The Phonological Loop: The Two-Second Buffer Behind Every Number You Repeat

Introduction

Someone reads you a phone number. You do not write it down. Instead you start saying it to yourself, quietly, on a loop, and you keep saying it until you can type it in. If anyone interrupts you, it is gone.

That small, slightly desperate act is one of the most studied behaviours in cognitive psychology. The system doing the work has a name. Alan Baddeley and Graham Hitch called it the phonological loop, and in 1974 they put it at the centre of a model of memory that has now survived fifty years of attack [1].

Almost every explanation of the loop you will find says the same thing. It holds sound for about two seconds. After that the trace fades unless you refresh it by repeating it.

That sentence is in textbooks, revision sites, university lecture slides and encyclopedia entries. It is repeated so consistently that it reads like a measurement, as though someone once watched a memory trace empty and clicked a timer at the moment it went.

Nobody did.

The two seconds was never a stopwatch reading. It came out of a study about word length, and it describes how much a person can say in roughly two seconds rather than how long a trace survives. That distinction sounds like pedantry. It is not. It changes what the number means, and once you see where it came from, you also see why a large group of researchers now argue that the decay it describes may not exist at all.

This article is about the loop, and about that number. Where it came from, what evidence built the case for the system underneath it, which parts of that evidence have held up, and which parts quietly failed to replicate while the textbooks kept printing them.

A Model Built to Replace a Simpler One

Before 1974, the standard account of memory was tidy. Richard Atkinson and Richard Shiffrin had described a system where information moved from a sensory register into a single short-term store, and from there, if it was rehearsed enough, into long-term memory. One short-term box. One job.

Baddeley and Hitch thought the box was doing too much. They ran experiments where people held a string of digits in mind while simultaneously reasoning about sentences. If there really were one general-purpose short-term store, loading it with digits should have wrecked the reasoning. It did not. Performance dropped, but far less than a single-store model predicted.

So they proposed something with parts. A central executive that directs attention, and two subsystems it commands. One for visual and spatial material, the visuospatial sketchpad. One for speech-based material, the phonological loop [1].

The model kept growing. In 1986 Baddeley split the loop itself into two pieces, which is the version most people learn. In 2000 he added a fourth component, the episodic buffer, a limited store that binds material from the other systems and from long-term memory into single episodes [2].

By 2012 Baddeley was drawing a distinction that matters for everything that follows. The overall framework, he argued, is stable. The specific models inside it are not, and were never meant to be [3]. That is an unusually honest thing for a researcher to say about his own life's work, and it gives permission to do what the rest of this article does, which is take the specific claims apart.

If you want the wider system this sits inside, we have written separately about how working memory behaves under pressure in clinical decisions, and about the role attention plays in what gets remembered at all.

The Two Parts: A Store and a Voice

The loop, in the 1986 version, has two components.

The first is the phonological store. Baddeley described it as an inner ear. It holds speech-based material passively, and it does not hold it for long. Nothing you do to it makes it last. It simply fades. What feeds it is the even briefer auditory trace we describe in sensory memory and the quarter second you miss.

The second is the articulatory rehearsal process. This is the inner voice, the thing you are using when you repeat that phone number under your breath. It has two jobs. It refreshes the fading contents of the store, which is why repetition keeps a number alive. And it converts written material into a sound-based form, which is why you can read a phone number off a screen and still end up rehearsing it as sound.

That second job explains something that surprises people. The loop is deeply involved in reading, even silent reading, because visual words get recoded into a phonological form on the way in.

The relationship between the two parts is a cycle. The store holds a decaying trace. The voice reads it out and puts it back in. Round and round, for as long as you keep paying attention, and no longer.

A note on names, because the search results are inconsistent. You will see this system called the articulatory loop. It is the same thing. The older literature used articulatory loop more often, and some study guides still lead with it. Phonological loop is the term Baddeley settled on, and it is the more accurate one, for reasons that become clear when we get to deaf signers.

Where the Two Seconds Actually Came From

In 1975, Baddeley, Neil Thomson and Mary Buchanan published the study that produced the number.

They gave people lists of words to recall in order. Some lists were made of short words. Some were made of long ones. The short lists came back far better. People who could manage a run of one-syllable words like hit, take and sum fell apart on five-syllable words like refrigerator, opportunity and university.

