Working memory: why you can hold only about four things

Miller's famous seven was never the real limit. Strip away rehearsal and grouping and the mind holds three to five chunks, which is why chunking, not willpower, decides how much you can juggle.

A Rigveda manuscript leaf in Sanskrit written in Sharada script, from Kashmir
Photo: Ms Sarah Welch, CC0, via Wikimedia Commons

Working memory capacity in young adults is about four chunks, usually 3 to 5, once you stop people from rehearsing or grouping the items. George Miller's "seven plus or minus two" from 1956 measured something looser, a span inflated by silent repetition and chunking. The practical upshot: you cannot raise the four, but you can make each of the four slots carry far more by packing information into larger chunks.

Key takeaways
  • About 4 chunks (3 to 5) is the central limit in young adults once rehearsal and grouping are blocked (Cowan).
  • Miller's 7 ± 2 (1956) measured a compound span inflated by silent repetition and chunking.
  • About 1.5 vs 3.0 items: 7-year-olds hold half what adults do in the same array task.
  • 4 parts in Baddeley's model: phonological loop, visuospatial sketchpad, central executive, episodic buffer.
  • From 7 to about 80 digits: one student's span after 250+ hours of practice, by chunking, not by a bigger store.

Working memory is the small workspace where you hold a phone number while you dial, the first half of a sentence while you read the second, or the carry digit in a sum. Its limit shapes almost every mental task, from following a lecture to comparing three bikes on four specs. This post covers how the limit was measured, why the number fell from seven to four, how Alan Baddeley's model splits the workspace into parts, and how chunking, the trick behind Vedic recitation and the multiplication tables Indian schoolchildren chant, gets around it.

3–5chunks held in young adults when rehearsal and grouping are blocked (Cowan 2010)
~2 sof speech that silent rehearsal can keep alive
~1.5 vs ~3items held by 7-year-olds vs adults in the same array task
1974year Baddeley and Hitch proposed the multicomponent model

Is the magical number seven still accepted?

Not as a capacity limit. Miller's 1956 paper showed that people can repeat back about seven random digits, letters or words, and he used "seven plus or minus two" partly as a joke about an integer that seemed to follow him around. Later work showed the seven depends on the material: young adults recall only 3 or 4 longer verbal units such as idioms or short sentences.

Nelson Cowan's 2001 review in Behavioral and Brain Sciences argued that the real, central limit averages about four chunks, and that seven appears only when people can quietly rehearse items or merge them into bigger units. His point was not that Miller was careless but that a digit span test measures a compound of several mechanisms:

  • a short speech-based store;
  • rehearsal, repeating items silently;
  • grouping items into bigger units;
  • long-term knowledge that turns a group into one familiar thing.

Take those away and a stable number appears. Cowan's later summary in Current Directions in Psychological Science (2010) also notes that mathematical models of reasoning tend to fit best when the number of items in working memory is set at about four.

The story

At Carnegie-Mellon University in the late 1970s, William Chase and K. Anders Ericsson paid an undergraduate, known in their papers as SF, by the hour to do one dull thing: listen to random digits read at one per second, then repeat them in order. He started at the normal span of about 7 digits. He was a good long-distance runner, and on day 5 he began hearing digit groups as race times; by day 20 he coded four-digit groups, so 3492 became a near-world-record mile, 3 minutes 49.2 seconds. Groups that could not be times became ages or years. After about 25 months and over 250 hours, his span stood at about 80 digits. Yet when the researchers switched to random consonants, his span was around 6. He had not enlarged his store; he had learned to pack three or four digits into each chunk, and to hang the chunks on a structure in long-term memory.

Chase W.G., Ericsson K.A.: Skilled Memory, Carnegie-Mellon University Technical Report No. 5 (1980)
How many items working memory holds, by method
See the data as a table
Typical items held (items)
Miller 1956 digit span7
Cowan central limit (young adults)4
Learned pairs, rehearsal blocked3
Array task, adults3
Array task, 7-year-olds1.5

Source: Cowan, Current Directions in Psychological Science 2010; Cowan, Behavioral and Brain Sciences 2001 · Central limit shown as the midpoint of 3 to 5 chunks.

