Commit To Disk

In 1885 Hermann Ebbinghaus measured his own forgetting by learning lists of nonsense syllables and relearning them later, timing how much easier the second pass went. Over the winter of 2011 a twenty-two-year-old Dutch student did the whole thing again on himself, about seventy hours of it, at the same seven intervals Ebbinghaus had used: twenty minutes, an hour, nine hours, a day, two days, six days, a month.

Half of the famous result held. Most of what the student lost, he lost inside the first hour, exactly as advertised. The other half did not. A month out, Ebbinghaus had still found relearning about a fifth easier than starting fresh, while the student was down to four percent, which is close to starting over. The cliff is solid, the long flat tail printed in textbooks is softer than advertised, and both curves come from a sample size of one person.

That is long-term storage. This series treats thinking as a panel of twelve separate functions rather than one score, and this is the one where the hardware metaphor stops helping and starts lying to you. A drive holds what you wrote and hands it back unchanged. This does neither.

Working memory is the RAM, processing speed is the clock, visuospatial processing is the graphics card. Storage is the drive, the part that keeps what the rest of the machine has finished with. On the panel it is called long-term storagelong-term storageOn the Cattell-Horn-Carroll map of mental abilities this is Gl: how efficiently new material gets learned and consolidated into durable memory. Its neighbour, Gr or retrieval fluency, covers how fast that material comes back out. The field treated the two as one construct until recently., and its one job is keeping what happens to you available later.

Three things have to go right. You take the thing in, you keep it, and you get it back. This is about the middle one, which nobody notices because it happens while you are doing something else.

The keeping runs on two clocks. Over minutes and hours the synapses carrying a new memory physically stabilise, which is why a blow to the head can erase the last few minutes and leave the rest untouched. Over weeks and years the hippocampus, a pair of curled structures deep in the temporal lobes, gradually hands the memory over to the cortex. A three-month scanning study caught that in progress: as the memories aged, hippocampal activity during successful recall kept falling while activity in a region of the frontal cortex kept rising.

We know the hippocampus does this job because of Henry Molaison, who in 1953 had most of his medial temporal lobes removed to stop his seizures. It worked. It also left him unable to form a new lasting memory for the remaining fifty-five years of his life, with his intelligence, his language and his childhood intact. He could hold a conversation. He could not remember having had it. The tidy version of that story oversells itself: a reconstruction of his brain after his death in 2008 found more surviving hippocampal tissue than his lifetime scans had suggested, plus undocumented damage in the frontal cortex and the cerebellum.

The cleaner demonstration that storage splits into kinds came from a Canadian patient known as K.C., who after a motorcycle accident kept his vocabulary, his general knowledge and the technical details of his old job, and lost the ability to recall a single episodeepisodeEpisodic memory is memory for specific experienced events, the kind you re-live rather than merely know. Its counterpart, semantic memory, holds facts stripped of the occasion on which you learned them. of his own life from any point in it. He knew things. He remembered nothing.

The handoff itself is where the field splits, and the disagreement matters more than it looks. One camp holds that a memory eventually goes fully cortical and the hippocampus is done with it. The other holds that anything still episodic needs the hippocampus forever, and that what becomes independent is only the flattened, factual version of the event. Both camps have patients and scans, neither has won, and estimates for how long a transfer takes run from a few years to never.

Sit with the second position for a moment, because it is the first hint that the drive analogy is in trouble. On that account consolidation is not a copy operation at all. What ends up in the cortex is a different object from what went in, a summary rather than the episode, and the original does not survive the move intact. Storage that edits its contents on the way through is not storage in the sense a disk means it.

One number to refuse outright is the brain’s capacity in bytes. The 2.5 petabyte figure comes from a magazine answering a reader’s question with rough arithmetic about neurons and connections, and the nearby one-petabyte figure traces to a press release about a study that reconstructed a piece of rat hippocampus the size of a red blood cell. The number in the headlines is not in the paper, and nobody has measured this in a human.

Everyone has heard that memories are filed during sleep, and the proposed mechanism is elegant. In deep sleep the cortex produces slow waves, the thalamus fires spindles, the hippocampus emits fast bursts called ripples. When the three lock into the right rhythm, the hippocampus replays the day and the cortex writes it down.

The replay is real and was seen directly in rats in 1994, when place cells that had fired together while an animal ran a track fired together again in its sleep. The human evidence is thinner than the confidence around it. Nobody records from a healthy human hippocampus for weeks, so the human work comes from small groups of epilepsy patients with electrodes already implanted for surgery, and it shows patterns reactivating rather than cells firing in sequence.

