Tip of the Tongue Phenomenon

Introduction

You know the word. You know what it means. You could pick it out of a lineup in half a second. You can even tell yourself it starts with an S, that it has two syllables, that it sounds a bit like "secant." And yet the word will not come.

This is the tip of the tongue phenomenon, and it is one of the strangest things memory does. Most memory failures are quiet. You either recall something or you do not, and the gap leaves no trace. But this failure announces itself. It comes with a feeling of imminence, a sense that the answer is millimetres away, and often with a small burst of frustration that psychologists Roger Brown and David McNeill described in 1966 as looking like mild torment, something close to the brink of a sneeze, followed by considerable relief when the word finally arrives [1].

Here is the core of it, stated plainly. A tip of the tongue state is not a storage failure. The word is in memory. What breaks is the link between what you mean and how the word sounds. Meaning arrives. Sound does not. That single dissociation is why researchers have spent six decades studying a moment that lasts, in most cases, only a few seconds.

And the research has gone somewhere genuinely surprising. Some of what you feel during these moments turns out to be less reliable than it seems. The oldest explanation for why they happen was widely considered dead by 2006, then came back to life in 2025 on the strength of brain imaging data from 576 people. Most interesting of all, and almost never written about outside the journals, is what these states do to your memory afterwards. Struggling with a word and losing does not leave you where you started. It leaves you slightly worse off. Resolving the struggle repairs the damage.

That last finding has practical consequences for anyone who studies. We will get there.

The Harvard experiment that started everything

Before 1966, the tip of the tongue state was an anecdote. William James had written about it in 1890, describing the gap left by a forgotten name as intensely active, a hollow shaped like the missing word. Sigmund Freud had his own theory involving repression. Neither could measure anything.

Brown and McNeill changed that with a method so simple it still gets used. They read participants the dictionary definitions of rare words and asked them to supply the word. Not common words. Words chosen from a narrow frequency band, common enough to be known but rare enough to be shaky: apse, nepotism, cloaca, ambergris, sampan, caduceus, sextant.

If this one causes a problem for you, apologies in advance. "A navigational instrument used in measuring angular distances, especially the altitude of the sun, moon and stars at sea."

The word is sextant.

Fifty six Harvard and Radcliffe undergraduates sat through three evening sessions. Forty nine target words. When a participant felt the word was close but would not come, they signalled it, and the experimenters immediately started asking questions. How many syllables? What letter does it start with? What other words are coming to mind instead?

Across the study, 360 tip of the tongue states were signalled, of which 233 were scorable [1]. Nine participants had none at all. One had eight.

What came back from those interrogations is the finding the whole field is built on. People stuck in the state were not blank. They were holding fragments. Brown and McNeill called this generic recall, and the numbers were well above what guessing would produce. First letter, correct 57 percent of the time. Syllable count, so accurate that the rank order correlation between what people guessed and what the words actually were came out at 1.0.

They also looked at what wrong words people produced. Two kinds showed up. Words similar in meaning, and words similar in sound. The sextant definition pulled up astrolabe, compass and protractor from the meaning side, and secant, sextet and sexton from the sound side. That split matters, because it suggests the search process can get hold of a word's shape without getting hold of its identity.

One detail deserves to be carried forward honestly, because it usually is not. Brown and McNeill also reported that people could locate a word's primary stress above chance, but they flagged a statistical problem with their own analysis and wrote that they were not sure of it. The syllabic stress result was never as solid as the first letter result, and the original authors said so.

Alan Brown's 1991 review in Psychological Bulletin pulled twenty five years of follow up work together and set the modern agenda [2]. It also sharpened a methodological split that still shapes the literature.

There are two ways to study these states, and they disagree.

The first is the diary study. Give people a notebook, ask them to record every tip of the tongue moment for four weeks, and count. Burke, MacKay, Worthley and Wade did exactly this across young, middle aged and older adults [4]. Diaries capture the real thing in the real world, which is their whole appeal. But the bias runs in an obvious direction. Older participants worried about their memory are more motivated to write things down. A moment that resolves is more satisfying to record than one that does not. Frequency counts and resolution rates from diaries should be read as characterisations, not measurements.