Then they did the thing that made the study famous. They measured how fast each person could read the words aloud, and found that reading speed predicted memory span. Their own summary of the finding is worth quoting exactly, because it is not what most people think it says:

"Span could be predicted on the basis of the number of words which the subject can read in approximately 2 sec."

Read that again. The claim is about how many words fit into two seconds of speech. It is a statement about articulation rate. The two seconds is the size of the rehearsal window, inferred from the fact that people remember roughly as much as they can say in that time.

From there the reasoning goes: if you can only keep alive what you can refresh, and you can refresh about two seconds of material, then whatever is not refreshed within about two seconds must be gone. So the store must hold something like two seconds' worth before it decays.

That is an inference, and a reasonable one. It is not an observation. A cross-linguistic review by Chincotta and Underwood states the standard model plainly: the store maintains information for roughly two seconds, after which it becomes irretrievable through trace decay. That is a fair summary of the textbook position. It is also the position that a substantial group of researchers now reject.

The honest version of the two-second claim is this. It is approximate. It is derived rather than measured. It varies with how fast you speak. And whether it reflects genuine time-based decay is one of the live arguments in the field.

The Four Effects That Built the Case

The loop was not proposed on the strength of one study. Four findings, taken together, make the argument, and each one closes a door that a simpler explanation might have escaped through. Most explainers list them as separate curiosities. They work much better read in order, as a single case being built.

One: similar sounds interfere with each other

In 1964 R. Conrad and A. J. Hull showed people strings of letters and asked them to write them back in order. Lists made of letters that sound alike, B, C, T, E and G, were recalled in the right order less accurately than lists of letters that do not, X, Z, A, R and W [4].

Two years later Baddeley extended it to words. Phonologically similar sets like cat, fad, pan and map were harder to hold in order than dissimilar sets like bar, kid, sun and toe. Crucially, when he compared this with semantic similarity, meaning made comparatively little difference to immediate serial recall [5].

That is the first door closed. Whatever this store is holding, it is holding it by sound, not by meaning. Confusions happen between items that sound alike.

And it happens with material you only ever saw, never heard, which is the evidence that written words are being recoded into sound on the way in.

Two: long words crowd out short ones

The word-length effect, described above. Short words are recalled better than long ones, and span tracks how fast you can articulate.

That is the second door. The limit is not a fixed number of slots. If it were a slot count, word length would not matter. Something time-shaped is going on.

Three: block the inner voice and both effects vanish

This is the most elegant of the four. Ask someone to repeat an irrelevant sound continuously, "the, the, the", while they try to hold a list. It is called articulatory suppression, and it occupies the rehearsal process.

Two things happen. The word-length effect disappears. And the phonological similarity effect disappears too, but only for material presented visually. For material people actually heard, similarity still bites.

That asymmetry is the third door, and it is a beautiful piece of reasoning. If suppression kills the similarity effect for written words but not spoken ones, then the route from written word to phonological code must run through the articulatory process. Block the voice and written material never reaches the store as sound. Spoken material gets in anyway, because it arrives as sound already.

Four: irrelevant sound gets in whether you like it or not

Colle and Welsh showed in 1976 that background speech damages recall of visually presented lists even when people are told to ignore it. Salamé and Baddeley followed it up in 1982 and interpreted it as unattended speech gaining automatic access to the store [6].

They also tested whether meaning mattered, by making the irrelevant speech semantically identical to the target items. The increase in disruption was very small and not statistically significant, which again points at a phonological rather than a semantic locus.

This fourth door is the one that did not stay shut. Dylan Jones and Bill Macken showed that irrelevant tones produce the effect too, and that what matters is change. A single repeated sound barely disrupts anything. A changing sequence disrupts a lot. Jones built an alternative account around that, in which the damage comes from changing acoustic state interfering with the process of holding order, rather than from sound forcing its way into a dedicated store.

So three of the four effects support the model cleanly. The fourth is genuinely contested, and the honest position is that the irrelevant sound effect is real while its explanation is not settled.

The Word-Length Effect Has a Replication Problem

Here is the part almost no page on the first page of Google will tell you.

The word-length effect is the load-bearing evidence for the two-second window. And it is fragile.