You can also intervene, and this is the strongest part of the case. Pair a smell with a learning session, release the same smell during deep sleep, and recall improves the next morning. Jan Born’s group showed that in 2007, and the techniquethe techniqueTargeted memory reactivation: re-presenting a cue that was paired with learning (a smell or a sound) while someone is in slow-wave sleep, to bias which memories get consolidated. It does not work in REM sleep or while awake. has since been tried ninety-one times. Pooled across all of it the benefit is real and small. Note what that shows: nudge which memories get replayed and you change what survives. It does not follow that more deep sleep means more memory, which is a different question with a worse answer.

That question was put by Björn Rasch, who co-wrote the definitive review of sleep-dependent consolidation and then ran the study that should make everyone nervous. His group put 159 people through two naps each and asked whether the amount of deep sleep a person gets predicts how much they remember. It does not. The correlation was essentially zero, and the statistics returned strong positive evidence for no relationship at all, which is rarer and stronger than failing to find one. His reading is that the earlier positive findings came from samples too small to trust, and his review of the wider literature now carries the title “How robust are sleep-mediated memory benefits?”.

The largest synthesis pools 823 comparisons and finds that sleeping after learning does beat staying awake, by a moderate amount. Correct for the tendency of positive results to reach print and the same authors' estimate drops by more than a third. The effect also depends heavily on how the experiment was built: the design that keeps people awake all night and then lets them recover before testing finds no advantage for sleep at all. A separate group asked in print in 2022 whether sleep’s role here is simply overrated, since the mechanisms invoked for it are not unique to sleep and quiet wakefulness engages much of the same machinery.

What survives all of this is not about the night after. It is about the night before. Losing sleep before you learn something costs roughly twice what losing it afterwards costs, because the hippocampus is impaired while it is trying to take the material in. One study scanned people who had been awake around thirty-five hours as they memorised pictures, found the hippocampus visibly quieter during encoding, and measured recognition two days later down by about a fifth, after everybody had caught up on sleep.

The tests are old and good. Someone reads you a list of sixteen words five times over and asks what you can recall after each pass, hands you a second list to interfere, then comes back twenty minutes later and asks again. Or reads you a short story to retell immediately and after a delay. Or shows you faces with names attached, patterns to rebuild, a line drawing to copy and then reproduce from memory.

What matters is what gets scored, because a memory test does not produce a memory number. It produces a learning curve across the five trials, a delayed recall, a figure for how much survived the delay, a count of words you produced that were never on the list, and a recognition score for whether you can pick the words out when shown them. Those come apart, and a single composite hides which part broke.

Four traps, and this function has the nastiest set on the panel.

The first is the one the whole article turns on. Failing to produce something does not mean it was never stored. It may be in there and unreachable. Clinicians separate the two with a cue: if you cannot recall the word alone but produce it instantly on being told it was a kind of fruit, your storage is fine and your retrieval is not. One test is built entirely around that logic, teaching you the category first so it knows the material went in, then measuring how much the cue rescues. In Alzheimer’s disease the cue stops helping, which is what a genuine storage failure looks like. This is why the panel splits this function from the next one, and why somebody saying “my memory is going” has not yet said anything specific.

The second is practice, which is the same problem wearing a lab coat. Memory tests are the worst offenders on the panel for improving simply because you have seen them before. Reuse the same word list a month later and people gain a modest but real amount; switch to a matched alternate list and the gain nearly vanishes. Story recall is worse. Followed across five annual sittings, repeated story recall drifts upward year after year, and by more than the word-list test it was measured against. Retest yourself on the same material and you are measuring your familiarity with the material. The act of measuring this function alters it, which is a nuisance for the clinician and a clue about the machinery.

The third is that a bad memory score is frequently not about memory. Hearing loss is the cleanest case: give someone with hearing loss the spoken version of a word-list test and they look impaired, give them the same test on paper and they look normal, and every other score they produce is fine. Depression costs about a third of a standard deviation on episodic memory, pooled across two hundred studies. Long-term benzodiazepine use costs roughly twice that. Anticholinergic drugs drag verbal learning down, and so does a bad night before the test.

The fourth is that abnormal scores are normal. Give a neurologically healthy adult a long battery and more than half will produce at least two scores in the impaired range. Correcting for age and education does not fix that. In the study that checked, correcting made it worse. Run enough measurements on anyone and some come back low.

Line up people of different ages and episodic memory looks like it starts sliding in your thirties. Follow the same people for years and no reliable decline appears before sixty. Both pictures come out of the same Swedish project, which ran the comparison deliberately, and the disagreement is the one that runs through this whole series. A snapshot across ages carries the fact that older people grew up in a different world. Following one person carries the fact that they are getting practised at the test. Meanwhile the knowledge you have accumulated keeps rising into your fifties and holds up far better than the episodes do.

