Chunking

Also called memory chunking. Also in Psychology

Quick answer

Chunking, in cognitive psychology, is grouping separate items into one familiar unit so that working memory stores a single unit instead of many. It works because working memory is limited by the number of units, not by how much each unit contains. A person who knows the pattern holds a whole phone number or chess position as one chunk.

~20 sec read

011 min

Where chunking shows up

A product designer is reviewing a sign-up form in a critique. The form asks people to type a 16-character activation code. It appears as one unbroken line: K7Q2M9X4B1T8W3ZD. A colleague points out that anyone who looks away from the screen to type loses their place and has to start again.

The designer changes only the display. The code now reads K7Q2 M9X4 B1T8 W3ZD. The characters are the same, but a person now copies one group, glances back, and finds the next group. They work with four units instead of sixteen.

Now change one input. The product issues a code made of four common words: tiger lamp river stone. That code is 22 letters long, longer than the first one. A person still copies it with ease, because each word is a unit they already know.

The pattern is this: people handle information in units, and the size of a unit depends on what the person has already learned. A string the reader knows as one thing costs about as much attention as a single item. A string the reader does not know costs attention for every character.

022 min

Why chunking works

Chunking depends on two memory systems and the way they interact. Working memory is the small store that holds what a person is thinking about right now. Long-term memory is the large store of everything learned. A chunk is a unit that a person has learned to treat as one thing, and the learned pattern lives in long-term memory.

When a person sees a familiar pattern, working memory needs to hold only that one unit. The details of the pattern come from long-term memory when they are needed. This is why a word costs less attention than the same letters in random order.

The idea rests on a distinction Miller drew between bits and chunks. A bit is a unit of information in the technical sense: one yes-or-no answer. A chunk is whatever unit the person has learned to treat as one. Miller's central finding was that the limit on immediate memory is a number of chunks, and the information inside each chunk can keep growing.

The span of immediate memory seems to be almost independent of the number of bits per chunk, at least over the range that has been examined to date.

The operation that builds bigger chunks is called recoding: replacing a long input with a shorter code that stands for it. A person who has learned that the three binary digits 101 stand for the octal digit 5 can hold a long binary string as a short octal one.

Cowan's review of the evidence puts the central capacity limit at about four chunks, not seven. The exact number matters less than the structure of the limit. It counts units, so the way to hold more is to make each unit contain more.

A reader can use this to predict new cases. Chunking will help whenever the input contains patterns the person already knows. It will not help with items that have no pattern for that person, such as a random password.

032 min

Where chunking comes from

George Miller described the idea in a 1956 paper in Psychological Review. He had noticed that two different experiments produced a limit near seven. In one, people judged tones by pitch, and could sort them into only about seven categories. In the other, people repeated back a list of items they had just heard, and could recall about seven.

Miller argued that the two limits are different. He wrote that the span of absolute judgment and the span of immediate memory are quite different kinds of limitations. Absolute judgment is limited by the amount of information. Memory span is limited by the number of chunks.

He tested this with a demonstration run by Sidney Smith, which Smith reported to the Eastern Psychological Association in 1954. Before training, the 20 subjects had a span of 9 binary digits and 7 octal digits. Octal digits are the digits 0 to 7, and each one stands for exactly three binary digits. Smith then taught subjects schemes for recoding binary into octal, and later into larger groups.

Smith then trained himself on each scheme. The table shows what he could repeat back without error.

RecodingBinary digits per chunkBinary digits Smith could repeat
None19 (group average)
2:1224
3:1336
4:1 and 5:14 to 5about 40

Smith's span stayed at 12 chunks in every row. Only the amount of information in each chunk changed.

Miller gave a second example from outside the laboratory. He described a beginner at radiotelegraphic code who hears each dit and dah as a separate chunk. Soon the beginner treats letters as chunks, and later words and whole phrases. The amount of message the operator can remember grows each time the units get larger.

041 min

Chunking in chess expertise

The most cited case of chunking in a real skill is chess. In 1973, William Chase and Herbert Simon published "Perception in chess". They compared chess players of different strength as they looked at and then rebuilt board positions from memory. Their proposal was that stronger players do better because they recognise familiar groups of pieces, and each group acts as one chunk.

Later work put a size on the idea. Simon and Gilmartin (1973) estimated that the results of grandmasters may be explained by a repertoire of 10,000 to 100,000 configurations stored in long-term memory. The figure of 50,000 is the one quoted most often. It is an estimate, not a count.