The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage CapacityNelson Cowan · Behavioral and Brain Sciences 24(1) · 2001

A target article reviewing decades of short-term memory studies. Cowan argues Miller's seven was a rough estimate and that the "pure" capacity limit is about four chunks, visible under four conditions: information overload that limits chunks to single items, steps that block recoding into bigger chunks, sharp discontinuities in performance at the limit, and indirect effects of the limit on other tasks. He separates this pure limit, measured in chunks, from "compound" limits where the number of chunks is unknown. The limit: the paper is a synthesis with open peer commentary, and critics argued that a continuous resource, not a fixed slot count, fits some data as well.

Read the paper (free abstract)

How do you measure the real limit?

You measure it by removing the tricks people use to cheat it. Cowan lists four boundary conditions under which about four items show up:

  1. a brief array of items shown all at once;
  2. sounds presented while attention is elsewhere, with attention turned to them only after they end;
  3. lists read while the participant keeps saying one word aloud, which blocks silent rehearsal;
  4. "running span", where a list stops without warning and you must recall the last few items.

In running span people recall only the last 3 to 5 digits, less than two seconds of speech, even though silent rehearsal could in principle hold about two seconds' worth.

Two experiments make the case sharp. In one, Cowan's group taught people new word pairs until each pair acted as a single chunk, then mixed pairs and single words in lists while participants repeated "the" to stop rehearsal.

People recalled about three units whether those units were single words or learned pairs, so capacity was counted in chunks, not words. In another, people saw coloured squares while hearing spoken digits from four loudspeakers. With a mask that wiped sensory memory, they held about four items in total, split between squares and digits, rather than four of each. The central store was shared across senses.

The limit also develops with age. Using a child-friendly array task, Cowan and colleagues estimated about 1.5 items stored for 7-year-olds against about 3.0 for older children and adults, and the advantage of attending to the right items was just as large in the youngest children. Storage capacity itself, not only attention control, separates a 7-year-old from an adult.

Telephone operators at switchboards in the Chinese Telephone Exchange in San Francisco in 1918
Operators at the Chinese Telephone Exchange, San Francisco, 1918, connected calls by name and number from memory. Long numbers stay usable only because we split them into groups of three or four digits.Photo: AnonymousUnknown author, Public domain, via Wikimedia Commons
The Magical Mystery Four: How Is Working Memory Capacity Limited, and Why?Nelson Cowan · Current Directions in Psychological Science 19(1) · 2010

A short review of the evidence for a central limit of 3 to 5 chunks in young adults. It describes the learned-pairs experiment (about three units recalled with rehearsal blocked, whether singletons or pairs), the cross-modal squares-and-digits experiment (about four items in total), and the developmental comparison (about 1.5 items in 7-year-olds versus about 3.0 in adults). It lays out two camps for why the limit exists: a cost of keeping many neural patterns active at once, or a benefit, since groups of about 3.5 items are efficient to search. The limit: the review is the author's own synthesis, and alternative "resource" models remain in play.

Read the paper (free)

What is Baddeley's model of working memory?

Baddeley's model treats working memory as a team of parts rather than a single box. Alan Baddeley and Graham Hitch proposed it in 1974 with three components, and Baddeley added a fourth, the episodic buffer, in 2000. Each part has its own limits, which is why you can hum a tune while picturing a route but struggle to read while someone talks to you.

ComponentWhat it holds or doesEveryday exampleSignature evidence
Phonological loop (1974)Speech-based traces that fade within seconds, refreshed by inner rehearsalRepeating an OTP under your breathSaying an irrelevant word aloud (articulatory suppression) wrecks verbal recall; similar-sounding words are harder
Visuospatial sketchpad (1974)Shapes, colours, locations and movementPicturing which lane to take at a flyoverSpatial tasks interfere with each other more than with speech tasks
Central executive (1974)Attention control: focusing, switching, inhibiting, updatingDropping the shopping list to answer a question, then resumingLess unified than first proposed; splits into several control functions
Episodic buffer (2000), "awareness buffer" (2026)A limited space that binds sound, image and long-term knowledge into one episodeHolding the gist of a sentence, not its exact wordsRecall of meaningful prose exceeds what the loop alone could hold

The four-chunk limit maps most naturally onto the buffer and the focus of attention, the part that holds integrated, conscious items. The loop and sketchpad are add-ons that let you exceed four when the material allows rehearsal or vivid imagery, which is exactly what Cowan's methods were designed to strip out.

