What Chunking Actually Means
Chunking is recoding: you take several small units of information and rebuild them as one familiar unit. A phone number stops being ten digits and becomes three groups. A chess position stops being twenty-odd pieces and becomes a castled king with a broken pawn structure. Nothing about your memory changed. What changed is the size of the thing you're counting.
The term comes from George Miller's 1956 paper in Psychological Review, "The Magical Number Seven, Plus or Minus Two," which opens with one of the best first lines in psychology: "My problem is that I have been persecuted by an integer. For seven years this number has followed me around, has intruded in my most private data, and has assaulted me from the pages of our most public journals."
Miller's real contribution wasn't the number. It was the distinction between a bit and a chunk. He wrote that "the span of immediate memory seems to be almost independent of the number of bits per chunk," and then drew the conclusion that matters for anyone trying to learn something: "Since the memory span is a fixed number of chunks, we can increase the number of bits of information that it contains simply by building larger and larger chunks, each chunk containing more information than before."
His example is a radio operator learning Morse code:
A man just beginning to learn radiotelegraphic code hears each dit and dah as a separate chunk. Soon he is able to organize these sounds into letters and then he can deal with the letters as chunks. Then the letters organize themselves as words, which are still larger chunks, and he begins to hear whole phrases.
The operator's memory span never grows. His chunks do. Miller called the process recoding and thought it was underrated, arguing that "the kind of linguistic recoding that people do seems to me to be the very lifeblood of the thought processes." Seventy years later, most study advice still treats memory as a container to be filled rather than a code to be rewritten.
One more detail from the paper is worth keeping, because it shows how slippery a chunk is. Miller cites a memory span of five words drawn at random from a thousand English monosyllables. Those same five words contain about fifteen phonemes. Was the span five or fifteen?
Miller's own answer was that "intuitively, it is clear that the subjects were recalling five words, not 15 phonemes, but the logical distinction is not immediately apparent." That unresolved distinction is the point. The unit is a property of the person, not of the material.
The Number Everyone Quotes Is Wrong
Miller's seven plus or minus two escaped into popular culture and hardened into a law. It isn't one, and Miller never presented it as one.
Nelson Cowan reopened the question in 2001 with "The magical number 4 in short-term memory," published in Behavioral and Brain Sciences. Pulling together evidence from tasks designed to block rehearsal and grouping, he argued for a real capacity limit of 3 to 5 chunks, averaging about four. Cowan's own later reviews and much of the work since have supported a limit in that range, though the mechanism behind it is still argued over. If you're planning how much to hold in mind at once, four is the honest planning figure.
So why did seven hold up for so long? Fabien Mathy and Jacob Feldman answered that in 2012 in Cognition, and their answer is what reconciles the two numbers. They defined a chunk in terms of compressibility and concluded that four is the capacity in maximally compressed units, while Miller's seven is the length of an uncompressed sequence of typical complexity, which compresses at a ratio of roughly 7 to 4.
Both numbers are right. They're measuring different things.
| Claim | Source | What it measures | Value |
|---|---|---|---|
| Span is about seven items | Miller (1956), Psychological Review | Items in a typical list, before compression | 5 to 9 items |
| Capacity is about four chunks | Cowan (2001), Behavioral and Brain Sciences | Chunks held when rehearsal and grouping are blocked | 3 to 5 chunks |
| Both numbers describe one limit | Mathy and Feldman (2012), Cognition | How much a typical list compresses | 7 items compress to about 4 chunks |
The practical reading: you get a handful of slots, and the variable you control is what you put in them. It's a reasonable explanation for why a summary crammed with nine bullet points is harder to hold than one well-named idea. If this framing is new to you, our guide to cognitive load theory works the same limit from the reader's side, on material nobody designed for you.
Why Chess Masters Can't Fake It
The cleanest demonstration that chunks live in long-term memory comes from chess. Adriaan de Groot showed in the 1940s that masters could rebuild a real game position after seeing it for only a few seconds, while weaker players couldn't. That looks like superhuman memory until you change one variable.
William Chase and Herbert Simon followed up in 1973 with "Perception in Chess" in Cognitive Psychology. When the pieces were scattered randomly instead of arranged as a plausible game, the master's huge advantage mostly evaporated. Their explanation: masters have ordinary working memory and extraordinary chunks. A real position decomposes into patterns they've seen thousands of times. A random one doesn't decompose at all.
