Miller on the limits of working memory
📚SOURCE
💎PERMANENT NOTES
🧵SYNTHESIS
Two different limits share this paper, and most people who cite it merge them. The first is absolute judgement: ask someone to label tones by pitch, or points by position on a line, and they manage about six or seven categories before errors climb. Miller measures this as channel capacity in bits. Add a second dimension and people distinguish more in total, though less finely on each.
The second is the span of immediate memory: how many items you can repeat back after one hearing. That also comes out near seven. But here the unit is not information. People hold about as many words as digits, although a word carries far more. What is limited is the number of chunks, and a chunk is whatever you have learned to treat as one thing.
That opens the escape route he calls recoding: regroup the input into bigger familiar units and the same span holds more. His example is a colleague who learned to read strings of binary digits as larger numbers and could then repeat back about forty.
He ends by doubting his own title. The sevens may be a coincidence.
✨KEY IDEAS
People can sort a one-dimensional stimulus into about seven categories reliably, and not many more.
Memory span is fixed in chunks and nearly indifferent to how much each chunk contains.
Recoding into larger units is how the limit is stretched, and it takes learning.
Using several dimensions at once raises the total you can tell apart.
He treats the recurrence of seven with open suspicion.
📝READING NOTES
Paraphrase, not quotation. Psychological Review, 1956. It began as a talk and reads like one: he opens by saying a number has been following him about.
Why I read it
To argue for a shorter exception list. Rafael's dispatchers at the pilot opened the day with 214 items each. I wanted something better than that seems like a lot.
What it does and does not give me
It does not say a list should have seven things on it. A list is on a screen and does not have to be held in the head, which is the point of a list. 💡 Miller's seven counted chunks, not items, and he was half joking about the number. Later work puts the span of unrehearsed memory lower, nearer four.
What it does give me:
- The comparison is the bottleneck. To triage, a dispatcher holds several exceptions in mind at once to decide which is worst. That set is small. Past it they are not triaging, they are scrolling.
- Chunking is the design question. Twelve late containers on one vessel are one thing, if we present them as one thing. Grouping is recoding done by the software.
- Experts chunk differently. A first-week dispatcher and a ten-year one do not see the same screen.
Where I part from him
He writes as though the limit is a property of the person. In a control room it is also a property of the morning: phones, a driver at the desk, a customer on hold. Nobody measures span under that. I would guess three.
Where it went
- 🧩 Working memory limits, the concept note.
- 💎 Working memory holds a handful of things so the list must hold the rest.
- The argument section of 🔍 Exceptions digest by e-mail: five things at 07:00.
- Sits oddly well next to 📖 Matuschak on evergreen notes: a note title that states a claim is a chunk.
Where it meets the Inbox
One line for the spec: the product should do the chunking, so the dispatcher does not have to.
- Group by cause before listing by time.
- Never ask for a comparison across more than a handful of groups.
- Let the digest carry the top five, and the full list carry the rest.
🗣️IN SHORT
- In : The Magical Number Seven, Plus or Minus Two, Psychological Review, 1956
- Miller argues that : both our ability to judge along one dimension and our immediate memory are limited to a small number of units
- by stating that : the memory limit counts chunks and not information, so that recoding material into larger familiar units lets the same span carry far more