Chunking

Abstract

Chunking is the process of grouping individual items into larger, meaningful units so that a limited short-term memory can hold more than its raw item capacity allows. Miller introduced the idea in 1956, distinguishing the informational bit from the psychological chunk and arguing that recoding many bits into a few chunks is how people escape the narrow span of immediate memory. Chase and Simon later showed that the recall feats of chess masters rest not on larger memories but on a store of learned board patterns that let them chunk a position at a glance. Modern estimates place the pure capacity limit near four chunks, and current accounts recast chunking as data compression that trades storage against encoding effort. This article defines chunking, traces its history, and reviews its mechanisms and limits.

Keywords: chunking, short-term memory, memory span, capacity limit, expertise

The string of digits 1 4 9 2 1 7 7 6 1 8 6 3 is twelve items long, comfortably beyond the handful most people can hold in mind at once. Read instead as 1492, 1776 and 1863 — three familiar dates — it becomes three items, and the same twelve digits are suddenly easy to keep and repeat. Nothing about the raw information changed; what changed is how it was organised. Chunking is the name cognitive psychology gives to that reorganisation: the grouping of separate elements into larger units that memory then treats as single things (Miller, 1956). It is the mechanism that lets a fixed, small memory store carry a large amount of structured material.

The construct matters because it dissolves an apparent paradox. Immediate memory is strikingly limited — only a few independent items survive a brief delay — yet skilled people routinely reproduce long strings, whole chessboards or pages of technical material after a single exposure. The resolution is that capacity is measured not in raw items but in chunks, and expertise consists largely in having learned the patterns that pack more raw material into each chunk (Cowan, 2001). The sections below define chunking, follow it from Miller's magical number through the chess studies that grounded it, set out the modern four-chunk capacity limit, examine how expertise builds chunks, and review the data-compression accounts that now formalise the idea.

## What Chunking Is

Chunking is the recoding of several individual elements into a single higher-order unit that short-term memory can hold as one item. The defining move is organisational: the raw material is not reduced or discarded but regrouped, so that what were many separate things become few composite things, each retrievable as a whole (Miller, 1956). A chunk is therefore defined functionally rather than physically — it is whatever the memory system treats as a single retrievable unit, whether that is a digit, a word, a familiar date, or an entire configuration of chess pieces. The size of a chunk is set by what the person already knows: the more structure a pattern has in long-term memory, the more raw material one chunk can absorb.

It is worth being precise about what the term does and does not claim. Chunking does not enlarge the capacity of short-term memory, which stays small and roughly fixed; it changes the currency in which that capacity is spent, from raw items to chunks (Cowan, 2001). Nor is a chunk merely a rehearsal grouping, though grouping and rehearsal often accompany it. A genuine chunk is bound together in memory as a unit, so that recalling any part tends to summon the rest. Holding this line keeps the construct doing specific work: it names the interface between a limited immediate memory and a vast long-term store, the process by which knowledge already held reshapes what can be taken in.

## The Magical Number Seven

The modern concept of the chunk begins with Miller's 1956 paper, which observed that across many tasks the span of immediate memory hovered around seven items — seven digits, seven letters, seven words — plus or minus two (Miller, 1956). The striking part of his argument was not the number itself but what it counted. Drawing on information theory, Miller distinguished the bit, the objective unit of information, from the chunk, the psychological unit, and pointed out that the span limit applied to chunks regardless of how many bits each contained. A person who can hold seven decimal digits can hold roughly seven unrelated words too, even though each word carries far more information, because the limit is on units, not on bits.

The practical consequence is that recoding raw material into richer chunks is the route around the bottleneck. Miller's own illustration was binary digits recoded into larger groups: a string of bits that overflows the span can be reorganised into a handful of multi-bit chunks that fit within it, dramatically increasing how much raw information the same seven slots carry (Miller, 1956). Later work revised the headline number downward — the true limit on independent chunks is smaller once grouping and rehearsal are controlled — but Miller's central insight survived intact: immediate memory is limited in chunks, and intelligent recoding is how the mind makes each chunk hold more (Cowan, 2010).

