Desmond: In eighteen eighty-five, a German psychologist spent seven months memorizing nonsense syllables at himself. Alone. Metronome-paced.
Nora: That man is the reason every memory-decay chart you've ever seen exists.
Desmond: One subject. Him. And the whole field is still standing on it.
Nora: Today we find out what he actually measured, what a very famous integer never actually meant, and why almost none of it survives contact with the paper.
Nora: Welcome to The Seams Are The Point. I'm Nora.
Desmond: I'm Desmond. This is episode one of five, and it's the origin story — Ebbinghaus, Miller, and why capacity was never the number you think it is.
Nora: Five episodes, building toward one question: what does human memory science actually tell an engineer building memory for an agent?
Desmond: Today's answer starts with a man who found Byron too interesting to forget properly.
Nora: Set the scene. Eighteen eighty-five, Hermann Ebbinghaus, twenty-nine years old.
Desmond: He tried tones first. Too cumbersome. He tried digits. Unsuitable. Then he tried stanzas of Byron's Don Juan.
Nora: And rejected those too.
Desmond: Because the meanings varied too much. Byron kept meaning things to him, and that was contaminating the data.
Nora: So the founding curve of memory science exists because a man found Byron too interesting to forget on command.
Desmond: He landed on consonant-vowel-consonant nonsense syllables. Three sessions a day, roughly seven months, metronome-paced, one subject: himself.
Nora: Now here's the part I got wrong before we started researching this. What does the curve actually plot?
Desmond: Not recall. Savings. How much cheaper relearning was, compared to learning it fresh.
Nora: So his twenty-four hour figure, thirty-three point seven percent —
Desmond: — is thirty-three point seven percent savings. Not 'you forgot seventy percent of it overnight.' It's 'a day later, relearning cost a third less.'
Nora: That's a statement about residue. Not absence.
Desmond: Recall failure was never the same thing as the memory being gone. That distinction is baked into the instrument from day one, and the corporate-training slide decks miss it every time.
The Record: The entire foundation of forgetting research rests on two subjects, a century and a third apart.
The Record: The nineteen eighty-five curve was replicated in two thousand fifteen by Jaap Murre and Joeri Dros — also n of one. Dros, a twenty-two-year-old Dutch student, spent about seventy hours relearning.
The Record: The shape held. But at thirty-one days, his savings measured four point one percent against Ebbinghaus's twenty-one point one. A fivefold gap between the only two people who have ever properly run this experiment.
Nora: Jump forward. Nineteen fifty-six, George Miller, Psychological Review.
Desmond: He opens the paper by saying he's been persecuted by an integer.
Nora: The magic number seven.
Desmond: Which almost nobody in the industry read correctly. He was not describing working memory. The construct didn't exist yet.
Nora: Then what was he doing?
Desmond: He had two sevens from two completely unrelated literatures. Absolute judgment of a single stimulus dimension topped out around two point six bits. Immediate memory span ran about seven items.
Nora: Two different sevens.
Desmond: His actual thesis was the dissociation between them. Absolute judgment is constant in bits. Span is constant in chunks — regardless of how much information each chunk holds.
Nora: And the shared number?
Desmond: He calls it a coincidence and moves on. He never defines 'plus or minus two' anywhere in the paper.
Nora: So the title everyone remembers is the joke, and the finding everyone forgot is the argument.
Desmond: Psychology spent seventy years building on the part he was joking about.
The Record: Go and read 'The Magical Number Seven, Plus or Minus Two' in full.
The Record: The phrase in the title never appears as a defined quantity in the body of the paper. Seventy years of citation rest on a number Miller dismissed as coincidental.
Nora: This is where it gets useful for us. Cowan revised the seven down to about four.
Desmond: And gets misread the exact same way Miller did. Cowan's four is a lower bound, measured under conditions designed to block chunking and rehearsal — the strategies people actually use in real tasks.
Nora: So it's a bottleneck estimate.
Desmond: Not an operating capacity. And even the container framing is contested.
Desmond: Ma, Husain and Bays, two thousand fourteen, argue capacity is a shared resource spread across items, not fixed slots.
Nora: So adding items degrades the fidelity of everything, rather than dropping items outright.
Nora: None of these are numbers we can put in a config file.
Desmond: Correct. But there's one claim underneath all of it that is precise, and it's the one worth keeping.
