How Working Memory Works

Baddeley and Hitch proposed in 1974 that working memory is not one store. It is a small system with separate parts.

• A loop for words and sounds. It works like a voice you can hear inside your head.
• A sketchpad for pictures and places.
• A controller that shares your attention between the two.
• A binder, added later, to explain how sights and sounds get joined into a single memory.

It has been challenged, revised and partly replaced several times. It survives in textbooks because you can test it on yourself in ten minutes.

Here is the test. If words and places really use different parts, then one person's word span and place span should be able to come apart. They do.

The loop has two parts. One holds sound for a couple of seconds. The other refreshes it by silently saying it again. That is why you mutter a phone number under your breath, and why an interruption mid-mutter destroys it completely.

Three classic findings pin the loop down.

• Similar-sounding items are harder to hold than different-sounding ones, even when you read them rather than hear them. They collide in a store built for sound.
• Long words are remembered less well than short ones, because saying them to yourself takes real time.
• Repeating a pointless word out loud while you try to remember a list wrecks your score, because it keeps the rehearser busy.

There is no inner voice for 'third from the left, then bottom right'. That is why Corsi block spans usually land one to two units below digit spans in the same person.

That gap is what the model predicts, and you can produce it yourself in five minutes.

Miller's 1956 seven-plus-or-minus-two is about chunks, not items, and a chunk is defined by what you already know.

Cowan argued in 2001 that the real limit is closer to four. Block rehearsing and grouping in the laboratory and four is what you get. That view is now the more widely held one.

Both figures describe the same system from different angles. Four is roughly the raw capacity. Seven is roughly what an adult reaches using the grouping that reading and arithmetic install without asking. That is exactly why a digit span of nine almost always means good grouping rather than a big store. What counts as a good digit span sets that out in numbers.

A span test asks how much you can hold. The n-back task asks something different. It asks how well you can keep swapping what you hold while still making decisions. Every new letter demands an answer and wipes out the item you were rehearsing.

That is why span and n-back scores track each other less closely than people expect, and why researchers treat them as related but separate.

If you score well on one and poorly on the other, nothing is wrong. You have found the seam between holding and updating, which is more interesting than either score alone.

The loop-and-sketchpad picture is a useful teaching model, not a settled account. Two challenges to it are worth knowing, because they change how you should read your own scores.

The first says working memory may not be a set of stores at all. On this view there is only long-term memory plus a spotlight of attention. What looks like a capacity limit is really a limit on how much that spotlight can keep lit at once.

It makes a prediction that turns out to be right. Anything competing for your attention shrinks your measured capacity. That is why taking these tests with a podcast on costs you a unit or two.

The second challenge is about what span tests predict. Engle and colleagues argue that the interesting difference between people is holding on despite interruption, not how much fits.

Their evidence is that harder span tasks predict reading comprehension and reasoning better than plain ones do. A harder span task makes you hold items while a second task keeps butting in.

On that account your forward digit span is the least useful memory number you can collect. That is a slightly awkward thing for a site with a digit span test to admit.

Both views agree on the practical point. A span is a snapshot of a system under load, not a fixed quantity engraved in your head. The load includes everything else competing for your attention at the moment you took the test.

The evidence says two things. Your scores will go up. The underlying capacity mostly will not.

Training on a working memory task reliably improves that task. It helps a little on tasks that look similar. Beyond that the benefits fade. Give the comparison group a real activity rather than nothing, so nobody is simply expecting to improve. Gains in reasoning, comprehension and school results then mostly vanish.

The most cited demonstration is Redick and colleagues in 2013. The broadest review is Simons and colleagues in 2016, which checked the brain-training industry's claims against the evidence and found a wide gap. The Federal Trade Commission reached a similar conclusion about one such company in 2016, at a cost of a two million dollar settlement.

What does help is taking the load off rather than growing the store. Write things down. Group things on purpose. Cut the interruptions. Practise the actual material until parts of it run themselves and stop taking up room.

That advice is unglamorous, it is well supported, and nobody can sell it to you.