## The Chess Studies

The idea that chunks are learned units, not fixed groupings, was put on an empirical footing by Chase and Simon's studies of chess memory. Building on earlier observations, they showed masters and novices a chess position for five seconds and asked them to reconstruct it. On positions from real games the masters were vastly superior, reproducing far more pieces from the brief glimpse — but on boards with the same pieces placed at random, the master's advantage almost entirely vanished (Chase & Simon, 1973). The masters did not have better visual memory; they had a large repertoire of familiar configurations, and a random board offered none of them to recognise.

Analysing the timing of the reconstructions, Chase and Simon found that pieces were placed in bursts: within a burst the intervals were short, and between bursts they were long, and the pieces within a burst formed a meaningful chess relation — a defensive cluster, a pawn chain. Each burst, they argued, was the retrieval of one chunk (Chase & Simon, 1973). From this Simon estimated how much raw material a chunk holds and how many chunks immediate memory keeps, concluding that expertise multiplies the content of each chunk rather than the number of chunks available (Simon, 1974). Later work found that experts also hold larger, more abstract structures than the original small chunks could explain, prompting a template extension in which frequently seen patterns develop slots for variable detail (Gobet & Simon, 1996).

## The Capacity Limit

How many chunks can immediate memory actually hold? Miller's seven proved to be an overestimate of the pure limit, because ordinary span tasks let people rehearse and sub-group the material, inflating the count. When Cowan reviewed studies that blocked those aids — preventing rehearsal, using stimuli that resist grouping, or measuring the point where performance breaks — the recurring figure was closer to four (Cowan, 2001). On this account the capacity of the focus of attention is about three to five chunks, and the larger spans seen in everyday tasks reflect chunking and rehearsal layered on top of that core limit (Cowan, 2010). Table 1 sets the principal estimates side by side.

The estimates converge once the terms are matched. The higher numbers count chunks in tasks that permit recoding; the lower ones count independent chunks with recoding stripped away. Direct scoring of the chunks in expert chess recall likewise lands in the low single digits, suggesting the units people actually manipulate are few and dense rather than many and thin (Gobet & Clarkson, 2004). The interactive span demonstration below shows the same material crossing the limit or fitting within it depending only on how it is grouped.

## Expertise and the Building of Chunks

If capacity in chunks is fixed but chunk content is learned, then expert memory is a matter of acquiring richer chunks. The clearest demonstration is the case of a subject trained on digit span by Ericsson, Chase and Faloon: over some two hundred hours of practice the subject raised an ordinary span of about seven digits to nearly eighty, entirely by learning to recode incoming digits into meaningful units — running times, dates, ages — and to organise those units into a retrieval structure (Ericsson et al., 1980). His span for consonants, which he could not recode this way, remained ordinary, proving that the gain came from learned chunking and not from a general expansion of memory.

The mechanism generalises across domains. Grouping structure imposed by the experimenter helps too: when digit strings are presented in rhythmic groups, or when a hierarchical structure is supplied, recall improves in step with how well the grouping matches units the learner can encode (Bower & Winzenz, 1969). Formal models tie these findings together by treating chunk formation as ordinary learning: patterns that recur are gradually bound into units, and with enough exposure those units acquire internal slots that absorb variable detail, letting a single template stand for a whole family of configurations (Gobet et al., 2001; Gobet & Simon, 1996). Expertise, on this view, is a large, organised library of chunks and templates, built slowly, that lets the expert perceive and hold in a glance what the novice must take in piece by piece. The demonstration below contrasts a novice's item-by-item reading of a string with an expert's chunked parse.

## Chunking as Data Compression

The most active current framing casts chunking as data compression: the mind exploits redundancy in a sequence to encode it in fewer units than its raw length, exactly as a compression algorithm replaces a repeated pattern with a short code. On this account the capacity limit is best understood not in raw items but in compressed units, and a sequence with more internal structure — more regularity to exploit — should be held better because it compresses further (Mathy & Feldman, 2012). Mathy and Feldman showed that the apparent magic number falls out of this logic: when memory span is measured on maximally compressible sequences, the number of compressed chunks people hold is close to the low pure-capacity figure, even though the raw span is larger.