The story got refined rather than overturned. Fernand Gobet and Simon reported in 1996 in Psychonomic Bulletin & Review that strong players do keep some superiority on random positions, though the gap is far smaller than on game positions. A pure chunk-counting model can't explain that residue, which is why Gobet and Simon extended the theory the same year, in a separate paper in Cognitive Psychology, to larger and more flexible long-term structures they called templates.
Two things follow for anyone learning anything.
First, a chunk you haven't built yet is worthless to you. Reading someone else's neat summary of a field gives you their chunks, not yours. This is the mechanism behind the illusion of competence: fluent material feels chunked while you're looking at it, and the feeling disappears when the page does.
Second, chunks come in two flavors. Gobet and colleagues made this explicit in Trends in Cognitive Sciences in 2001, distinguishing deliberate chunking that is strategic and goal-directed from automatic chunking that runs continuously off perceptual experience. You can force the first. The second is what thousands of hours buys you, and there's no shortcut.
From 7 Digits to 79: How Far Chunking Goes
In 1980, K. Anders Ericsson, William Chase, and Steve Faloon published a two-page paper in Science called "Acquisition of a Memory Skill." Their participant, known as SF, practiced digit-span recall in the lab. After more than 230 hours of practice, his span went from 7 digits to 79.
That number deserves a moment. Not 9 digits, not 15. Seventy-nine, read once, repeated back. Chase and Ericsson's later accounts of the same participant put him at 82 digits, after about 250 hours spread over two years.
His memory did not improve. The paper's own conclusion is that with an appropriate mnemonic system, "there is seemingly no limit to memory performance with practice," and the wording of the abstract is careful about the scope of what he gained: his performance on other memory tests with digits matched that of memory experts with lifelong training.
The clincher is what didn't move. When the sequences were made of letters instead of digits, his span stayed in the ordinary range, about six or seven. The skill lived in the coding scheme he'd built for digits, not in some general upgrade.
That scope limit is the real lesson, and it's easy to get backwards. Chunking doesn't give you a better memory. It gives you a better code for one specific kind of material, and you have to build a new code for each new kind. The person who can recall 79 digits and the person who can recall a dense paper after one read are using the same mechanism on different corpora.
A 2025 model suggests a circuit can learn to do this on its own. Aneri Soni and Michael Frank at Brown simulated chunking in a prefrontal cortex and basal ganglia circuit in eLife, and the network learned to reuse the same prefrontal populations to hold several items at once, buying quantity at the cost of precision. Their reading is that the ceiling is software rather than hardware: a network that never learns a good gating policy performs below the capacity it was handed. It's a simulation, so treat it as a hypothesis about the mechanism, not a measurement of anyone's brain.
Chunking Frees Capacity You Didn't Know You Had
There's an obvious objection to all of this: maybe chunks don't reduce load, they just relabel it. Maybe you're still doing the same work and counting it differently.
Mirko Thalmann, Alessandra Souza, and Klaus Oberauer tested that directly in the Journal of Experimental Psychology: Learning, Memory, and Cognition in 2019, in a paper titled "How does chunking help working memory?" Across four experiments they found that the benefit of chunking showed up not only in recall of the chunked material, but also in recall of other, unchunked items held at the same time. Chunks genuinely cost less. The capacity they free is real, and it's available for something else.
Two limits from the same paper are worth carrying. The spillover came from chunks early in the list, not from chunks at the end. And the benefit stopped being independent of chunk size once the chunks shared elements, which is why the authors conclude that working memory "is not limited to a fixed number of chunks regardless of their size." Bigger chunks still cost more. They just cost less than their parts, and that is the honest version of the four-slot picture: the slots are not literal and a bigger chunk is not free. What holds is the direction, not the arithmetic.
This is why chunking sits underneath every other study technique rather than competing with them. Recall practice on a well-chunked page tests a few ideas, not a long list of facts. Spaced review gets shorter for the same reason: fewer items. And a hard paragraph stops feeling like juggling once you're holding two units instead of nine clauses.
It also explains a frustration most readers know. Dense material isn't hard because the sentences are long. It's hard because the author's chunks aren't yours yet, so every clause spends a slot. Reading slower doesn't help. Stopping to name the unit does.
How to Chunk What You Read
Chunking while reading is a deliberate act with a physical trace. Here's the method that works, in five steps.
