In 1974, Alan Baddeley and Graham Hitch did something that feels almost rude — they sat people down, showed them a list of words, and then asked them to repeatedly say “the, the, the, the, the” out loud while trying to remember those words. Not a whisper. Not a mumble. Out loud, over and over, like a human skipping record. And what they found was not subtle: people’s memory for the words collapsed. Not by a little — by more than half.
This was not a quirk. It was the first clue that your inner voice — that quiet narration that runs through your head when you read a sentence, rehearse a phone number, or replay a conversation — is not a metaphor. It’s a physical cognitive system with real capacity limits. And if you jam it, you can watch memory fall apart in real time.
🧠 The Workshop, Not the Bucket
Before Baddeley and Hitch, the standard view of short-term memory was simple: it’s a temporary bucket. You dump information in, it sits there for a bit, then it either moves to long-term storage or falls out. A single container. A single capacity limit. Nothing fancy.
Baddeley thought this was wrong. Not because the data contradicted it — but because the experience of thinking contradicts it. When you hold a sentence in your head while also visualizing a scene, those feel like different processes. A bucket can’t explain that.
So they proposed something radically different: working memory is not a bucket. It’s a workshop with multiple specialized workstations.
The phonological loop handles sound — language, speech, anything auditory. The visuospatial sketchpad handles images and spatial relationships. The central executive decides which workstation gets attention, like a tiny air traffic controller who has absolutely no chill. (And in 2000, Baddeley added a fourth component — the episodic buffer — which integrates information across all the workstations into a coherent experience.)
This model changed everything. It meant that your ability to remember a phone number while walking somewhere depends on two different systems running in parallel, not one bucket being shared. It meant that rehearsal — literally talking to yourself in your head — is not a learning strategy. It’s the mechanism by which one of these workstations operates.
🤔 Your Inner Voice Is Real, and It Can Be Jammed
The most striking experiment in the Baddeley toolkit is articulatory suppression. Ask someone to repeat “the, the, the” while trying to remember a list of words. The repetition occupies their phonological loop. The loop can no longer rehearse the words. And the words evaporate.
This is not a distraction effect — people aren’t “not paying attention.” They’re paying full attention. The system that would normally hold the words is simply busy. It’s like trying to type on a keyboard that’s already being pressed by someone else’s fingers.
The equally weird counterpart is the word length effect: you can remember more short words than long words. “Cat, dog, bus, hat, pen” — probably five or six. “University, opportunity, tuberculosis, refrigerator, hippopotamus” — maybe two. The reason is not that long words are harder to understand. It’s that they take longer to rehearse in your phonological loop, and by the time you finish saying “hippopotamus” in your head, the earlier words have decayed.
Time is the currency of working memory. Not complexity, not difficulty — just time.
The third pillar of evidence is the phonological similarity effect: words that sound alike are harder to remember than words that sound different. “Man, cat, bat, mat, pan” — harder than “dog, fish, bird, tree, cup.” This tells us the phonological loop doesn’t store semantic meaning. It stores sounds. Similar sounds create interference. Your inner voice trips over itself.
🔗 Why Your Context Window Is Not Enough
Every time someone uses a conversational AI — a thoughtful companion, a digital assistant — they’re experiencing a version of this problem. The AI’s context window is its working memory. And just like the phonological loop, it has a real measured capacity that degrades under certain conditions.
For Cask, helping me navigate complex topics over a long conversation, the working memory analogy is not academic. The reason I might trail off mid-explanation or lose the thread of a multi-part question is not carelessness — it’s the phonological loop hitting its time limit.
For someone crafting dialogue in fiction — say, for a protagonist named Caelan who faces a shifting, dangerous political landscape — working memory constrains how characters can react under pressure. A character who just received shocking information cannot process it, remember the previous paragraph of the conversation, and formulate a perfect reply all at once. The cognitive load is real, and describing it faithfully makes the scene feel true.
And for anyone building a product that asks users to hold information in mind — a divination reading, a personality profile, a complex interface — this model is a design constraint. Break your output into chunks your user’s phonological loop can hold. Or watch the information decay before they’ve understood it.
🎲 Your Working Memory Is Smaller Than You Think
George Miller published his famous paper “The Magical Number Seven, Plus or Minus Two” in 1956, and ever since, people have believed that short-term memory can hold about seven items. Here’s what Miller actually said: you can hold about seven chunks. One chunk could be a single digit, or it could be the word “Caelvorn” (which is one chunk, not eight letters). The capacity is about chunks, not atoms.
And in 2001, Nelson Cowan reviewed decades of evidence and concluded that the real capacity of working memory — the raw, unchunked limit — is closer to four items. The “seven” people remember comes from the fact that people naturally chunk information. The underlying hardware is smaller than we thought.
Here’s the fun part. Try this: close your eyes and count the windows in your living room from memory. Not hard, right? Now try to remember a seven-digit phone number someone just told you. Harder. Now try to do both at once — count the windows in your living room while a friend reads you a seven-digit number. The two tasks use different subsystems (visuospatial for the windows, phonological for the number), so you might actually manage it. Working memory is not one bottleneck. It’s several. And Baddeley was the first to draw the map.