The compression view has since been sharpened and, in part, contested. A close test of whether verbal short-term memory truly compresses redundant material found that people benefit from familiar chunks but do not appear to perform general-purpose compression on arbitrary regularities, suggesting the mechanism is retrieval of known units rather than online re-encoding (Norris & Kalm, 2021). A complementary account models visual working memory as lossy compression, in which a limited store is allocated to minimise expected error, so that items are grouped and their detail sacrificed in whatever way best preserves overall accuracy (Nassar et al., 2018). The demonstration below makes the compression idea concrete by contrasting a structured sequence, which collapses to a short description, with a random one that does not.

## Chunking in the Brain

Deliberate chunking leaves a measurable signature in the brain. When people encode a sequence using an organising strategy rather than rote maintenance, activity in the lateral prefrontal cortex tracks the strategy rather than the raw memory load: imposing structure that reduces the number of chunks lowers demand on the maintenance system while recruiting prefrontal regions that build and apply the organisation (Bor et al., 2003). The finding fits the behavioural picture — chunking is an active encoding operation, not a passive property of storage — and locates the operation in the control circuitry that governs strategic cognition.

Chunking is not confined to verbal or visual memory; it also organises action. In studies of motor-sequence learning, practised movements come to be produced in stable groups, and the emergence of these motor chunks reflects a computation that trades the cost of storing a longer program against the cost of computing each step afresh (Ramkumar et al., 2016). The same principle — package recurring structure into a retrievable unit to spend a limited resource efficiently — appears whether the material is digits held for seconds or a movement sequence executed over months, which is part of why chunking is regarded as a general principle of cognitive organisation rather than a quirk of memory span.

## Worked Example

The power of chunking can be made concrete by counting units before and after recoding. Take the twelve-digit string 1 4 9 2 1 7 7 6 1 8 6 3. As twelve independent items it far exceeds both Miller's span of about seven and the stricter pure-capacity limit of about four chunks, so a single reading leaves most of it unrecoverable (Cowan, 2001). Now regroup the identical digits into three four-digit fields — 1492, 1776, 1863 — each of which a reader familiar with the dates recognises as one unit. The item count collapses from 12 to 3.

That regrouping does two things at once. It reduces the number of units from twelve to three, bringing the load from well above the four-chunk limit to well within it, and it does so without losing any information, because each date can be unpacked back into its four digits on demand (Figure 1). The compression ratio here is four to one: twelve raw digits carried in three chunks. Crucially, the trick works only because the dates already exist as units in long-term memory; the same digits grouped as 1 492 177 618 63, or presented to someone who knows none of the dates, would offer no such purchase and would stay near the limit. This is the whole logic of chunking in miniature — recoding raw items into fewer known units, so that a fixed small capacity carries far more structured material than its item count would suggest.

Figure 1

Twelve Digits Recoded into Three Chunks

Discussion

Chunking has proved to be one of the most durable ideas in cognitive psychology because it answers a question every account of memory must face: how a demonstrably tiny immediate store supports the large-scale cognition of literate, skilled adults. Miller's contribution was to reframe the limit as a limit on units rather than on information, which turned the puzzle of expert memory into a tractable question about how units are built (Miller, 1956). Chase and Simon then supplied the empirical anchor, showing that the units are learned patterns held in long-term memory, so that the apparently boundless memory of the expert and the narrow span of the novice are the same limited system operating on differently organised material (Chase & Simon, 1973; Simon, 1974).

The refinements since have not overturned this picture so much as formalised it. The capacity limit has been pinned closer to four independent chunks once recoding is controlled (Cowan, 2001), the units themselves have been elaborated from small chunks into structured templates (Gobet & Simon, 1996), and the whole process has been recast in the language of compression, which makes precise the intuition that structure is what buys extra capacity (Mathy & Feldman, 2012). What remains constant across every version is the core claim: the mind escapes a fixed, small memory limit not by enlarging the store but by packing more into each unit, using knowledge it already holds.