This case adds something the number experiments do not. The chunks are not arbitrary groupings. They are meaningful patterns that a player learned over years of play. A master does not have a better working memory. A master has a larger library of chunks, so a real position takes fewer units to store.

The same logic limits the skill. A chunk helps only where its pattern applies, so a master's stored patterns give little help on a board position that does not resemble real games.

052 min

How chunking shows up in product, design, and AI

Forms and inputs. Long identifiers are easier to read and copy when they are split into groups. Phone numbers, card numbers and activation codes are routinely shown that way. The designer's choice is where to put the breaks. Breaks that match a pattern the user knows, such as a country code or a word boundary, make each group one chunk.

Navigation and settings. A settings page with thirty ungrouped options asks the reader to evaluate thirty units. The same options under four labelled sections give the reader four units to scan first. The labels do the work only if they name categories the reader already understands. A section called "Advanced" tells the reader little.

Code. An engineer reading unfamiliar code sees a loop, a condition and three assignments as separate items. A more experienced engineer recognises the same lines as "filter, then sum" and holds one unit. Named functions and common patterns exist partly to make code readable in chunks. Naming a block after what it does lets the next reader skip its internals.

Data and AI work. A data scientist reading an evaluation report with thirty metrics can group them into a few families, such as accuracy, cost and latency. The families become the units that go into the discussion.

The word also has an unrelated use in software. In retrieval systems, chunking means splitting documents into passages before indexing them, as in retrieval-augmented generation. That is a text-processing step. This article is about the memory process in people.

Chunking in design is almost always deliberate and helpful. The accidental failures are the opposite case: a product that presents many small, unrelated items because nobody grouped them.

062 min

How to use chunking deliberately

The goal of using chunking is to make the units match what the reader already knows.

  1. Group by meaning first, size second.

    Put items together because the reader already sees them as one thing. Groups of a fixed size help a reader keep their place, but a group of meaning is a true chunk, because the pattern is already in the reader's long-term memory.

  2. Name the group before listing its parts.

    A label such as "Billing" lets the reader decide to skip or read. The label becomes the unit in working memory, and the parts are retrieved only if needed.

  3. Keep the number of top-level groups small.

    Cowan's review found a central limit of about four chunks when people cannot recode the items. Designing around four to five visible groups is a reasonable target for unfamiliar material. It is a guide, not a rule.

  4. Teach the pattern, then use it.

    If a product has its own vocabulary, introduce it once with an example. After that, the term works as a chunk for the user.

  5. For your own learning, practise the recoding until it is automatic.

    Miller reported that the translation from one code to the other must be almost automatic, or the person loses part of the next group while remembering the translation of the last one. Smith needed repeated drilling before his results matched his span for octal digits.

A caution on the first point. A grouping that makes sense to the team that built the product is not necessarily a chunk for a new user. Test groupings with people who do not know the product.

071 min

Common misunderstandings about chunking

"People can hold seven items." This is the best-known misreading. Cowan notes that Miller meant his number seven more as a rough estimate and a rhetorical device than as a real capacity limit. Miller himself wrote about chunks, not items, and the number of chunks depends on what the person knows.

"Chunking increases memory capacity." It does not. The number of chunks stays about the same. What grows is the information inside each chunk. Smith held 12 chunks throughout his drills.

"Any grouping is a chunk." In Miller's account, a chunk is a familiar unit, and a great deal of learning has gone into forming these units. Splitting a random string with spaces helps a reader keep their place. It does not make the string a chunk until the reader knows the pattern.

Chunks are personal. A chunk for an expert may be several unknown items for a novice. A chess diagram that a master reads at a glance is a crowd of pieces to a beginner. Writing for experts with novice-level grouping wastes their time. Writing for novices with expert-level grouping loses them.

There are also cases where hiding detail inside a chunk is a mistake. A report that rolls ten different failure types into one label called "Errors" is easy to read and useless for diagnosis. Group for the reader's task, not just for brevity.

081 min

Chunking vs. nearby concepts

Chunking sits next to several concepts that people merge with it.

ConceptWhat it isHow it differs from chunking
PsychologyMiller's LawA statement of the limit. Chunking is the strategy for working within itWhat it is: The claim that people can hold only a limited number of items at once
PsychologyWorking MemoryA system. Chunking is a way of packing content into itWhat it is: The store that holds information in use
PsychologyMnemonicsDeliberate and taught. Chunking also happens without effort as people gain expertiseWhat it is: Deliberate memory aids such as rhymes and acronyms
PsychologySchema TheoryA broader theory of knowledge. A stored schema can supply the chunkWhat it is: Knowledge organised into structures that guide understanding
PsychologyCognitive Load TheoryA design theory. Chunking is one mechanism it draws onWhat it is: A theory of how instruction should account for working memory

The axis that separates chunking from the first two is simple. The limit and the store are fixed. Chunking is the part a person or a designer can change.