In 2000, Peter Lovatt, Steve Avons and Jackie Masterson looked hard at the original stimulus lists. Long and short words in the 1975 study differed in syllable count, but they also differed in other ways that affect memory: how common they are, how familiar, how many phonemes they contain, how much they sound like each other.

When Lovatt and colleagues built new word sets matched on those properties and compared short-duration disyllables with long-duration ones, the advantage did not reliably appear. The effect showed up with Baddeley's original words and largely vanished with matched ones [7].

Ian Neath, Tamra Bireta and Aimée Surprenant reached a compatible conclusion in 2003. The time-based version of the word-length effect, they argued, is specific to particular stimulus sets rather than a general property of memory [8]. Caplan and colleagues in 1992, and Service in 1998, had also failed to find length effects with properly matched materials.

Baddeley has defended the original interpretation, and the word-length effect is still routinely described as a benchmark finding that any model of verbal memory has to explain. It has not been withdrawn and it is not fraudulent. But its status as clean evidence for time-based decay is disputed in a way that revision guides never mention.

This matters more than a footnote, because if the word-length effect is stimulus-specific, the two-second window loses the study it was derived from.

Welsh, Mandarin, and Why Your Language Changes Your Span

If span really is limited by how much you can say in a couple of seconds, then people whose number words take longer to pronounce should remember fewer digits. Not because of anything about them. Purely because of their language.

That is testable, and it was tested.

In 1980 Nick Ellis and Richard Hennelly worked with Welsh-English bilinguals, the same people doing the same task in two languages. Digit span in Welsh was significantly smaller than in English, and the reason is unglamorous: Welsh digit words take longer to articulate [9].

The mirror image shows up in Mandarin. Chinese speakers have unusually long digit spans, and the standard explanation is that Chinese number words are short and fast to rehearse. Stigler, Lee and Stevenson documented the advantage in 1986.

There is a neat confirmation. Chincotta and Underwood showed in 1997 that cross-linguistic digit-span differences shrink under articulatory suppression. Take rehearsal away and the languages converge, which is what the loop predicts.

Two cautions, because this is where popular accounts overreach. First, the pattern generalises across many languages, with span tracking the articulation rate of number words across Arabic, English, Finnish, Greek, Hebrew, Japanese, Malay, Swedish and Spanish. Second, and less conveniently, at least one Mandarin-versus-English comparison found the advantage extends to word span and is not fully explained by rehearsal speed. Something beyond articulation rate is contributing.

Rehearsal speed is a large part of the story. It is not the whole story, and anyone telling you it is has stopped reading too early.

While we are on capacity, it is worth retiring another familiar number. The famous seven plus or minus two comes from George Miller in 1956. Nelson Cowan reviewed the evidence in 2001 and argued the real limit, once you strip out rehearsal and grouping, is closer to three to five items, about four [10]. Cowan also points out that Miller intended seven more as a rough estimate and a rhetorical device than as a measured constant. The way around that limit is chunking, which does not enlarge the store but packs more into each item.

The Loop as a Language-Learning Device

The most consequential claim about the phonological loop is not about phone numbers at all. It is that the loop is how you learn words.

The evidence runs through a task called nonword repetition. You hear something that obeys the sound rules of your language but means nothing, and you repeat it back. Because it has no meaning, you cannot lean on what you already know. You have to hold the sound.

Susan Gathercole and Baddeley followed children longitudinally and found that nonword repetition predicted later vocabulary size [11]. A companion study looked at children learning new names and found the same relationship.

In 1998 Baddeley, Gathercole and Costanza Papagno pulled this together into an explicit hypothesis: the phonological loop is a language learning device [12]. The association with vocabulary is strongest for nonword repetition, with correlations typically in the range of about .4 to .6, and weaker for digit span at roughly .25 to .45.

The logic is straightforward once stated. A new word is, at first hearing, a nonword. You have no meaning to attach it to. If you cannot hold its sound long enough to build a stable representation, you cannot learn it.

The same problem turns up in ordinary adult life. A person's name is phonologically arbitrary in exactly the way a nonword is, which is part of why we remember faces but forget names.