The second is laboratory induction, which controls everything and loses the natural context. Heine, Ober and Shenaut ran both approaches across three age groups in the same study to see how far apart they land [5]. Researchers even invented artificial creatures for this purpose. The TOTimals method taught participants names and attributes of imaginary animals, then tested retrieval, so that everyone had identical prior exposure to every target [6].

The commonly repeated figure that young adults get one of these a week and older adults get one a day comes from the diary tradition. It is a reasonable generalisation across that literature. It is not a number anyone measured in a lab.

What actually comes out when the word will not

The partial information finding is the most cited result in this field. It is also the one most in need of an update.

Later work extended the list of retrievable fragments well beyond first letters. Miozzo and Caramazza tested Italian speakers, where nouns carry grammatical gender, and found that people in a tip of the tongue state could report a word's gender at above chance rates while having no access at all to its sound [8]. Gender and phonology came apart. Whatever the retrieval system is doing, it is not a single lookup that either succeeds or fails. It has layers, and the layers can fail separately.

Koriat and Lieblich framed the question that organises all of this back in 1974, asking what a person in this state knows that a person who simply does not know the answer lacks [7].

Here is the update, and it complicates the classic story.

Huebert, McNeely-White and Cleary published a set of experiments in 2023 that separated what people report freely from what they produce under pressure [9]. When free to describe their experience, people in the state reported more partial recollection than people not in it. That matches Brown and McNeill. But they were not more accurate. And when forced to guess the first letter, participants showed a strong tendency to feel confident they knew it while getting it wrong.

Read that again, because it is the part that changes things. Part of the felt experience of almost having the word appears to be an inference the mind constructs, not a readout of something genuinely retrieved. Cleary's laboratory connects this to work on déjà vu, where people similarly confabulate reasons for a familiarity signal they cannot explain.

So both things are true. Real fragments do come out, reliably, and Brown and McNeill measured them correctly. And the subjective certainty attached to those fragments runs ahead of their accuracy. The feeling is not a clean instrument.

It is also worth separating this state from a close relative. A tip of the tongue state is the acute sense that a word is arriving right now. A feeling of knowing is a colder prediction that you would recognise the answer if someone offered it. Schwartz and Metcalfe argue these are metacognitive experiences with different properties rather than the same thing at different intensities [10]. Brain imaging has since supported the separation.

A sixty year argument, in order

The shape of this field is easier to see chronologically. Ideas that looked settled were overturned, and at least one that looked dead has come back.

Notice the gap between 1991 and 2015. The mechanism question dominated for decades while the question of what these states do to the learner sat almost untouched.

Five explanations, and an argument nobody has won

Schwartz's 1999 paper gave the field its organising distinction [3]. Either you have some direct access to the target sitting in memory, or you are inferring its nearness from clues. Everything since sorts into those two camps.

**Blocking.** A related word arrives first and gets in the way. Jones and Langford tested this by feeding participants phonologically similar words and reported that these increased the rate of failure [11]. The intuition is strong. Everyone has had a wrong name jam itself into the slot where the right one should go.

**Incomplete activation.** The target is there but has not crossed the threshold needed to produce speech. You sense its presence because activation is weak rather than absent.

**Transmission deficit.** Burke, MacKay, Worthley and Wade proposed that meaning and sound are stored as separate layers of nodes, and that the state occurs when activation reaching the meaning layer fails to pass down to the sound layer [4]. Three things weaken that connection: not using a word often, not using it recently, and getting older. This became the field's default account because it predicts so much of the data at once.

**Cue familiarity.** Metcalfe, Schwartz and Joaquim argued that the feeling can be triggered by how familiar the question feels, independently of whether the answer is accessible [13]. Familiar cue, strong feeling, no target required.

**Accessibility heuristic.** The volume and vividness of whatever comes back drives the sensation, whether or not any of it is correct. Cleary's illusory partial access results fit here comfortably.

Now the argument.

Blocking took a serious hit in 2007. Kornell and Metcalfe compared states where an intruding word showed up against states where nothing came at all [14]. If blockers block, the two kinds should behave differently, and a delay should help the blocked ones more by letting the intruder fade. Neither happened. The two types looked the same, and delay helped both equally. Harley and Bown had already pushed from the other direction, arguing that words from dense phonological neighbourhoods are actually easier to retrieve, and that it is words with few sound neighbours that get stuck [12].