Central executiveattention control Phonological loopabout 2 s of speech Episodic bufferabout 4 bound chunks Visuospatial sketchpadshapes, places Long-term memory: language, visual knowledge, episodes
Baddeley's model after 2000. The executive directs attention, three limited stores hold information, and all of them draw on long-term knowledge, which is where chunks come from.

Why is the limit about four?

Nobody knows for certain, and Cowan's 2010 review frames two camps. In the "weakness" camp, holding more items is biologically expensive: if the neural patterns for each item must fire in turn within a short cycle, perhaps every 100 milliseconds, too many items take too long or start to blur, so a red square and a blue circle get misremembered as a red circle and a blue square.

In the "strength" camp, a small store is a feature: simulations suggest searches are most efficient in groups of about 3.5 items, and a small set of active ideas can all be linked to one another without confusion.

Cowan's review also cites imaging work showing that one parietal region appears capacity-limited, at least for visual items. Newer models put the limit in learning as well as storage. A 2025 Brown University model of basal ganglia and thalamus circuits learned to chunk similar colours to save space, and broke down when asked to hold too many items, which suggests part of what we call capacity is a learned strategy.

How does chunking get around the limit?

Chunking works because the limit counts chunks, not raw items. "4, 0, 7, 1, 5, 2" is six items; "407" and "152" are two. Cowan's own example: to remember bread, milk and pepper, picture bread floating in peppery milk, and three items become one image. The chunks are supplied by long-term memory, so the more you know about a domain, the more each working memory slot can carry. That is why experts seem to have bigger memories only inside their field.

Indian oral tradition built this into its methods. The Vedas were preserved by reciting each text in several fixed patterns:

  • Padapatha: each word is said separately.
  • Kramapatha: words are recited in overlapping pairs (1-2, 2-3, 3-4).
  • Jatapatha: each pair goes forward, back and forward again (1-2, 2-1, 1-2).
  • Ghanapatha: longer groups are woven back and forth, such as 1-2, 2-1, 1-2-3, 3-2-1, 1-2-3.

Each pattern turns a long sequence into small, overlapping, heavily rehearsed chunks, and because every text existed in several versions, they cross-checked one another. UNESCO inscribed the Vedic chanting tradition on its intangible heritage list in 2008.

Same 12 digits, two ways to hold them As single digits: 12 items, about 4 slots 3492 1 4 5 0 2 5 0 8 ← no room, lost As chunks from long-term memory: 3 items, room to spare 34923:49.2, near-record mile 145014:50, a 5 km time 250825:08, a 10 km time The slot count does not change. What changes is how much each slot carries, and that depends on what long-term memory already knows. SF's running times worked for digits only; his letter span stayed near 6.
Chunking in one picture: four slots hold four loose digits, or three meaningful groups of four. The first group is SF's own example from Chase and Ericsson's report; the other two time labels are illustrative, drawn by the author.
Palm leaf manuscript of the Vedas in Sanskrit written in Grantha script from Tamil Nadu
A Vedic palm-leaf manuscript in Grantha script from Tamil Nadu. Writing came late to the Vedas; the recitation patterns did the work of preservation for centuries.Photo: Ms Sarah Welch, CC BY-SA 4.0, via Wikimedia Commons

How many chunks is that number?

Does having your phone nearby shrink working memory?

Probably not by its mere presence. A 2017 study by Adrian Ward and colleagues reported that people whose phones sat on the desk scored lower on a working memory test than those whose phones were in another room, an effect nicknamed "brain drain". A pre-registered direct replication of Ward's second experiment by Ana Ruiz Pardo and John Paul Minda in 2022 found no difference across six conditions (phone on desk, in a pocket or bag, or outside the room, each switched on or off) on either the operation span task or a go/no-go task.

The phone on the desk is not the thief. The notification that pulls your focus is.

What clearly does cost working memory is attending to the phone: a notification pulls the focus of attention, and with only about four slots, the item it displaces may be the one you needed. I cover that cost in attention residue and task switching, and the evidence on Indian screen time in India's screen time and attention.

