Procedural memory and muscle memory

Try it

Can you feel a skill forming?#

This is mirror drawing, the task behind the first evidence for procedural memory. You trace a star with up and down swapped. Your own times show up at the end, and if you come back tomorrow it compares them.

It’s a rough version. Milner’s task used a pencil and a real mirror, with the star’s double outline on paper, and this uses a mouse, a finger or arrow keys, with the pen moving by how far you move. It’s a demonstration, and one run can’t say anything about your memory. Your last ten runs (times, out-of-band counts and the input used) stay in this browser until you clear them, and the demo reports only that it was started.

If you can’t use it: the task depends on seeing the pen and moving it, and the keyboard route is there for anyone who can’t drag. The next section describes what Milner found.

How was it found?#

H.M. was a man with profound memory impairment after brain surgery in 1953. In 1962 Milner showed him a five-pointed star with a double outline and asked him to trace it while seeing his hand and the star only in a mirror. He acquired the skill during ten trials and retained it well across three days, and at the end he had no recollection of having done the task. Squire calls this “the first hint” that there was more than one kind of memory in the brain.

For a time, motor skills looked like a special case. Then in 1980 Cohen and Squire found that memory-impaired patients learned to read mirror-reversed words at a normal rate, with poor memory for the task and the words. So skills that aren’t about movement belonged too. They drew the distinction between declarative and procedural knowledge, which Simply Psychology traces in everyday terms to Gilbert Ryle’s “knowing how” and “knowing that” in 1949.

The list kept growing, so the broader word became nondeclarative: an umbrella for skills and habits, priming, simple conditioning and more.

How does a skill get built?#

The classic account is Fitts and Posner’s, from 1967, which we know through Simply Psychology. It has three stages:

1. Cognitive: you follow instructions, slowly and with errors, and need attention for almost every part.
2. Associative: errors drop, and the movements start to run together.
3. Autonomous: it takes little monitoring and stands up to doing something else at once.

So a skill usually doesn’t start as procedural memory. Your first mirror round is mostly the cognitive stage, with you working out each move. Diedrichsen and Kornysheva describe the change as going from the effortful selection of single movements to producing them together fast and accurately, with premotor areas coming to encode short chunks.

Early gains seem to come in the pauses

Bönstrup and colleagues had 27 people tap a five-key sequence (4-1-3-2-4) with the non-dominant hand, 10 seconds on and 10 seconds off, and looked at speed inside each stretch. Early on, speed rose by 2.37 key presses a second overall. Changes during practice came to about nothing (−0.32), and the gains across the rest periods came to 2.69. They read that as a rapid form of consolidation. They allow that recovery from low-level fatigue could conceivably explain the gains, and they cite earlier work that reads such gains as consolidation. The overnight gain, 0.73, didn’t correlate with those early gains.

A follow-up by Buch and colleagues recorded brain activity from outside the head, by magnetoencephalography, during those rests. They report fast replay of the practised sequence, about 20 times faster than the skill, which predicted how much was consolidated, and which they locate in the hippocampus and entorhinal cortex as well as sensorimotor cortex. That location is inferred from the recording, and they call replay a “possible” binding mechanism.

What about sleep?

Key-sequence tasks were long the standard example of sleep improving a skill. Brawn and colleagues found that with morning training, performance fell over the waking day and recovered after sleep, and that with evening training it stayed stable across sleep and the next day. Simply Psychology reports a pooled analysis by Schmid and colleagues of 48 studies, with what it calls a small overall advantage for sleep (g = 0.43) and a medium one for mirror tracing (0.62).

Where in the brain is it?#

• Basal ganglia. The Queensland Brain Institute describes them as particularly involved in coordinating sequences of movement. Squire and Dede call the dorsolateral striatum “crucial for the development of habits”, and report that in mice it became more engaged as performance became automatic.
• Cerebellum. Krakauer and Shadmehr review evidence for it in two kinds of associative motor learning, eyelid conditioning and the eye reflex that adapts to head movement.
• Motor cortex. The same review points to primary motor cortex for skilled finger movements. Kawai and colleagues complicate this in rats. Lesions after training left a learned sequence intact, and lesions before training stopped it being learned, so they suggest a role in “tutoring” circuits below it.