There is a complication worth stating, because it cuts against a purely mechanical reading. Nonword repetition is not a pure measure of storage. Children repeat nonwords that sound more like real words more accurately, which means long-term phonological knowledge is feeding into the task.

This connects directly to how bilingual brains store more than one language, and it is worth contrasting the kind of repetition described here with elaborative rehearsal, which works on meaning rather than sound and does something quite different to long-term retention.

The Patient Who Could Not Learn Russian

The strongest single piece of evidence for the language-learning claim comes from one person.

Giuseppe Vallar and Baddeley described a patient known as PV, an Italian woman with damage to the left hemisphere [13]. Her auditory verbal span was around two items. She showed no recency effect for spoken lists. Her visual short-term memory was intact.

Then came the finding that mattered. PV could learn pairs of words in her own language at a normal rate. She could not learn Russian vocabulary. The system that had failed was precisely the one needed to hold unfamiliar sound patterns long enough to attach meaning to them.

PV is often paired with another patient, ELD, who showed the opposite profile, a visuospatial short-term memory deficit with verbal short-term memory intact. Together they form a double dissociation, which is the neuropsychological argument that two systems are genuinely separate rather than two aspects of one.

One honest caveat, which most summaries skip. This is single-case evidence. PV is one person and ELD is one person. Single cases can be extraordinarily informative, and this one is, but a pattern in one patient is not a population estimate.

What Happens When the Loop Does Not Work Well

Beyond single cases, phonological short-term memory shows up across several clinical groups.

In developmental language disorder, also called specific language impairment, poor nonword repetition is one of the most reliable markers. A meta-analysis by Katharine Graf Estes, Julia Evans and Nicole Else-Quest pooled 23 studies and found children with SLI performed on average 1.27 standard deviations below children without it [14]. That is a large group difference, and it is why nonword repetition has been proposed as a phenotypic marker.

Laurence Archibald and Gathercole added an important qualification: the nonword repetition deficit in SLI is larger than, and persists beyond, a general short-term memory deficit. Something more specific is going on than a simple across-the-board memory problem.

The trail even reaches genetics. Work from the SLI Consortium linked variation in two genes on chromosome 16q, CMIP and ATP2C2, to nonword repetition performance in language impairment.

In dyslexia, reduced verbal short-term memory is a consistent associated finding, on digit span and nonword repetition, and it persists into adulthood. What it means is argued about. Steve Majerus and Cowan have made the case that the deficit is specifically about serial order rather than about phonological storage in general, which would reframe a lot of the literature. Imaging work by Steinbrink and colleagues in dyslexic thirteen-year-olds found reduced activation in left premotor regions associated with rehearsal.

A final contrast that illustrates the loop's partial independence from general ability. Phonological short-term memory is markedly impaired in Down syndrome and relatively preserved in Williams syndrome, despite significant cognitive difficulties in both.

A boundary worth drawing clearly. These are research findings about groups. Nothing here is a self-assessment. A short digit span is not a diagnosis, and the variation among people without any diagnosis is wide. Verbal span also changes across the lifespan, which we cover separately in how aging changes memory.

Where the Loop Lives in the Brain, and Whether It Lives Anywhere

For a long time this section of any article would have been short and confident.

In 1993 Eraldo Paulesu, Chris Frith and Richard Frackowiak published brain imaging work that appeared to localise the two components separately. The phonological store mapped to the left supramarginal gyrus in the inferior parietal lobe. The rehearsal process mapped to Broca's area and neighbouring premotor and supplementary motor regions [15].

Two boxes in the psychological model, two regions in the brain. It was a satisfying result and it was corroborated through the following decade.

Almost every explainer on the internet stops here. The field did not.

In 2008 Bradley Buchsbaum and Mark D'Esposito published a review whose title tells you where it lands: the search for the phonological store, from loop to convolution [16]. Their argument is that phonological short-term memory emerges from the combined action of the neural processes underlying speech perception and speech production, and does not correspond to a single brain region at all.

They are pointed about why the older result was so appealing. A dedicated store region gave neuroimaging a physical receptacle to match the psychological box. That is attractive, and it is not evidence.

Their alternative is a region called area Spt, sitting in the posterior planum temporale, which acts as an interface between hearing speech and producing it, within the dorsal auditory stream described by Gregory Hickok and David Poeppel [17].

