Chunking is the only lever that moves the limit on how much a person can hold. Everything else about that limit is fixed, and roughly awful. This one is free.
The word is Miller's, from 1956, and the mechanism is duller than the word suggests. A group your memory already treats as one thing costs the same to hold as a single letter. 4KP9 is four things the first time you see it and one thing the moment you have grouped it. Nothing was removed. The material is identical down to the character. What changed is the number of separate objects being tracked, and that number is the one with a ceiling on it.
So below is a twelve-character reference code, the kind that comes back from a booking or an order. You get four seconds, then it goes, then you type back whatever you still have. Six of them. Three are shown as one unbroken run of twelve. Three are shown as three groups of four. Same length, same alphabet, same four seconds.
Which kind comes first is a coin flip, and then they alternate. That matters more than it reads. The back half of any memory test is done by someone more practised and more bored than the person who started it, and if one condition always went second it would be carrying both of those on its own.
Six twelve-character codes, four seconds each, typed back from memory. Three are shown as one run of twelve. Three are shown as three groups of four. The page counts how many characters come back in the right place from each kind. It needs scripting to run, and scripting is off.
How it is counted
A character scores if it is the right character in the right place. Out of twelve. Drop one early and everything behind it shifts left and scores nothing, which is harsh, and it is harsh in exactly the same way for both kinds of code. Spaces you type are ignored, so grouping your own answer costs you nothing and gains you nothing.
Three pairs is a small number and the page says so at the end rather than rounding it up. A single lucky four seconds moves the average by more than half a character, and one code with a run of familiar digits in it will do the same.
What it costs you
The group size is part of the design, not a detail of it. Three and four work. Five is where a group stops being one thing and starts being a small list, which is the problem you were solving. A sixteen-digit number in two groups of eight has been formatted and not chunked.
Grouping against the meaning is worse than leaving it alone. A phone number split after the wrong digit, a sort code broken in halves, a date rendered as 2026-0 8-13 by a column that ran out of room. The reader has to undo your grouping before they can use theirs, and undoing costs more than the gap ever saved.
It is not only for numbers. Eleven form fields in one column is eleven things. The same eleven under three headings is three things and a place to stop. Nothing was removed there either.
The field that rejects the spaces is rejecting the chunking. Whoever grouped the code in the email and whoever wrote the validator for the field are usually two different people, and the person retyping it is a third one who did nothing wrong.
I group my own wifi password in fours when I read it out to someone. They write it down in threes and ask me to say it again slower. The grouping was mine the whole time.