## Current Directions

Recent work is testing how far the compression metaphor can be taken and what exactly a chunk buys. A direct experimental question is whether a chunk simply occupies one slot or actively frees capacity for other material. Isolating this, Thalmann, Souza and Oberauer found that a pre-learned chunk does more than count as a single item: holding information in a familiar chunk released resources that improved memory for other, unchunked items in the same set, evidence that chunking offloads to long-term memory rather than merely relabelling a slot (Thalmann et al., 2019). This sharpens the classical account, which had left open whether chunks and single items compete for the same limited capacity in the same way.

A second front concerns the limits of the compression analogy itself. Careful tests find that verbal short-term memory does not compress arbitrary redundancy the way an ideal algorithm would; the benefit comes specifically from recognising familiar units, so the operative mechanism looks more like retrieval of known chunks than general re-encoding (Norris & Kalm, 2021). In parallel, rational-analysis models treat chunking in visual working memory as lossy compression tuned to minimise expected error, predicting when detail will be sacrificed to preserve gist (Nassar et al., 2018). The convergent trajectory is toward accounts that specify not just that chunking helps but how much, for which material, and at what cost in lost detail — moving chunking from a qualitative principle to a quantitative theory of how a limited memory allocates itself.

Common Misconceptions

Chunking increases the capacity of short-term memory.

It does not. The number of units the store can hold stays small and roughly fixed; chunking changes what counts as a unit, packing more raw material into each, so the same capacity carries more information (Cowan, 2001).

Experts remember more because they have better memories.

Their general memory is ordinary. Masters reconstruct real chess positions far better than novices but lose almost all of that advantage on random boards, showing the gain comes from learned patterns, not superior storage (Chase & Simon, 1973).

The magical number seven is the true capacity limit.

Seven counts chunks in tasks that allow rehearsal and grouping. With those aids removed, the limit on independent chunks is closer to four, and Miller's figure reflects recoding layered on that smaller core (Cowan, 2010).

Glossary

Bit.

The objective unit of information from information theory, the amount needed to decide between two equally likely alternatives; the quantity Miller distinguished from the psychological chunk.

Capacity limit.

The maximum number of independent chunks that can be held in immediate memory at once, estimated near four once rehearsal and grouping are prevented.

Chunk.

A collection of elements bound in memory as a single retrievable unit; whatever the memory system treats as one item, from a digit to an entire chess configuration.

Chunking.

The process of grouping separate items into larger meaningful units, so that a limited short-term memory holds more structured than unstructured material.

Data compression.

The encoding of information in fewer units by exploiting its redundancy; the framework in which modern accounts formalise chunking as the exploitation of structure to save capacity.

Immediate memory.

Memory for material just presented, held over a span of seconds; the limited store whose unit is the chunk and whose capacity chunking spends efficiently.

Lossy compression.

Compression that discards some detail to fit a limited store, allocating precision so as to minimise expected error; a proposed model of grouping in visual working memory.

Memory span.

The longest sequence of items a person can reproduce in order after a single presentation; the classic measure whose limit Miller placed near seven chunks.

Motor chunk.

A group of individual movements that practice welds into a single stored unit executed as a whole, the action counterpart of a memory chunk.

Recoding.

The transformation of material from one representation into another, richer one, such as replacing many bits with a few chunks; Miller's term for the route around the span limit.

Retrieval structure.

A learned organisational scheme that assigns incoming items to known locations for rapid later access, central to the skilled-memory feats of trained digit-span experts.

Skilled memory.

Exceptional memory performance within a practised domain, achieved by recoding material into meaningful units and organising them with a retrieval structure rather than by general capacity.

Template.

A large, structured chunk with slots for variable detail, developed through extensive exposure, that lets an expert encode a whole family of configurations as one unit.

Working memory.

The system that holds and manipulates information over short intervals in the service of ongoing cognition, and whose limited capacity chunking helps to stretch.