Lesion evidence supports the reframing. In a study of fourteen patients with conduction aphasia, the region of maximal lesion overlap circumscribed area Spt, not the supramarginal gyrus that the classic account names [18].

Two broader challenges push in the same direction. Bradley Postle has argued that the standard model of dedicated buffers can no longer accommodate the findings, and that working memory is better understood as an emergent property that appears when attention is directed at representations held in the ordinary perceptual and language systems [19]. Daniel Acheson and Maryellen MacDonald have argued that the serial ordering the loop is supposed to provide is actually supplied by the machinery of language production, and that classic effects like phonological similarity are better read as production errors [20].

So the current state is genuinely open. There is a large body of imaging work consistent with separable store and rehearsal regions, and a serious body of work arguing that no dedicated store exists and that the phenomena emerge from speech systems doing what they normally do.

When Do Children Start Rehearsing?

The traditional answer has a specific age attached to it, which should always make you slightly suspicious.

John Flavell, David Beach and Jack Chinsky reported in 1966 that spontaneous verbal rehearsal appears around age seven and is close to universal by about ten. Supporting evidence came from the reported absence of word-length and visual phonological-similarity effects in younger children. If they are not rehearsing, those effects should not appear, and they did not.

The claim became a fixture. Age seven, rehearsal switches on.

Two lines of work have made it look much less like a switch. Christopher Jarrold and Deborah Hall argued that the apparent developmental change may be secondary to increases in how much children can recall at all, rather than a genuine qualitative shift in strategy.

Then in 2021 a seventeen-laboratory registered replication of the original Flavell study was published [21]. Registered replications are about as good as this evidence gets: the design and analysis are locked in before data collection. The broad age pattern held up. But a substantial proportion of five and six year olds were observed verbalising, and the smooth increase from seven to ten was not consistently present.

The defensible version is that rehearsal becomes more common and more organised across childhood, with wide variation between children at any age, and that a sharp age-seven onset is not supported.

Deaf Signers Have a Loop Too

If the loop were fundamentally about speech, a language with no sound should not engage it.

Margaret Wilson and Karen Emmorey tested deaf users of American Sign Language and found a phonological similarity effect for signs, with signs sharing formational features harder to hold in order [22]. Manual articulatory suppression, occupying the hands with irrelevant movement, eliminated the effect for pictures that had to be recoded, exactly as vocal suppression does for hearing participants.

They followed it with a sign-length effect, the manual counterpart of the word-length effect. Longer signs, worse recall.

The architecture, in other words, transfers. A store plus a rehearsal process that refreshes it, with the same signature effects, in a language made of hands rather than sound.

The most recent evidence pushes further. A 2026 study of deaf native signers of Israeli Sign Language concludes that phonological working memory is largely modality-free, with similar constraints operating in signers and speakers [23].

This is why the name matters. Phonological here does not mean acoustic. It refers to the structured sublexical units a language is built from, which in a signed language are handshapes, locations and movements. The loop is a language system, not a hearing system.

Is There Any Decay At All?

We come back to the two seconds, and to the largest disagreement in this field.

The classical account needs decay. A trace fades with time, rehearsal refreshes it, and forgetting is what happens when refreshing fails to keep up.

Stephan Lewandowsky and Klaus Oberauer have spent years arguing that time-based decay does not exist. Their 2008 paper takes direct aim at the foundation, arguing the word-length effect provides no evidence for decay [24]. In 2015 they went further, arguing that rehearsal is an unworkable solution to a problem that is not there [25].

Their alternative is interference. Items degrade because other items overwrite and confuse them, not because a clock runs down. Their computational model fits serial recall data without any decay parameter at all, which is a strong argument: if you can reproduce the phenomena without the mechanism, the mechanism is not doing necessary work.

On the other side, Pierre Barrouillet and Valérie Camos have defended temporal decay within their time-based resource-sharing framework, in which maintenance happens through attentional refreshing that competes for the same resource as ongoing processing.

Several other models sit nearby. Nairne's feature model explains item and order memory through feature overlap and interference. Jones's account treats serial memory as a by-product of general perceptual and motor processes. Cowan's embedded-processes model dispenses with dedicated buffers entirely, treating working memory as activated long-term memory plus a limited focus of attention. Ericsson and Kintsch described how skilled performers exploit long-term structures to achieve effective capacity far beyond any short-term limit.

