely.

Four receptor systems. One converging deficit. Encoding is not a single molecular process that can only be broken one way. It is a coordinated state that several independent mechanisms can pull apart.

The drunk who was awake and the man who lost his hippocampus

Two non pharmacological cases make the same point, and both are more familiar than anything above.

An alcohol blackout is an encoding failure in someone who is fully conscious. Aaron White's review lays out the essentials. Alcohol interferes primarily with the formation of new long term memories while leaving established memories and short term retention intact. There are two forms. Fragmentary blackouts, sometimes called greyouts, are partial and can often be filled in with cues. En bloc blackouts are complete, permanent, and cannot be recovered under any circumstances, because there is nothing to recover [35]. Blackouts tend to begin at blood alcohol concentrations around 0.16 percent, roughly twice the legal driving limit in many jurisdictions, and rapid rises in concentration make them more likely. Mechanistically, alcohol suppresses hippocampal pyramidal cell activity.

The crucial detail is what the person is doing during a blackout. They are walking, talking, making decisions, holding conversations that other people remember perfectly well. They are awake. There is simply no recording. This is the same dissociation the isolated forearm studies found in the operating theatre, occurring every weekend in ordinary life.

At the other extreme sits Henry Molaison. In 1953 he underwent bilateral medial temporal lobectomy for intractable epilepsy, removing the hippocampal formation, most of the amygdala and the entorhinal cortex. Scoville and Milner published the case in 1957, reporting a persistent and severe impairment of recent memory whose degree tracked the extent of hippocampal removal, alongside intact early memories and intact technical skills [36]. Later MRI work confirmed the anatomical extent of the lesion.

What H.M. retained is as informative as what he lost. Working memory functioned. He could hold a conversation. He could learn new motor skills, improving at mirror drawing across days while insisting each morning that he had never done the task before. Conscious autobiographical recording was gone. Everything else was there.

Put the three side by side. A surgical lesion, a toxin, and a benzodiazepine. Permanent, transient and reversible. Structural, metabolic and pharmacological. All three produce a person who is present, responsive and unable to write the record. That convergence, from three completely unrelated directions, is the strongest available argument that encoding is a genuinely separable stage rather than a convenient abstraction.

What anesthesia and memory research still cannot say

This is where the article gets careful, because the most common question about anesthesia and memory is not the one this piece has been answering.

People do not usually want to know whether encoding is a separable stage. They want to know whether anesthesia damages memory afterwards. That is a different question with a much messier evidence base.

The landmark study is ISPOCD1, published in the Lancet in 1998. It tested 1218 patients aged 60 and over before and after major non cardiac surgery. Postoperative cognitive dysfunction was present in 25.8 percent of patients at one week, with a confidence interval of 23.1 to 28.5, and in 9.9 percent at three months, with a confidence interval of 8.1 to 12.0. Control figures from UK participants were 3.4 percent and 2.8 percent respectively. Risk factors for early dysfunction included increasing age, longer anesthesia duration, less education, a second operation, postoperative infections and respiratory complications. For dysfunction persisting to three months, only age remained a risk factor. Hypoxaemia and hypotension were not significantly associated with the long term outcome [37].

A large population study reached similar territory from another direction. Within the Mayo Clinic Study of Aging, 1819 participants aged 70 to 89 at enrolment underwent cognitive assessment roughly every 15 months. Exposure to anesthesia and surgery after enrolment was associated with a small decline in cognitive scores, slightly steeper than the decline expected from normal ageing [38].

Here is the plain statement that most coverage of this topic leaves out. The causal contribution of the anesthetic, as distinct from the surgery itself, and as distinct from whatever underlying illness made the surgery necessary, cannot currently be separated. Every one of these studies is observational. People who have major surgery differ systematically from people who do not, in ways that are themselves associated with cognitive trajectory. Surgery produces inflammation, pain, sleep disruption and physiological stress, any of which could account for what is observed. The authors of the Mayo analysis say so explicitly. Nobody has run, and for obvious ethical reasons nobody can run, the trial that would settle it in adults.