Simple sequences in rats are a long way from a violin, and how far that result carries to people isn’t something these papers settle.

Is there any memory in the muscles?#

For skills, the evidence above points to the nervous system. But there’s a separate finding in muscle tissue, and physiologists call it muscle memory too. It’s one proposed explanation for why muscle can come back more readily than it was first built.

Seaborne and colleagues had eight previously untrained men do 7 weeks of resistance exercise, 7 weeks off, then 7 weeks back on. Muscle mass returned to baseline during the break, yet some genes that lost methylation, a chemical mark on DNA, in training stayed that way. Across the genome, about twice as many sites had lost it after retraining as after the first training, 18,816 against 9,153, from eight men at the first count and seven at the second. It was a small study, and it looked at the thigh. Earlier work in mice given testosterone points to a second candidate: nuclei added to muscle fibres during growth, kept after muscle mass returns to normal.

How is it different from other kinds of memory?#

Episodic and semantic memory are the two kinds of declarative memory, and working memory has its own page.

What do people get wrong about it?#

That you can’t put it into words or bring it to mind

Simply Psychology’s summary defines it as skills “that a person cannot bring to mind or explain”. Much of a skill is like that, and the same page adds that a skill starts as “effortful, conscious knowledge” before it becomes procedural memory. Yet the line isn’t that clean once it is one. The field’s standard key-sequence task is one you’re told in advance, and Brawn and colleagues call it “explicit” motor-sequence learning. Squire and Dede’s wording is safer: abilities expressed through performance “without requiring conscious memory content”.

That it has nothing to do with the hippocampus

Simply Psychology says the striatum and cerebellum handle procedural memory “while the hippocampus handles declarative memory”, the classic picture from H.M., whose skill learning, Squire says, suggested that motor skills “must lie outside the province of the medial temporal lobe”. Buch and colleagues now describe “hippocampo-neocortical replay” during rests in a sequence task, attributed to hippocampal sources from a recording outside the head. Their task is one people are told outright, so in our reading declarative knowledge may explain some of that. Squire and Dede, discussing imaging reports of medial temporal activity in unaware sequence learning, add that fMRI “cannot establish a necessary role”. Activity there is a different claim from being needed.

That sleep is what makes skills stick

Simply Psychology’s summary says REM sleep following slow-wave sleep “significantly enhances” procedural consolidation, and it reports the more modest pooled result further down. Brawn and colleagues do argue for sleep-dependent consolidation, and they add that the classic pattern depends on when training happens. Bönstrup’s gains in 10-second rests show speed rising in the same session, awake, which is a separate question from what lasts. The papers we read on sleep don’t single out one stage, and we didn’t read the older studies behind that line.

That you never forget how to ride a bike

The saying is everywhere, and even a research review opens with “we apparently never forget how to swim, ride a bicycle or drive a car” (Krakauer and Shadmehr), with the “apparently” doing some work. The classic number is Hill, Rejall and Thorndike’s, through Wikipedia: savings in relearning typing after 25 years without practice. Savings means it came back quicker than it was first learned, which implies some of it had gone. Julius and Adi-Japha also note a loss from the last trial of one session to the first of the next in mirror drawing, back to Snoddy in 1926, so a slower start on your second day doesn’t mean nothing was kept.

Do AI systems have procedural memory?#

The term is in use, with a different meaning. The CoALA framework gives language agents two forms of procedural memory: “implicit knowledge stored in the LLM weights, and explicit knowledge written in the agent’s code”. The weights are a fair parallel to knowing how, since a model can do far more than it can explain. CoALA calls fine-tuning them “a costly form of learning”, though, done on a schedule, so a model in use doesn’t normally improve by repeating a task. We’d put an agent’s written code and instructions closer to knowing that.

Counterparts, the memory system whose site this is, keeps what the AI writes down in its own words, including notes on how it likes to do things. Those are written descriptions and don’t change the model’s weights. The Field Guide follows the mechanisms AI systems do borrow, and the directory compares the systems.

Questions people ask#

Questions about how any of this carries over to AI? Ask in r/AIMemorySystems.