For nearly twenty years, that looked like the end of blocking.

Then Gong, He, Wang and Wang published a diffusion tensor imaging study in 2025 using data from 576 people aged 18 to 87 [15]. They mapped the structural connectivity of the phonological processing network and asked which properties predicted how often someone got stuck. The answer went the wrong way for transmission deficit. Higher global efficiency and higher mean degree centrality in the phonological network predicted more failures, not fewer. At the level of individual regions, nodal efficiency in the bilateral posterior superior temporal gyrus and clustering in the left premotor cortex predicted more of them, while degree centrality in the left dorsal superior temporal gyrus and clustering in the left posterior supramarginal gyrus predicted fewer.

The authors read a richer phonological network producing more failures as evidence for competition, which is to say for blocking.

So where does that leave things? Honestly, unresolved. Transmission deficit still explains the widest range of behavioural findings, particularly ageing and word frequency effects. Blocking has a fresh structural result behind it that transmission deficit does not obviously predict. Behavioural evidence and structural evidence are currently pointing in different directions, and a single imaging study does not settle a dispute that behavioural work spent two decades on. Anyone who tells you this question is closed is ahead of the data.

The pipeline these theories argue about looks roughly like this.

That final branch, the one on the right, is where the learning research lives. It gets its own section later because almost nobody covers it.

What brain imaging can and cannot tell you

A wave of imaging work between 2001 and 2010 tried to find the state in the brain. It found activity. Whether it found the state is a harder question.

Kikyo, Ohki and Sekihara used functional MRI to track the time course of retrieval and reported activity in the anterior cingulate cortex and dorsolateral prefrontal cortex during failed access [16]. Maril, Simons, Weaver and Schacter then compared the tip of the tongue state directly against feeling of knowing and found they produce different signatures, supporting the argument that these are separate phenomena rather than one thing at two intensities [17].

The most informative result came from Cambridge. Shafto, Burke, Stamatakis, Tam and Tyler found that the rate at which older adults got stuck on words tracked grey matter loss in the left insula, a region tied to phonological production [18]. What makes this finding useful is the control comparison. Performance on Raven's Matrices also declined with age, but did not correlate with insula grey matter. Word finding was dissociating from general cognitive decline. A follow up showed that during successful naming, insula activity was unaffected by age. The age difference appeared only during failures, where younger adults produced an activity boost that older adults did not [19].

Electrophysiology added its own picture. Galdo-Alvarez, Lindín and Díaz reported prefrontal over recruitment in older adults during face naming, consistent with a search that keeps running after it should have stopped [20]. Kozlovskiy and colleagues used source localisation and reported reduced parahippocampal activation during these states [21].

The most recent synthesis is Xie and Wang's 2026 review, which pulls the cognitive, neural and neurochemical strands together [22]. It describes a left lateralised frontotemporal network handling retrieval monitoring, links higher failure rates to reduced integrity of the arcuate and uncinate fasciculi, and proposes that transient lexical blocks reflect a local imbalance between excitation and inhibition in language circuits.

That proposal is worth stating carefully. The GABA and glutamate evidence in the review comes from spectroscopy studies of related language tasks, not from spectroscopy performed during tip of the tongue states themselves. It is an extrapolation, and a reasonable one, but nobody has yet measured the neurochemistry of the moment itself.

Two further limits belong on the record. Sample sizes in the imaging literature are small, often a couple of dozen people. And the reverse inference problem is severe here. The anterior cingulate cortex activates during conflict, error detection, pain, effort and emotional arousal. Observing it light up during a failed word search does not establish that it is doing conflict detection in that moment. The conflict interpretation is plausible. It is not proven by the activation.

Pharmacology has produced two results that are more informative than they look. Bacon, Schwartz, Paire-Ficout and Izaute gave participants lorazepam and found something odd: the drug did not stop the underlying retrieval process, but it stripped away the subjective experience [23]. People produced wrong answers without the feeling that would normally warn them. The cognitive process and the phenomenology came apart. And Lesk and Womble gave people 200 milligrams of caffeine, roughly two cups of coffee, before a general knowledge test [24]. Caffeine reduced failures when the priming words were phonologically related to the answer, and increased them when the priming words were unrelated. Alertness alone cannot explain a result that goes in both directions.