Terminology has moved to reflect this uncertainty. In 2018, a consensus working group led by Evered published new nomenclature simultaneously in six journals, replacing the old catch all "postoperative cognitive dysfunction" with a family of terms under the heading perioperative neurocognitive disorders, covering preoperative neurocognitive disorder, postoperative delirium, delayed neurocognitive recovery up to 30 days, and postoperative neurocognitive disorder up to 12 months [39][40][41]. The renaming was not cosmetic. It was an admission that the field had been using one label for several distinct phenomena with different time courses and probably different causes.

The children's literature has actually resolved more cleanly, largely because it was possible to randomise. The GAS trial compared just under an hour of sevoflurane general anesthesia in infancy against awake regional anesthesia and found no difference in neurodevelopmental outcome at five years, using full scale IQ as the primary measure, having found no difference at two years either [42]. The PANDA study compared 105 sibling pairs where one sibling had a single anesthesia exposure before 36 months, averaging 80 minutes, and tested IQ between ages 8 and 15. No significant difference emerged on IQ or on secondary neurocognitive and behavioural measures [43]. The MASK study found no association between exposure and general intelligence, though multiple exposures were linked to modest reductions in processing speed and fine motor skills.

The reasonable summary of the paediatric evidence is that a single brief exposure in early childhood does not appear to harm general cognitive development, while multiple exposures carry a small, domain specific signal that is worth watching.

None of this changes the core argument of the article. The dissociation between awareness and encoding is a fact about how these drugs act during administration. What happens weeks or years later is a separate question, with weaker evidence and much greater uncertainty, and it deserves to be described that way rather than folded into the same confident story. For the wider picture of how memory shifts across the lifespan independently of any medical intervention, the account of how ageing changes memory provides the necessary baseline against which surgical effects have to be judged.

Conclusion

Start again with the volunteer in Cambridge.

They are sitting up. They can tell you their name, the date, what they had for breakfast. Show them a list of words and they will hold the last few in mind and repeat them back. Ask them to name birds and they will produce sparrow, robin, crow without hesitation. Ask them to squeeze your hand and they will. By every behavioural measure available in that moment, the lights are on.

Twenty minutes later, none of it exists.

That gap is the whole point. For most of the history of thinking about the mind, consciousness and memory were treated as more or less the same faculty. To experience something was to record it. If you could not recall an event, the reasonable inference was that you had not been there.

Midazolam breaks that inference. So does an alcohol blackout, and so did the surgery performed on Henry Molaison in 1953. Three completely different interventions, one shared result: a person who is fully present and entirely unable to keep what is happening to them.

What the anesthesia literature adds is precision. Polster's timing experiments pinned the deficit to encoding rather than retrieval. Veselis showed the memory dose response curve pulling away from the sedation curve, drug by drug. The flumazenil work showed that you can strip away sedation and leave amnesia standing. The isolated forearm studies showed the whole thing playing out in real operating theatres at a rate of roughly one patient in twenty, with essentially none of them remembering. And in rodents, though not yet in people, the alpha-5 subunit of the GABA-A receptor supplies a plausible molecular address for where the block occurs.

Plenty remains open. Whether disrupted brain rhythms cause amnesia or merely accompany it is unresolved, and the authors of the leading review say so. Whether low dose propofol amnesia is purely an encoding failure is disputed by the same group who did most to establish the encoding account. Whether anesthesia contributes anything to long term cognitive decline in older patients, beyond what surgery and underlying illness contribute, cannot be answered with the study designs available.

But the central observation survives all of that, and it is worth carrying away. Being awake and building a memory are two separate jobs, performed by overlapping but distinguishable machinery, and a small molecule binding to the right receptor can leave one running while the other stops. Your sense that you were conscious for the last hour rests almost entirely on the fact that you remember it. Those are not the same thing. They only usually travel together.