Continuum Engine
Part One
The Missing Organs
Your brain has a retrieval algorithm. Psychologists spent forty years reverse-engineering it. Your search bar has about a third of it, and nobody seems to have noticed.
Parking
Where did you park three Tuesdays ago?
Give it an honest try before reading on. Picture the car, the street, the walk back to it. You have no idea.
Now try this morning, and then try where you usually park. Both come back at once, and "usually" arrives even though it is an average over hundreds of individual days you have completely forgotten.
That was a memory experiment, with you as the only subject. You did not search anything, and nothing came back as a list of parking spaces ranked by relevance.
This morning arrived instantly. The habitual one arrived with no particular day attached to it - a memory assembled out of a hundred forgotten ones. And three Tuesdays ago returned nothing, not a guess and not the nearest available Tuesday.
Now try that with a search box
Ask a search index for five results and you get five results, even when all five are wrong. It has no idea what "this morning" means, no idea what "usually" means, and it is structurally incapable of looking you in the eye and saying "I don't know." Those are missing organs.
The missing organs have all been built. Cognitive psychologists worked them out over four decades, fitted the equations against how long real humans take to answer real questions, and published every single one. They are sitting in a literature that almost nobody building AI memory systems reads.
ACT-R
Somebody wrote down how thinking works. Then they ran it.
It is called ACT-R - Adaptive Control of Thought, Rational - which is a name only a university could love. John Anderson and Christian Lebiere began building it at Carnegie Mellon in the 1970s. It is still going.
ACT-R is a scientific claim, written as runnable code.
Give it a task and it does the task, then tells you how many milliseconds a human would have taken - and you go and time actual humans with a stopwatch to see whether it was right.
That means it can be wrong, loudly, in front of everyone. Roughly seventy-three papers in its own bibliography exist to check whether it is. Very little in software is falsifiable that way. Your caching strategy has never had to survive an experiment.
You already own all the parts
| Piece | What it actually is |
|---|---|
| A chunk | One fact, or one memory of something happening. You fetch it by describing it, never by looking up an ID. There is no primary key in your head. |
| A buffer | The one thing you are holding in mind right now. This is why you lose your train of thought the moment somebody says a number at you. |
| A production | An if-this-then-that rule. Exactly one fires at a time. Thinking, in this model, is a very fast sequence of these. |
| Activation | A single number on every memory: how easy it is to reach right now. The rest of this series is about this number. |
Picture a very deep swimming pool
Every memory you have is an object floating somewhere in it. The things you use constantly bob near the surface where you can grab them without looking. The things you have not touched in years have drifted down into the dark.
Activation is depth. Nothing is ever taken out of the pool. Given a strong enough pull, something from the very bottom can come straight back up.
Your parking spaces are in there too. This morning's is bobbing at the surface. Three Tuesdays ago is somewhere near the bottom, intact, and far too deep for a casual reach to find.
There is also a line painted down the side of the pool, some way below the surface. Reach for a memory sitting under that line and your hand comes back empty, which is how three Tuesdays ago produced nothing at all instead of a guess.
Facts and skills live in different places
Knowing that Paris is the capital of France, and knowing how to ride a bike, are stored in separate systems with separate rules. Facts get easier to reach the more you use them. Skills get stronger the more they work.
That is why you can forget a fact you learned yesterday and still ride a bike you have not touched in fifteen years.
Remembering is a dice roll
A memory can be a perfect match for what you asked and still not come back, because it sat too deep or the dice went against it.
The theory claims exactly that. Memory is noisy, it fades, and everything competes with everything else. The architecture's job is to say how noisy, how fast, and how competitive.
Why Trust Any Of This
The theory can be checked with a stopwatch.
Psychology has had a rough decade. Plenty of famous results turned out not to replicate, which is a fair reason to hesitate before building software on top of a theory about brains.
This particular corner of it gets checked with a stopwatch. What is being timed is the gap between asking somebody a question and hearing the answer, which the literature calls reaction time.
Name two capitals and feel the gap
Name the capital of France. Then name the capital of Australia.
Paris arrives before you have finished reading the question. Canberra takes a beat, and for plenty of people Sydney surfaces first and has to be sent back down - a near miss, which part three explains.
That beat is the swimming pool seen from outside. Nobody can look into your head and measure how deep Canberra sits, but anybody with a stopwatch can time how long it takes to come up.
The model's predictions take the form "a person will take 847 milliseconds to produce it" - a little under a second. Then somebody sits a hundred undergraduates down, measures them, and compares the two numbers.
Forty years of that, across hundreds of experiments, is what turns these equations from a nice metaphor into something with actual load-bearing capacity. When we get to the constant that governs how fast memories fade, it will be 0.5, and the reason that number is worth anything is that thousands of stopwatches say so. A decay constant an engineer picked on a Thursday has no such defence.
What the other four parts cover
Four mechanisms, a rival theory, and then what happens when you try to build the thing.
- 2 Fading. Why memories sink at exactly the rate they do, and why that rate turns out to be a bet about the future rather than a failure of storage.
- 3 Cueing. How the things already in your head drag other things up with them - and why that makes knowing more about a subject actively unhelpful.
- 4 The argument. A rival theory that made different choices everywhere, and the three guardrails both of them ended up building anyway.
- 5 The build. Every dial in the real thing, what measuring it changed, and what is still unbuilt.
Next
Part 2 · Forgetting is a forecast
Two researchers went and counted how often words appear in newspaper headlines. What they found says your memory is running a betting operation, and it is unnervingly good at it.