Getting older, and the line that actually matters

These states become more frequent with age. That part is not in dispute and it appears across both diary and laboratory methods.

Why proper names are the worst category has a mechanistic answer. Most words have redundant routes into them. If you cannot retrieve "hammer," the concept of a tool for driving nails still connects to the word through several semantic paths. Proper names have no such backup. Knowing that a man is a baker does not help you retrieve the surname Baker, because the link between the person and the sound of their name is arbitrary and singular. One weak connection, no alternative route.

Barry, Ferrer, Lerma-Usabiaga and Paz-Alonso tested this systematically in 2025 with 80 adults naming famous faces and places [26]. Three factors each independently predicted more failures: learning the name later in life, retrieving it less often, and encountering it less recently. Navarrete and colleagues had earlier shown that words learned early in life resist these failures, an age of acquisition effect that persists across the lifespan [27].

Now the question people actually want answered. Is this a warning sign?

Salthouse and Mandell put 718 adults aged 18 to 99 through naming tasks covering famous places, common nouns and famous people [25]. Failures increased with age across all three categories, as expected. Then they controlled statistically for episodic memory, the system that deteriorates in Alzheimer's disease. If word finding difficulty were an early expression of that same decline, controlling for it should have shrunk the age effect substantially.

It barely moved. Their conclusion was that although ageing is associated with both lower episodic memory and more frequent word finding failures, the two phenomena appear to be largely independent of one another.

So the frequency of these moments, on its own, is a poor signal.

What carries more information is resolution. Kim, Kim and Yoon compared 30 older adults with subjective memory complaints against 30 without, all aged 50 to 79, using a celebrity naming task [29]. Despite scoring normally on objective neuropsychological testing, the complaint group not only got stuck more often but resolved fewer of those states, including after being given syllable cues. Campos-Magdaleno and colleagues followed a longitudinal cohort and found that these measures help distinguish cognitively unimpaired adults from those with mild cognitive impairment [30].

Genuine word finding pathology also looks different in kind, not just in degree. In anomia following stroke or brain injury, the failure is pervasive and covers everyday vocabulary, and the person may not recognise the word even when they hear it. In the semantic variant of primary progressive aphasia, the meaning itself erodes. That never happens in a benign tip of the tongue state, where the concept stays perfectly intact and the person knows exactly what is missing. Beeson, Holland and Murray used famous name retrieval to compare these clinical groups directly [28]. Elevated rates also show up in dyslexia, where the difficulty sits specifically at the phonological step while meaning remains accessible [31].

The practical distinction, then, is not how often it happens. It is whether the words keep coming back, whether the concept behind them stays intact, and whether anything else is changing alongside. A gradual increase in frequency with age is the expected pattern. A pattern of states that stop resolving, appearing together with other cognitive or language changes, is worth raising with a clinician. Anyone genuinely worried about their own memory should talk to a doctor rather than count incidents against a chart.

The same experience in hands, noses and fifty one languages

If this is a basic property of retrieval rather than a quirk of English speakers, it should show up everywhere. Mostly, it does.

Schwartz surveyed idioms across 51 languages and found that 45 of them, roughly 88 percent, have an expression referring to the tongue, mouth or throat [3]. The Korean expression translates as sparkling at the end of the tongue, which gave his paper its title. Five languages in the survey had no such idiom: American Sign Language, Amharic, Icelandic, Kalenjin and Kiswahili.

This figure gets repeated online as evidence that 90 percent of people experience the phenomenon. It is not that. It is a count of languages that happen to have a particular idiom, which is a fact about vocabulary rather than a measurement of how many humans have the experience.

The distinction matters, and Brennen, Vikan and Dybdahl demonstrated why. They worked with speakers of an unwritten Guatemalan Mayan language who had no dedicated verbal expression for the state, and induced it at rates comparable to those found among speakers of Western languages, resolved reliably by giving initial letters [36]. No idiom, same experience.