Set that beside the sections above and the pattern is uncomfortable for the textbook story. The two-second window rests on the word-length effect. The word-length effect is stimulus-set specific. And a serious research programme argues the decay it was taken to demonstrate is not there.

None of this means the phonological loop is fiction. The phenomena are solid and reproducible. What is contested is the mechanism underneath them.

What Is Settled and What Is Not

It is worth separating these cleanly, because most sources blur them.

Reasonably settled:

The phonological similarity effect is real and well replicated, and it appears with visually presented material, which implies recoding into a sound-based form. Articulatory suppression reliably removes the word-length effect and removes the similarity effect for written but not spoken material. Immediate serial recall of verbal material is coded largely by sound rather than meaning. Memory span tracks articulation rate across languages, and the direction of that relationship is not in dispute. Nonword repetition correlates with vocabulary growth more strongly than digit span does. Poor nonword repetition is a reliable group-level marker of developmental language disorder. Selective verbal short-term memory impairment exists as a clinical dissociation. Deaf signers show similarity and length effects in signed language. And the multicomponent framework has survived fifty years even as the models inside it have been revised.

Genuinely unsettled:

Whether the word-length effect holds with properly matched stimuli, with Baddeley on one side and Lovatt, Neath and colleagues on the other. Whether there is time-based decay at all, with Barrouillet and Camos against Lewandowsky and Oberauer. Whether a dedicated phonological store exists as a brain region, with Paulesu and the classical imaging literature against Buchsbaum, D'Esposito, Postle, Acheson and MacDonald. What causes the irrelevant sound effect, store access or changing state. When rehearsal emerges in children. And whether the Mandarin span advantage is fully explained by rehearsal speed.

Six substantial open questions in a topic that most pages present as finished.

Things You Will Read That Are Not Right

A short list, because correcting these is more useful than repeating them.

The store does not hold sound for exactly two seconds. The figure is approximate, derived from articulation rate rather than measured directly, and the decay it implies is disputed.

The loop does not store general auditory information. It handles speech-based and phonological material. Pure tones and music behave differently, and signed language engages it, so the right word is language rather than sound.

The store does not hold a fixed number of items. It is time-shaped rather than slot-shaped, which is exactly why span differs between languages.

The phonological loop and the phonological store are not the same thing. The store is one of two components. Leaving out the rehearsal process removes half the system and all of the explanatory power.

Short words are not always remembered better than long ones. Match the words properly on frequency, familiarity, phoneme count and similarity, and the advantage largely goes away.

The store is not simply located in the supramarginal gyrus. That was the 1993 result. Lesion evidence from conduction aphasia centres on area Spt instead, and several researchers argue no dedicated store region exists.

Children do not start rehearsing at seven. A seventeen-lab replication found many younger children verbalising and no clean switch.

Working memory capacity is not seven plus or minus two. Cowan puts the underlying limit nearer four.

And one that should be said plainly: you cannot train your phonological loop to be bigger. Strategies like chunking and grouping work around the limit by reorganising what counts as an item. They do not raise the ceiling. Claims that a few weeks of exercises will expand your working memory run well ahead of the evidence.

What the Loop Actually Is

Strip away the disputed mechanism and something solid remains.

There is a system that holds language-shaped material for a short time. It is limited by how fast you can articulate, which is why the limit moves when you change language and why it collapses when you occupy your voice. It confuses things that sound alike. It is disrupted by changing background sound. It recodes what you read into what you would say. It predicts how quickly children acquire vocabulary, and when it breaks, learning unfamiliar words breaks with it.

Whether that system is a dedicated store with its own address in the parietal lobe, or a pattern that emerges when speech perception and speech production hold hands for a couple of seconds, is not resolved. Fifty years in, the framework has outlived most of its own specifics, which is roughly what Baddeley said would happen.

The next time a number arrives and you find yourself repeating it under your breath, you are running the one part nobody disputes. The inner voice, refreshing a fading trace, buying you a few more seconds. Whether the trace was fading because time was passing or because other things were crowding in is, remarkably, still an open question.