Reexamining the "brain drain" effect: A replication of Ward et al. (2017)Ana C. Ruiz Pardo, John Paul Minda · Acta Psychologica 230 · 2022

A pre-registered direct replication at Western University in Canada. Participants were assigned to one of six conditions crossing phone location (desk, pocket or bag, outside the room) with phone power (on or off), then completed an automated operation span task, a cue-dependent go/no-go task and a smartphone attachment questionnaire. There was no difference between location conditions on either task, so the "brain drain" effect did not replicate. The limit: one lab and one sample; a 2023 meta-analysis from the University of Augsburg counted studies supporting an effect, so the honest summary is that any mere-presence cost is small and contested.

Read the paper

What should you do with a four-item mind?

Design tasks so you never ask working memory to hold more than about four new things at once. Write lists instead of rehearsing them. When studying, learn the building blocks until they become single chunks before stacking new material on top, which is the logic behind spaced practice in study techniques ranked by evidence. When comparing options, put them in a table so the comparison happens on paper. And protect the consolidation step that turns today's effortful items into tomorrow's free chunks, which happens largely in sleep (see how sleep turns learning into memory).

  • Group numbers into threes and fours: OTPs, account numbers, PNRs.
  • Offload to paper anything you need to hold while thinking about something else.
  • Build chunks deliberately: definitions, formulas and tables recited until retrieval is instant.
  • Cut interruptions during complex work; each one costs a slot, sometimes the important one.

More on the mind, from Nyaya logic to meditation research, lives in Manthan.

Questions people ask

Is working memory the same as short-term memory?

They overlap. Short-term memory usually means passive holding of information for seconds; working memory adds the active manipulation, such as reordering, comparing or calculating. Baddeley's model puts both inside one system with separate stores and an attention controller.

Can you increase working memory capacity with training?

The raw limit of about four chunks has proved hard to move. A 2013 meta-analysis of 23 studies by Monica Melby-Lervåg and Charles Hulme found short-term gains on the trained memory tasks but no convincing transfer to reasoning, attention, reading or arithmetic. What reliably increases is effective capacity in a domain, because expertise makes each chunk bigger. Learn the material, and the same four slots carry more.

Why do I forget why I walked into a room?

The goal was held only in working memory, and walking through a new space brought in fresh items that displaced it. With about four slots and no rehearsal, a single strong distraction is enough. Saying the goal out loud uses the phonological loop to keep it alive.

How did Vedic reciters memorise such long texts?

By turning long sequences into overlapping, rehearsed chunks through fixed patterns such as krama, jata and ghana patha, and by learning from childhood until each pattern was stored in long-term memory. Working memory then only had to hold the current small group, not the text.

Sources

  1. Cowan N. The magical number 4 in short-term memory. Behavioral and Brain Sciences 24(1):87-114 (2001)
  2. Cowan N. The magical mystery four: how is working memory capacity limited, and why? Current Directions in Psychological Science 19(1):51-57 (2010)
  3. Ruiz Pardo AC, Minda JP. Reexamining the "brain drain" effect: a replication of Ward et al. (2017). Acta Psychologica 230:103717 (2022)
  4. Baddeley A. Working memory: looking back and looking forward. Nature Reviews Neuroscience 4:829-839 (2003)
  5. Wikipedia: Baddeley's model of working memory (components and history) (accessed 2026)
  6. Allen RJ, Baddeley AD, Hitch GJ. Awareness as the heart of working memory. Trends in Cognitive Sciences (2026)
  7. Wikipedia: Vedic chant (recitation styles, citing Filliozat 2006 and Staal 1986) (accessed 2026)
  8. UNESCO: Tradition of Vedic chanting, Representative List of the Intangible Cultural Heritage (2008)
  9. Chase WG, Ericsson KA. Skilled Memory. Carnegie-Mellon University Technical Report No. 5 (DTIC ADA114635); see also Ericsson, Chase and Faloon, Science 208:1181-1182 (1980) (source for SF: 7 to about 80 digits, 250+ hours, running-time codes, consonant span about 6)
  10. Melby-Lervåg M, Hulme C. Is working memory training effective? A meta-analytic review. Developmental Psychology 49(2):270-291 (2013)
  11. ScienceDaily: How humans learn to optimize working memory (Soni and Frank, eLife) (2025)
  12. eLife reviewed preprint: Synaptic theory of chunking in working memory (2026)
working memoryNelson CowanAlan Baddeleymagical number sevenchunkingphonological loopepisodic bufferVedic recitation
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Mindblog.

Keep reading in Mindblog

From the other blogs