The physical version is measurable. A healthy adult loses somewhere between half a percent and one and a half percent of hippocampal volume a year, the rate rising with age. In Alzheimer’s disease it runs about three times that, and the two trajectories separate before anything is diagnosable.

Hence the clinical attention this function gets over every other on the panel. In typical Alzheimer’s disease delayed recall goes first, and the biology precedes the symptom by a long way.

The clearest look at that ordering comes from families carrying a mutation that guarantees the disease, where researchers can count backwards from the age a parent fell ill. Spinal-fluid amyloid starts drifting about twenty-five years ahead of expected symptoms. Amyloid becomes visible on a scan, tau starts rising, and the hippocampus starts shrinking around fifteen years ahead. Metabolism drops at ten, and so does episodic memory, which is measurably impaired a full decade before anyone expects symptoms. Standard tests of global cognition do not catch it until five.

Those families are the clearest window available and also the wrong population to generalise from, since inherited early-onset disease need not run on the same clock as the ordinary late-onset kind. What the sequence does establish is that this function starts failing long before the failure is called a diagnosis.

Anyone building cognitive tests should sit with what came next. A blood test measuring a form of tau protein now identifies Alzheimer’s pathology with accuracy around ninety percent, in primary care as well as specialist clinics. In the same study the primary-care doctors managing those patients were getting it right about six times in ten, and the dementia specialists about seven. Note what that comparison is. Nobody has run the blood test head to head against a memory battery, so the beaten party is the clinician’s overall judgement, which already has cognitive testing inside it. The result is striking and it is not the result people quote it as.

The word list is not thereby useless, because it answers a different question. A biomarker tells you what you have. A repeated test tells you what you can still do, and whether that is moving.

Mild cognitive impairmentMild cognitive impairmentA diagnosis of measurable cognitive decline that has not yet cost someone their independence. It sits between normal ageing and dementia, and it is a description of a current state rather than a prediction. is not a waiting room, either. A meta-analysis of forty-eight studies puts reversion to normal cognition at roughly a third of everyone diagnosed. Where the diagnosis was made matters enormously: in one pooled analysis, one in four community-diagnosed cases went back to normal against about one in twelve of those diagnosed in a memory clinic.

The same split runs the other way for progression. Pooling forty-one cohorts, people diagnosed in a specialist clinic went on to dementia at about twice the annual rate of people identified in the community. Both facts describe the same thing from opposite ends, which is that a memory clinic is where the worse cases have already selected themselves. A diagnosis carries information about the person and also about the door they walked through.

The feeling that your own memory is worse carries more information than I expected going in. Across meta-analyses, people reporting it go on to be diagnosed with dementia at roughly twice the rate of people who do not. That is a doubling of a small number rather than a sentence, and most people who feel it never progress, but it is not the reassuring shrug I was ready to write.

Then the unglamorous list, which deserves to be checked first and usually is not. Untreated sleep apnoea. Heavy drinking, where each extra seven units a week above fourteen is associated with about seventeen percent more dementia risk. Anticholinergic load. Depression, which can produce a picture convincing enough to be mistaken for dementia and which largely lifts when treated, though a substantial minority of those people turn out to have been early in something degenerative anyway. These are the causes most likely to be treatable and least likely to be dramatic.

This function is the hardest on the panel to measure honestly at home.

Every well-validated memory instrument is gated: the word lists are licensed and ship with one or a few alternate forms, the story tests are copyrighted subtest material, and the face-name and pattern tests run on platforms sold to clinics rather than people. There is no equivalent of the paper rotation test that visuospatial processing gets, where the research instrument itself has circulated freely for decades.

So there are two honest options: do nothing, or run a within-person exercise and refuse to turn it into a percentile.

If you run it, build your own word lists, never reuse one, fix the delay and write it down, and record the whole shape instead of one figure. How many on the first hearing, how many by the fifth, how many after the delay, and how many came back when you gave yourself a category cue. That last comparison is the only one separating storage from retrieval.

What you cannot do at home is make any of it mean something against other people. The norms are the product and the alternate forms are the product. A single delayed-recall number carries enough noise that only a large change is worth reacting to, so several sittings under identical conditions, months apart, are what tell you anything.

Two halves to this, and they point in opposite directions. What you do while learning works better than almost anything else on this panel. What you buy to help with it mostly does not.

Test yourself instead of rereading. Students read a passage; some reread it, others put it away and tried to recall it. Tested immediately, the rereaders won. Tested a week later, the recallers won by a wide margin, remembering a little over half the material against a little over two fifths. Push it further and the students who studied once then recalled three times beat the students who studied four times, having felt considerably less prepared while doing it. Across the meta-analyses, retrieval practice beats rereading by about half a standard deviation, and the gap widens with the delay rather than closing.