Bilingual speakers produce a result with a clean internal contrast. Gollan, Bonanni and Montoya found that bilinguals get stuck more often than monolinguals on ordinary vocabulary, but at the same rate on proper names [32]. The standard explanation is a frequency lag. Splitting your speech across two languages means every individual word gets used less often, which weakens the connections that transmission deficit says matter. Proper names are already low frequency for everybody, so there is no additional penalty to add. Gollan and Silverberg found the same elevated rate in Hebrew and English bilinguals with comparable partial information access [33], and Ivanova and Costa documented slower picture naming in bilinguals even in their dominant language [34]. If you want the wider picture of how two languages coexist in one brain, our piece on how bilingual brains store languages covers the storage side.

Then a study that rules out a competing explanation entirely. Pyers, Gollan and Emmorey tested people fluent in both American Sign Language and English [35]. These two languages share no phonology whatsoever, one being manual and the other spoken. Blocking between similar sounding words across languages is therefore impossible. The bilingual disadvantage appeared anyway, which points at the frequency account rather than cross language interference.

Signers get their own version. Thompson, Emmorey and Gollan documented tip of the fingers experiences in 33 deaf signers [37]. The structure is the same. The states resolve spontaneously, they cluster on proper names, and they leak partial information. Signers stuck on a sign could more readily retrieve its handshape, location and orientation than its movement. A visual and manual language with no sound in it at all still splits meaning from form, and still fails at the same seam.

Odours behave differently, and the difference is informative. People can feel that the name of a smell is imminent without being able to produce it, but these tip of the nose states do not show the same partial phonological access. That asymmetry suggests the fragment leakage is specific to lexical retrieval rather than a general property of feeling close to an answer. Cleary and Schwartz's 2025 volume maps this family of related phenomena, including the connection to déjà vu [38].

The part almost nobody writes about

Everything so far treats these states as a symptom. Something breaks, you feel it, it passes. But there is a body of research asking a different question. What does the state do to your memory afterwards?

The answer is not neutral.

Warriner and Humphreys ran an experiment where each failed retrieval was assigned either a short struggle period of 10 seconds or a long one of 30 seconds before the answer was supplied [39]. Participants came back 48 hours later and were tested on the same words. Words that had been struggled with for the longer period were almost twice as likely to get stuck again.

Sit with that. Being told the correct answer did not prevent the repeat. And more effort spent failing made the repeat more likely, not less.

D'Angelo and Humphreys followed this with six experiments in Cognition and worked out what was happening [40]. Roughly 21 percent of states from a first test recurred, both at a five minute interval and a week later. The interpretation is implicit learning of an unsuccessful route. Every second spent searching and failing is a second spent practising a path that does not lead to the word.

But the same paper carries the fix, and it is a strong one.

When a state was resolved, the recurrence rate dropped sharply. The conditional probability of getting stuck again on a word ran around .26 to .30 for unresolved states against roughly .08 to .13 for resolved ones. In the 48 hour comparison, resolving made recurrence about five times less likely. Cued resolution worked as well as spontaneous resolution, and giving cues raised the resolution rate from 35 percent to 82 percent. The learning was implicit rather than strategic: telling participants a later test was coming made no difference, with recurrence at 12 percent for the informed group against 11 percent for the uninformed.

There is an upside too. Metcalfe, Schwartz and Bloom gave 46 undergraduates 82 general knowledge questions under a five second deadline, then told them they could look up only about 10 percent of the answers, forcing them to choose [41]. Being in a tip of the tongue state raised the probability of wanting the answer from 0.18 to 0.44. Even when participants had answered correctly, the state raised answer seeking from 0.03 to 0.24. A separate survey found that people are essentially unaware this is happening to them.

So the state is a curiosity signal that operates below conscious awareness. Bloom and colleagues followed the physiology and found that feedback delivered during one of these states produced enhanced centro parietal positivity between 250 and 700 milliseconds afterwards, and that this enhancement mediated a positive relationship between the state and later recall [42]. The moment the missing word arrives is a moment of unusually good encoding.

This sits inside a wider literature on learning from mistakes. Metcalfe's review of that field documents the hypercorrection effect, where errors made with high confidence are the ones most likely to be corrected afterwards [43]. Failing well is a real thing. Failing and never finding out is not.

