What Brings a Memory Back

Too Much Information

Knowing more about something makes it harder to remember.

The experiment behind that sentence has been replicated for fifty years.

Volunteers memorise a pile of simple sentences. The doctor is in the park. The lawyer is in the bank. The doctor is in the church. Later they are shown one sentence and asked: did you study this?

People are reliably slower to recognise "the doctor is in the park" when doctor turned up in lots of the sentences than when it turned up in only one.

It is called the fan effect, after the diagram: draw doctor in the middle of a page with a line to every sentence it appears in, and the lines spread out like a fan.

Shout a name across a crowded room

Shout "Priya!" and one head turns.

Shout "hey!" and nobody turns, because nobody can tell it was meant for them.

Whatever is currently on your mind is calling out, constantly, to everything it is connected to. Back in the swimming pool from part one, every memory that hears its name rises a little way towards the surface. That is why one thought drags the next one along behind it.

ACT-R calls this spreading activation. It sits on top of the history from part two: history sets how deep something has sunk, and the shout lifts it from there.

But a cue has one voice, and it gets shared out. The number of things a cue is attached to is its fan, and the wider the fan, the less any single one of them hears. Doctor attached to one sentence is a name shouted across a room. Doctor attached to twelve is "hey!"

You only get one spotlight

Thinking about more things spreads the same spotlight thinner. That is why rambling at somebody rarely helps them remember, and why a longer, vaguer question is not a more informative one.

Search engines use the same penalty

Information retrieval has the same idea under a different name, inverse document frequency: common words count for less than rare ones, because a word that appears everywhere points at nothing in particular. Two fields with no contact between them landed on the same penalty.

The setting does not travel

Psychology tuned this on words linked to two or three other words. At that setting, spreading dies completely once something has seven connections. A chapter in a document has fifty, which is a voice tuned for a dinner party being asked to fill a stadium. Borrow the constant without checking and your spreading silently does nothing at all.

Near Misses

There is a right way to be wrong.

Ask somebody for 7×8 in a hurry and you will sometimes get 49, but not 23 or 812. You get the answer to the neighbouring question, because 7×7 was sitting right there and was very nearly close enough.

It is the same reason you occasionally leave the house with your neighbour's keys and never with a banana. Human errors have a shape, and the shape is "adjacent."

Run it on yourself

Say "silk" out loud five times. Then, without stopping to think: what do cows drink?

Most people say milk. Cows drink water. Five rounds of silk had it calling out to every word that rhymes with it, cows was calling out to everything dairy, and the one word both of them were shouting at came up before the right answer did.

ACT-R models this on purpose, and calls it partial matching. A memory that fits the question slightly wrong is fined for being different, and the fine grows with the difference.

In the pool, the fine is a weight tied on before the memory is pulled up. A near miss gets a pebble and can still beat a correct answer that has sunk deep. A banana gets an anchor. Every number involved is negative, and the manual specifically warns you off making them positive, so nothing ever gets a bonus.

A system that fails by handing you something adjacent is a system you can debug.

The Bouncer

The most important thing a memory can do is not show up.

Everything so far produces a number. The final step is the one that makes this recall rather than search: there is a bar, and anything below it does not come back at all.

Back in the pool, the bar is how far down your arm can reach. Anything above it you can grab. Anything below it stays in the dark, however perfectly it matches what you were looking for, unless a loud enough shout lifts it within reach. ACT-R calls this the retrieval threshold.

Two names to try for

Try to name your teacher from when you were seven. Then try to name the child who sat behind you in that class.

For most people the first is either there or on the tip of the tongue: fingertips brushing something just under the bar, close enough to feel its shape and maybe its first letter. The second is a total blank, with no shape at all, because it is nowhere near. Those are different states and your brain does both.

Stacked on top there is noise, which in the pool is the water never being quite still. There is a small random wobble on every attempt, so the same memory sits a little higher on one try and a little lower on the next. There is also a fixed per-memory wobble set when the memory first formed. Some memories simply, and permanently, float a bit higher than their history says they should - which is both true of people and mildly insulting.

The same equations also predict how long remembering takes. And they predict that someone with a higher bar gives up faster.

On a failed attempt your arm stays in the water for as long as a memory sitting exactly at the bar would take to reach it, then comes out empty. Raise the bar and that wait gets shorter.

Nobody designed that in. It falls out of the maths, and it matches what people do: we abandon hopeless recall attempts quickly.

This is also the stopwatch from part one doing its work. A theory that predicts how long something takes can be checked against a hundred undergraduates and a timer - which is why these particular equations survived forty years of people trying to break them.

ACT-R turns all four off by default

All four mechanisms in this series - the fading, the spreading, the near-miss penalty, the noise - are switched off by default in ACT-R. Install it, run it, and none of this is happening.

The theory declines to assume they always apply. So anybody shipping all four switched on has made four decisions the science itself refused to make for them. That is fine, as long as it is on purpose.

We got this wrong, and only found out by reading the actual manual instead of the papers citing it. That story is in part five.

The other serious theory of thinking made different calls at nearly every fork. Its author is also the one who drew the line against anything editing its own memory scores.

The Rival Theory →