Spread the sessions out. Distributed practice beats cramming across hundreds of experiments, and the right gap scales with how long you need the material to last: revisit after a few days for a test in a few weeks, after weeks for something you want in a year. Produce the material rather than read it, which is worth about four tenths of a standard deviation on its own.

Then use a technique, and take it seriously. Trained memory competitors are not smarter than matched controls, and that has been checked: their general intellectual ability comes out the same, while their recall of a seventy-two item list runs near ceiling against about forty for the controls. They are running a spatial method on ordinary hardware. Put people with no prior training through forty days of half-hour daily sessions on that method and their recall roughly doubles, with their brain connectivity shifting toward the pattern that distinguishes the competitors.

The deflation matters more, because the interventions people actually buy are on this side of the ledger.

Exercise and the hippocampus is the case study in how a good result gets flattened by repetition. A year-long trial in older adults found the walking group gained about two percent of hippocampal volume while the stretching group lost about one and a half. A commentary in the same journal pointed out that by the end of the year the two groups' memory scores did not differ, and a later pooled analysis across many trials found no significant effect on total hippocampal volume, with the signal surviving on one side only and looking more like slowed loss than growth. Exercise still earns its place on other grounds, including the modest cognitive gains it does show. The specific claim that it grows the part of your brain that stores memories is thinner than its circulation.

Supplements are mostly a graveyard. Ginkgo went to three thousand people over seventy-five for six years and did not prevent dementia, with slightly more cases in the treated group than the placebo group. Omega-3 in healthy adults comes back at approximately nothing.

The one real exception is a plain daily multivitamin, which over three years in older adults produced a small but statistically solid benefit on memory and global cognition. Read that carefully. An ordinary multivitamin, in older adults whose baseline nutrition may have been marginal, and the headline figure of a couple of years of cognitive ageing avoided is the test-score difference converted through the trial’s own ageing slope rather than something anyone measured. It is a genuine finding about nutrition, not a memory drug.

It is certainly not an argument for the things sold as memory drugs. The two most heavily marketed memory products of the past decade were both pursued by the Federal Trade Commission. One settled for two million dollars in 2016. The other’s own supporting study found no benefit over placebo on its primary measures, with the marketed claims surviving only in after-the-fact subgroup analysis, and that case is still being litigated.

At population scale, modelling attributes about forty-five percent of dementia to fourteen modifiable risk factors across a lifetime. That is attribution arithmetic rather than trial evidence, and it supports the habits associated with lower risk without promising anything to any individual. The trials bundling those habits are honest about their size: the Finnish multidomain trial improved cognition slightly more than a control arm that also improved, and the recent American version beat its self-guided comparison by a small margin, with both arms getting better and no untreated group to measure either against.

Underneath all of it is the difference between performance and capacity. Everything in the first half of this section improves how much of what you attempt to learn actually sticks, and none of it has been shown to raise the ceiling. The competitors are the proof: same general ability as the controls, radically better recall, entirely through method.

In 2000 Karim Nader showed that reactivating a consolidated memory in a rat makes it briefly unstable, so that blocking protein synthesis right after the reminder damages a memory the same drug could not touch an hour earlier. Retrieval opens a windowRetrieval opens a windowReconsolidation: the finding that a reactivated memory returns to a labile state and has to be stabilised again, during which it can be strengthened, weakened, or altered. Solid in rodents, contested and much weaker in humans.. The memory has to be put back, and while it is out it can change. The human version is messier than its press coverage, with acute effects, failed replications, and at least one clean null on the claim that anything was permanently erased. Nobody has deleted a human memory.

The reconstruction is not in dispute. Ask people about a car crash they watched on film and the verb in your question changes the speed they report, and a week later a third of the people asked with the harsher verb remember broken glass that was never in the footage. Give people a list of related words and about four in ten lists will produce a confident recall of a word that was never presented.

A drive that altered a file slightly every time you opened it would be defective. A brain doing it is working to specification, because the job is not to reproduce the past accurately, it is to be useful the next time something similar happens, and useful means updated by everything since.

Which reframes everything the earlier sections were circling. Consolidation that arrives as a summary rather than the episode, a test score that climbs because you sat the test before, a week-old memory that survives because you pulled it out on Tuesday: those are not three quirks. They are one property seen from three angles.

I use the hardware framing across this series because it makes twelve functions legible as twelve, which is the whole argument. It is worth knowing which piece breaks it. This one does, in the most interesting place available, because here the read operation is a write.

So the practical version is smaller than the metaphor and stranger than a tip list. Sleep before you learn, not only after. Retrieve rather than reread, because the retrieving is what does the writing. And if you want to know whether this function is drifting, watch the gap between what you cannot recall on your own and what comes back the moment someone gives you a cue. Both are failures. Only one of them is a failure of storage, and the other one is the machine working exactly as designed.