092 min

Where the evidence on chunking is contested

The number of chunks is the part with an open debate. Miller's headline was "seven, plus or minus two". Cowan (2001) argued that a better estimate is about four. His review concluded that a single central capacity limit averages about four chunks, and that higher numbers come from conditions where rehearsal or long-term memory lets people combine items into larger units.

Cowan's argument is not that Miller was wrong about chunks. It is that the larger estimates mix the capacity limit with the effect of chunking itself. When chunking is blocked, the observed limit is lower. Others do not accept a single fixed number, and the literature includes several competing models of the limit.

The chess estimate is contested too. Holding (1985, 1992) proposed that the number of chunks needed could fall to about 2,500 if patterns are encoded independently of colour and location. The two figures differ by a factor of 20, which shows how much depends on what counts as the same chunk.

A third difficulty is the definition. Miller admitted in 1956 that we are not very definite about what constitutes a chunk of information. That uncertainty has not gone away. A chunk is identified by how it behaves in an experiment, which makes the claim that memory holds a fixed number of chunks hard to test without circularity.

For a practitioner the disagreement changes little. Both estimates say the limit is small and counts units. Design for a few groups, test with the actual audience, and treat any specific number as an approximate guide.

101 min

How chunking changed since 1956

Miller used the word chunk in 1956 for a unit of memory whose size depends on learning. In 1973, Chase and Simon turned it into an explanation of expertise: strong chess players store many learned patterns, and later work debated how many. In 2001, Cowan reduced the estimated capacity to about four chunks and tied it to the focus of attention.

The meaning changed in one respect. Miller treated chunking as a way to extend a short-term span. Later work treated chunks as stored knowledge, which makes chunking a theory of how experts differ from novices.

?7 questions

Questions people ask

What is chunking in psychology?

Chunking is grouping separate items into one familiar unit so that working memory stores one unit instead of many. A person who knows that 1-9-4-5 is a year holds it as one chunk, not four digits.

What is an example of chunking?

Reading a phone number as three or four groups is a common example. A stronger case is chess: the often-quoted estimate is that masters store about 50,000 familiar patterns, each acting as one chunk.

How many chunks can working memory hold?

About four, according to Cowan (2001), when people cannot recode the items into larger units. Miller's 1956 figure of seven was a rough estimate. The count is in chunks, so what each chunk contains can differ widely.

What is the difference between chunking and Miller's Law?

Miller's Law states that people can hold only a limited number of units in mind. Chunking is the process that lets those units carry more information. One describes the limit and the other describes a way to work within it.

Does chunking really increase memory?

It does not raise the number of chunks, but it raises the information in each one. In Smith's drills, the same 12 chunks were worth 24, 36 and about 40 binary digits as the recoding got larger.

How do you use chunking in design?

Group items by meaning the reader already knows, label each group, and keep the number of visible groups small. Then test with people who are new to the product, because a chunk depends on what a person has learned.

Is chunking the same as chunking in AI?

No. In retrieval systems, chunking splits a document into passages before indexing it. In psychology, chunking is how a person groups items into units in memory. The two share a word and a general idea of grouping, but they are separate topics.

▶3 videos

Watch

  • Understanding Chunking

  • Chunking: the secret to fluency?

    Steve Kaufmann - lingosteve

  • Speak English Fluently with CHUNKING

    The English Class

§4 sources

Sources

Show all 4 sources
  1. Miller, G. A. (1956). The magical number seven, plus or minus two: Some limits on our capacity for processing information. Psychological Review, 63(2), 81-97. Copy at Classics in the History of Psychology:

  2. Cowan, N. (2001). The magical number 4 in short-term memory: A reconsideration of mental storage capacity. Behavioral and Brain Sciences, 24(1), 87-114.

  3. Gobet, F. (1998). Expert memory: A comparison of four theories. Cognition, 66(2), 115-152. PDF copy hosted on a blog:

  4. Chase, W. G., & Simon, H. A. (1973). Perception in chess. Cognitive Psychology, 4(1), 55-81. (73)90004-2

Keep reading

More from Design

All of Design
All of Design

Missing a concept?

Tell us what to write next. We’ll email you when it’s live.

Several? Separate them with commas.

Only to tell you when it’s live.

How we use your email: privacy.