George Miller's 1956 paper put a number on working memory — about seven chunks. The number itself is fuzzy; the lesson is that memory is small, and the design's job is to group information so users can hold it without strain.
7 ± 2chunks in working memory
The Law · Miller, 1956
There's a small ceiling to how much a person can hold in working memory at one moment. The trick isn't to shrink interfaces to that ceiling — it's to group information into chunks the user can fit under it.
Coined by
George A. Miller (1956)
Capacity
~7 ± 2 chunks
Workaround
Chunking
1956 study
The Magical Number Seven
Miller surveyed studies on pitch perception, taste discrimination, and short-term recall — and the same ceiling kept showing up: about seven items. He never argued this should set menu length; that's a misreading that became gospel.
Working memory · the original ceiling
Original
Pitch · taste · recall
Range
5–9 chunks
Misread as
"7-item menus"
Compression
Chunking
The way around the limit. Group raw items into meaningful chunks and you fit more in memory. A 10-digit phone number is hard; (555) 867-5309 is easy — three chunks instead of ten digits, and the chunks themselves are familiar shapes.
Three chunks beat ten digits
Mechanism
Group into meaning
Classic
Phone · IBAN · dates
Effect
More info, same load
Related heuristic
Recognition Over Recall
Even when memory is bounded, recognition is far cheaper than recall. Show the options on screen and let the user point at the right one. Menus, autocomplete, breadcrumbs, recently-used lists, and saved searches are all Miller workarounds dressed as features.
Show, don't quiz
Cheaper
Recognition
Costly
Recall from blank
Apply
Surface, don't quiz
Classic case
Format Numbers Into Chunks
The textbook example. (555) 867-5309 is three chunks; 5558675309 is one ten-digit blur. The same logic applies to credit cards, IBANs, dates, reference numbers, and OTP codes — format for the eye and the memory, not for the database column.
Group what the user must hold
Pattern
Group digits
Apply to
Phone · card · IBAN · OTP
Avoid
Raw run-on strings
IA
Group Information Architecture
Navigation works best when each level has a small, scannable set. Card-sorting research lands consistently around five to nine categories per level. Beyond that, users start to lose the map of where things sit, and back-tracking spikes.
~ Seven per level, then nest
Target
5–9 per level
Research
Card sorting
Symptom
Lost in navigation
Pattern
Progressive Disclosure
Show the few choices that matter now, reveal the rest on demand. Settings panels, forms, and onboarding all benefit — keeping any one screen under the working-memory ceiling without hiding what advanced users need a click away.
A few now, the rest on request
Show
The 5–9 that matter
Reveal
On request
Watch for
Hiding the essential
Flow
Multi-step Wizards
Long forms collapse into wizards for Miller-shaped reasons. Each step asks for a chunk small enough to hold while deciding — the alternative is a forty-field page where users lose track of what they've already answered halfway down.
One chunk per step
Pattern
One chunk per step
Best for
Long forms · setup flows
Watch for
Step bloat hides total length
Myth
The Common Misuse
"Menus should have at most seven items" — Miller never said this. His work was about chunks in working memory, not interface element counts. Menus are scanned visually, not held in memory. The 7 ± 2 number is a guide to grouping, not a hard ceiling on what's on screen.
Visual scan ≠ memory load
Myth
7-item menus
Reality
Chunking, not counting
Apply
Group, then size
Take-away
Chunk Size, Not Item Count
The right read: when a user must hold something in mind across a task — a code, a step, a comparison — keep the chunk at five or fewer. When they're scanning with the screen as memory, the limit doesn't apply. Distinguish "carry it" from "see it."
Carry it vs see it
In memory
≤ 5–9 chunks
On screen
No hard cap
Test
Can they hold it?
Refinement
The Number Got Smaller
Decades of follow-on research narrowed the figure — Cowan's 2001 paper suggests closer to four chunks for genuine working memory. The principle holds, only sharper: memory is small, expensive, and the design's job is to protect it.
Seven → four
Original
Miller, 1956 · ~7
Refined
Cowan, 2001 · ~4
Practical
Plan around ≈ 4 safe
✦
Miller's Law in the Age of AI
The assistant can hold the chunks a user can't — but the user still has to verify what they couldn't keep in mind themselves.
✦ AI Era
The Assistant Holds Context for You
A model can carry the full set of variables a user can't. "What was that customer's plan?" or "remind me which file we last edited" no longer demands the user keep it in head — the assistant remembers across the session and surfaces it on cue.
Memory offloaded to the model
Shift
User → assistant memory
Helps
Long multi-step tasks
Watch for
Stale or wrong context
✦ AI Era
The Risk: Trusting What You Can't Hold
When the model handles the chunks the user can't, the user loses the ability to verify. Surface sources, intermediate steps, and undo so people can re-establish ground truth — don't let "I can't hold it all" become "I never had to check."