1. Read a whole section before you mark anything. You can't find the boundary of an idea from inside the first sentence. It's an easy mistake to make: people highlight in the first paragraph and then keep highlighting to stay consistent with the first mark.
2. Mark the one passage the section turns on. One per section, not five. A highlight is a claim about where a chunk begins and ends, so making it costs you a decision, which is the point. Our piece on the science of highlighting covers what the research says about doing this well versus badly.
3. Name the chunk in your own words. Add a short note to the highlight, three to seven words, no jargon borrowed from the page. Naming is the recoding step Miller described. Without a name you have a saved quote, not a chunk.
4. Give the chunk a handle you'll reuse. A tag, a topic, a recurring label. Chunks compound only if the same label keeps collecting new material, which is how a scattered set of highlights turns into a structure. Marking passages with Glasp's web highlighter and tagging them as you go gives each chunk an address you can return to.
5. Try to rebuild the section from the names alone. Cover the page. If four names bring back the argument, the chunks hold. If they don't, your boundaries were wrong and you can see exactly where. That test is active recall applied to structure rather than facts.
Two notes. Books work the same way: Kindle passages you've already marked are raw chunk candidates, and importing your Kindle highlights puts them somewhere you can name and group them instead of leaving them in a device. And when you want to check a chunk boundary against someone else's, the Glasp community shows what other readers pulled out of the same page, which is a quick way to find the sentence you skipped.
How to Chunk a Video or a Lecture
Video is the hardest format to chunk because it doesn't stop. Text lets you sit with a paragraph. A lecture keeps talking while you're still holding clause three.
The evidence here is unusually specific, and unusually honest about its own limits. Günter Daniel Rey and colleagues published a meta-analysis of the segmenting effect in Educational Psychology Review in 2019, covering 56 investigations and 88 pairwise comparisons. Cutting a continuous presentation into meaningful pieces produced:
- Better retention, d = 0.32, and better transfer, d = 0.36, both significant.
- Lower reported cognitive load, d = 0.23.
- A bigger gain for learners with high prior knowledge than for those with none or a little, and only on retention.
- Longer time on task. Segmented material took more time to get through, which is a genuine cost rather than a rounding error.
Then it gets interesting, because the pacing result flips depending on what you measure.
| Condition | What it means | Retention | Transfer |
|---|---|---|---|
| System-paced segmentation | Segmented material, delivered at a preset pace | d = 0.42, significant | d = 0.35, significant |
| Learner-paced segmentation | Segmented material, learner controls the pace | d = 0.19, not significant (p = 0.10) | d = 0.45, significant |
| All studies pooled | Both together | d = 0.32, significant | d = 0.36, significant |
It's tempting to read the top row as proof that a designer's cuts are what matter and your own pausing is worthless. The authors tested exactly that idea and couldn't confirm it. Their moderator analysis found no support for the hypothesis that the benefit comes from a designer cutting the material at meaningful joins. And on transfer, the learner-paced studies came out ahead.
The defensible reading is narrower than the usual advice. Pre-cut segments have the strongest evidence for straight retention. Controlling the pace yourself looks better when the test is whether you can use the material on a new problem. Nobody has shown which mechanism is doing the work. What survives either way is the boundary: a segment has to be a meaningful piece of the argument, since a stop in the middle of a clause doesn't produce a chunk.
So segment deliberately. Before watching, get the structure: a YouTube Summary with timestamped sections hands you a first draft of the segment boundaries, and you can then treat each timestamp as one chunk to name. Highlight one line of the transcript per segment rather than collecting quotes as they fly past. If the video has no structure worth segmenting, that's useful information about the video. For the broader workflow, see how to learn from YouTube.
Chunking Compared to Other Study Techniques
Chunking gets filed next to the popular study techniques, which obscures what it does. Rehearsal methods change how often you touch the material, and scheduling methods change when. Chunking changes the size of the units they all operate on.
| Technique | What it changes | When it's the right move |
|---|---|---|
| Chunking | The size of the unit you hold | Material feels overwhelming rather than unfamiliar |
| Active recall | Retrieval strength | You recognize the content but can't produce it |
| Spaced repetition | Timing of review | You knew it last month and it's fading |
| Interleaving | Order of practice | You can do each type but can't tell them apart |
| Dual coding | Number of channels used | The idea has a spatial or visual structure |
| Mind mapping | External layout of relations | You need to see how parts connect |
The ordering matters more than the list. Chunk first, then rehearse. Running spaced repetition over sixteen unrelated facts is a scheduling problem you invented for yourself by never doing the recoding step. Four named chunks are cheap to keep alive.