Key Researchers

Nelson Cowan (living). Cognitive psychologist at the University of Missouri; his embedded-processes model and 2001 review reset the pure capacity limit of immediate memory near four chunks, separating the core limit from the recoding and rehearsal that inflate everyday span. Homepage - Google Scholar - ORCID

K. Anders Ericsson (1947-2020). Psychologist at Florida State University; his skilled-memory studies, including the trained digit-span subject who reached a span near eighty, showed that exceptional memory is built from learned chunking and retrieval structures rather than raw capacity. Wikipedia - Google Scholar - ORCID

Fernand Gobet (living). Cognitive scientist at the London School of Economics; with Simon he developed the template theory of expert memory and the CHREST model, distinguishing implicit perceptual chunks from the structured templates that experts acquire through long practice. Wikipedia - Google Scholar - ORCID

George A. Miller (1920-2012). American psychologist and a founder of cognitive science; his 1956 paper coined the chunk, distinguished it from the informational bit, and located the span of immediate memory near seven units, framing recoding as the route around that limit. Wikipedia

Klaus Oberauer (living). Cognitive psychologist at the University of Zurich; his experimental work isolates how holding material in a pre-learned chunk frees working-memory capacity for other items, sharpening the classical account of what a chunk actually buys. Wikipedia - Google Scholar - ORCID

Herbert A. Simon (1916-2001). Polymath at Carnegie Mellon University and Nobel laureate; with Chase he demonstrated that expert chess memory rests on learned board patterns and estimated the size and number of chunks, placing chunking at the centre of the study of expertise. Wikipedia

Frequently Asked Questions

What is chunking in psychology?
Chunking is the process of grouping separate pieces of information into larger, meaningful units so that a limited short-term memory can hold more of them. Reading the digits 1 4 9 2 as the year 1492 turns four items into one, and immediate memory holds units, so the regrouping lets the same small store carry far more raw material (Miller, 1956).

Who came up with the idea of chunking?
The concept was introduced by George A. Miller in his 1956 paper on the magical number seven. Miller distinguished the informational bit from the psychological chunk and argued that the span of immediate memory is limited in chunks, so recoding material into richer chunks is how people carry more information in the same number of slots (Miller, 1956).

How many chunks can short-term memory hold?
Miller's famous estimate was about seven, but that figure counts chunks in tasks that allow rehearsal and sub-grouping. When those aids are prevented, the limit on independent chunks is closer to four, which is now the more widely accepted estimate of pure capacity (Cowan, 2001).

How does chunking explain expert memory?
Chase and Simon found that chess masters reconstruct real positions far better than novices but do no better on random boards. Their advantage comes from a large store of learned patterns that let them chunk a meaningful position at a glance, so expertise multiplies the content of each chunk rather than the number of chunks (Chase & Simon, 1973).

Does chunking make memory bigger?
No. Chunking does not expand the capacity of short-term memory, which stays small and roughly fixed. It changes what counts as a unit, packing more raw material into each chunk, so the same limited capacity carries more information without any increase in the number of units held (Cowan, 2001).

What is the difference between a bit and a chunk?
A bit is an objective measure of information, while a chunk is a psychological unit that memory treats as a single item. The span limit applies to chunks regardless of how many bits each holds, which is why seven unrelated words are about as hard to retain as seven digits despite carrying far more information (Miller, 1956).

Is chunking the same as data compression?
Modern accounts treat chunking as a form of compression, exploiting structure in a sequence to encode it in fewer units. The analogy is powerful but imperfect: verbal memory benefits mainly from recognising familiar chunks rather than compressing arbitrary redundancy, so the mechanism looks more like retrieval of known units than general-purpose compression (Norris & Kalm, 2021).

Can chunking be trained?
Yes. A subject trained by Ericsson and colleagues raised his digit span from about seven to nearly eighty over months of practice by learning to recode digits into meaningful units and organise them with a retrieval structure. The gain was specific to trained material, confirming that it came from learned chunking rather than a general memory improvement (Ericsson et al., 1980).

References