If retrieval attempts and their aftermath interest you, our article on retrieval practice and the brain goes deeper into why testing yourself changes memory more than reviewing does.

So how long should you struggle?

This is the obvious question, and honesty requires a clear answer about its status.

No study has directly measured an optimal waiting time. Nobody has run the experiment that varies delay across a proper range and reports a turning point. What follows is an inference assembled from three separate findings, and it should be treated as an inference rather than a result.

The first finding is that delay genuinely helps. Choi and Smith found higher resolution rates after an interval than on an immediate retest, with the effect stronger for the states that felt most intense [44]. Stepping away works. Kornell and Metcalfe found the same benefit of delay, though notably it helped all states equally rather than helping supposedly blocked ones more [14].

The second is that effortful, prolonged struggle is specifically what entrenches the error. In D'Angelo and Humphreys, the difference between short and long delays only appeared under instructions that pushed intense retrieval effort, where long delay recurrence hit 18 percent against 6 percent for short [40]. Grinding is the problem, not waiting.

The third is that resolution repairs the damage, and a cue counts as resolution.

Put those together and a defensible position emerges. Sitting with the gap while your attention drifts elsewhere is fine and may let the word surface on its own, which is the outcome that most strengthens the correct route. Actively hammering at it for an extended period is the one behaviour with direct evidence of a cost. If the word has not arrived after a genuine but brief attempt, getting it, whether through a partial cue or by looking it up, and then deliberately using it again, converts a trial that would have taught you the wrong path into one that repairs it.

Phonological cues in particular have solid support. Rastle and Burke showed that prior processing of related words primes resolution in both young and older adults [45]. Abrams and Rodriguez found that the syntactic class of a priming word affects whether it helps, and that position within a list of primes matters [46]. Ouyang, Cai and Zhang reported that older adults benefit more from phonological priming than younger adults, which is what a transmission deficit account predicts [47].

Gesture, often suggested as an aid, has weaker support. Beattie and Coughlan investigated whether iconic gestures assist lexical access during these states and the results were equivocal [48].

Things you have probably read that are not true

Three claims circulate widely and none of them survives contact with the source.

The first is that 90 percent of people experience this phenomenon. The real number behind it is Schwartz's finding that 45 of 51 surveyed languages, about 88 percent, contain a relevant idiom [3]. Somewhere in the retelling, a fact about languages turned into a fact about people. The evidence that the experience itself is close to universal is real, but it comes from induction studies like Brennen's work with speakers of an unwritten language [36], not from that percentage.

The second is the advice to clench your right fist to retrieve a stuck word. This traces to a 2013 PLOS ONE paper by Propper and colleagues on unilateral hand clenching and episodic recall [49]. Three problems. The final sample was 50 people. The key corrected score analysis did not reach significance, with F(4,45) = 2.20 and p = .08. And it was not a study of tip of the tongue states at all. Whatever it does or does not show about episodic recall, it is not a basis for advice about stuck words.

The third is that frequent word finding failures are an early sign of dementia. Salthouse and Mandell's data argue directly against reading it that way [25]. The increase with age is real and it is largely independent of the memory decline that matters clinically.

Conclusion

The tip of the tongue phenomenon looks like a small malfunction and turns out to be a window. It shows that producing a word is not one operation but at least two, that meaning and sound are stored separately enough to fail separately, and that the system running the search can report on its own progress while getting the details wrong.

Sixty years in, the honest summary is mixed. The partial information finding is solid and replicated, though the confidence attached to those fragments outruns their accuracy. The mechanism question is open, with a two decade behavioural consensus now facing structural imaging data that points elsewhere. The neuroscience has identified a network without establishing what any part of it is doing, and the newest neurochemical account is a well argued extrapolation rather than a measurement. The clinical picture is clearer than the popular coverage suggests, and considerably less alarming.

What has changed most is the framing. These moments were treated as errors for most of the field's history. The work since 2008 suggests they are more like unstable states that resolve in one of two directions. Resolve one and you have strengthened the right connection. Fail to resolve it and you have quietly rehearsed the wrong one.

The word on the tip of your tongue is not lost. It is a link that did not carry enough signal this time. And what you do in the next few seconds decides whether that link gets stronger or weaker.