And if you're weighing mind maps against this, note the difference in what each one claims. A mind map arranges units you already have. Chunking creates the units. Our review of whether mind maps work goes through the evidence for the arranging half.
Five Ways Chunking Goes Wrong
Chunking around the wrong boundary. The heading structure of an article is the author's chunking, not yours, and it's often shaped by word count. If a section's real idea starts two paragraphs in, cut it there.
Building chunks on parts you don't know. A chunk made of components you can't unpack is a label over a hole. Miller's radio operator built his letters out of dits and dahs he already knew. If you can't expand a chunk on request, it isn't compression, it's a bookmark.
Too many chunks. Fifteen well-named chunks on a chapter puts you right back over the limit. On most articles, three to five is the honest count, and being forced to choose is what makes the method work.
Confusing fluency with a chunk. This is the big one. A passage that reads smoothly feels chunked while it's in front of you, and the test is whether you can rebuild the section without it. See the illusion of competence for how reliably this fools people, including people who know about it.
Assuming the chunks transfer. SF's 79 digits didn't make him better at anything except digits. Your chunks for one field don't carry to the next one, though your habit of building them does.
Frequently Asked Questions
What is chunking in simple terms?
Chunking is grouping small pieces of information into one meaningful unit so that your working memory treats them as a single item. A ten-digit phone number recoded into three groups is the standard example. The number of things you can hold at once stays the same, but each thing carries more.
Is chunking the same as memorizing?
No. Memorizing is about getting material into long-term storage and keeping it there. Chunking is about how material is organized so that less of it is needed to represent the same idea. Chunking makes memorizing cheaper, which is why it pairs so naturally with recall practice and spaced review.
How many items should be in a chunk?
There's no fixed answer, because a chunk is defined by meaning rather than by count. A useful working rule: a chunk should be something you can name in a few words and unpack on demand. If you can't unpack it, it's too big. If it doesn't earn its own name, it's too small.
Does chunking work for reading?
It works for reading, and it's arguably more valuable there. The classic experiments used digits and chess positions because those are easy to score, but Miller's own framing was about recoding in general, including language. Reading a dense paragraph is a working memory task, and naming the unit is what stops every clause from consuming a slot.
What's the difference between chunking and the memory palace?
A memory palace is a retrieval system: it gives items distinct locations so you can walk back through them in order. Chunking is a compression system: it reduces how many items there are to place. The two combine well: the memory palace technique gives you the loci, and chunking decides how many you need. Nobody has tested the combination head to head, so treat that as a reasonable expectation rather than a finding.
What are some examples of chunking?
A phone number read as three groups instead of ten digits. A chess position seen as a castled king and a broken pawn structure instead of twenty-odd pieces. Morse code heard as words instead of dits and dahs. A card number, a chord name, a postal code, or a whole section of an article held as one claim instead of nine sentences. In every case the material is unchanged and the count went down.
Can AI chunk information for me?
Partly, and it's worth being precise about the limit. A summarizer can propose boundaries, which is real help on a long video or an unfamiliar paper. But the chess research says the chunk has to exist in your long-term memory to do any work, so a boundary someone else drew is a starting point for your own recoding, not a replacement for it. Use the machine for the structure, then name the units yourself, which is how Glasp's AI chat and YouTube Summary are meant to be used: they draft the boundaries, you own the names.
Start With One Chunk
Miller's number didn't survive, but his reframing did: what you can hold depends on what counts as a unit, and the unit is the part you control. Cowan lowered the count, Mathy and Feldman explained why both counts were right, Chase and Simon showed the units have to be yours, and Ericsson's participant demonstrated how far the lever moves when someone leans on it for more than 230 hours.
None of that requires a new app or a system. It requires stopping at the end of a section and naming what it was about, in your own words, before moving on.
Try it on the next article you read. Mark one passage per section with Glasp, give each one a name you'd actually say out loud, and at the end rebuild the piece from the names alone. Four names should do it. If they do, you've just compressed an article into something that fits in working memory, which is the entire skill. For the retention side of the same problem, read how to remember what you read.