Nora: Say it.
Desmond: Capacity is invariant to the information content per unit. A context slot holding a fully-resolved, pre-joined record costs the same as one holding a bare id.
Nora: That's the first design rule. And notice — it arrives with no number attached.
Desmond: Which is exactly why it survives where the numbers didn't.
Nora: Nineteen seventy-four. Baddeley and Hitch. You said this is the experiment that started the modern field.
Desmond: For the wrong reason, usually. Subjects had to verify statements about letter order while holding up to eight spoken digits at the same time.
Nora: If short-term memory were one unitary store, that should have been catastrophic.
Desmond: It wasn't. Response time rose about fifty percent at the heaviest load.
Nora: And error rate?
Desmond: Sat at roughly five percent. Refused to move. The model survived by failing to break.
Nora: They nearly didn't publish that.
Desmond: Baddeley said decades later it was still incomplete. But look at what it hands us — load degrades latency and place-keeping long before it degrades accuracy.
Nora: That's a diagnostic. If an agent gets slower and starts losing its place, that's not the same failure as an agent giving you a wrong answer confidently.
Desmond: Altmann, Trafton and Hambrick, two thousand fourteen, sharpen that further.
Desmond: Brief interruptions specifically inflate sequence errors — wrong step, repeated step, skipped step — while leaving the quality of work within a step untouched.
Nora: So you get a free error taxonomy out of this.
Desmond: Wrong content inside a step, versus lost its place in the procedure. Only the second one should respond to context-management changes.
Nora: Before we get to filtering — you keep saying the cost driver isn't item count.
Desmond: It's element interactivity. Sweller's Cognitive Load Theory. Ten independent lookup facts are nearly free. Four mutually-constraining ones are expensive.
Nora: Budget on an interactivity score, not on tokens.
Desmond: Right. And Sweller's own two thousand nineteen retrospective walks back two things people still get wrong. First, germane load — it used to be treated as a third kind of load that refills whatever space you free up.
Nora: And it doesn't?
Desmond: If it did, total load would stay constant no matter what you did. It doesn't stay constant.
Nora: So kill any policy that says 'we freed some tokens, let's spend them on enrichment.'
Desmond: Freed capacity isn't a pool you're obligated to refill.
Nora: What's the second correction?
Desmond: Measurement is unsettled. The field's central variable is mostly assessed by asking people how hard something felt.
Desmond: So don't build a cognitive-load score for an agent and tune against it — budget on structural proxies you can actually compute.
Nora: Now — filtering.
Desmond: Vogel, McCollough and Machizawa, two thousand five, Nature. The most useful counterintuitive result in the whole literature.
Nora: Go.
Desmond: Low-capacity individuals can hold more in working memory than high-capacity ones. They encode the irrelevant items too.
Nora: So the deficit isn't a smaller container.
Desmond: It's a leakier admission policy. Spend the budget on the gate, not on enlarging the window.
The Record: A scope note the popular retelling drops. The Vogel finding is a conditional inference, drawn from an EEG amplitude measure in visual working memory.
The Record: The original paper hedges accordingly. The industry retelling does not.
Nora: Last one, and it's the strangest. Expertise reversal. Kalyuga, Ayres, Chandler and Sweller, two thousand three.
Desmond: Worked examples help novices. Measurably help. And they make experts measurably worse.
Nora: That's not a diminishing return.
Desmond: It's a sign flip. Guidance that duplicates what you already know is just redundant material you now have to process anyway.
Nora: So any always-injected SOP, any fixed few-shot block, any static memory preamble —
Desmond: — is net-negative on exactly the tasks where your agent is already strong. Untested on agents, as far as we can find. Directly testable, though.
Nora: Let's land where we started. Ebbinghaus measured savings, not recall — the memory wasn't gone.
Nora: Miller's seven was never a working-memory limit, and he never even defines the plus-or-minus-two.
Nora: The one claim that actually transfers is that capacity ignores information density, and the one lever that actually matters is filtering, not window size.
Desmond: And expertise reversal means the fix for a weak agent can be the exact thing breaking a strong one.
Nora: None of that required deletion. Which is the wrong instinct entirely — and it's next.
Desmond: Episode two. There's no delete primitive in human memory. Every mechanism that looks like erasing something turns out, on inspection, to be a change in accessibility.
Nora: See you there.