Procedural Memory

Abstract

Procedural memory is the nondeclarative system that stores skills, habits, and learned sequences of action, expressed through improved performance rather than conscious recollection. It is the memory of knowing how, dissociated from the declarative memory of knowing that. The founding evidence came from amnesic patients who acquired motor and perceptual skills normally across days while retaining no memory of the practice, showing that skill learning survives medial-temporal damage. Procedural memory depends on the basal ganglia, cerebellum, and motor cortex rather than the hippocampus, is acquired gradually through repetition, and becomes rapid, automatic, and resistant to forgetting. This article sets out its definition, the declarative dissociation, the stages and neural circuitry of skill acquisition, three interactive demonstrations, and current work on consolidation and the skill-adaptation distinction.

Keywords: procedural memory, skill learning, nondeclarative memory

Procedural memory is the form of long-term memory that stores how to perform skills and habits: riding a bicycle, touch-typing, reading mirror-reversed text, or executing a well-practised sequence of movements. It is a type of nondeclarative memory, the broad category of long-term memory that operates outside conscious awareness and is expressed through changed performance rather than deliberate recollection. Its defining contrast is with declarative memory, the system for facts and events that can be consciously brought to mind and stated. Procedural memory is memory for knowing how; declarative memory is memory for knowing that (Cohen & Squire, 1980).

The distinction is not merely descriptive. Skills and facts dissociate in the brain, so that damage abolishing the capacity to form new declarative memories can leave the capacity to acquire new skills entirely intact. This is the phenomenon that gives procedural memory its status as a separate memory system rather than a subtype of a single general-purpose store, and it is the through-line of this article.

Knowing How Versus Knowing That

The clearest evidence for procedural memory as an independent system comes from dense amnesia. The patient H.M., who underwent bilateral medial-temporal-lobe resection to control epilepsy, could form almost no new declarative memories afterward, yet Milner showed that across three days of mirror-drawing practice his tracing error fell steadily, the normal learning curve, even though at the start of each session he had no memory of ever having done the task (Milner, Corkin, & Teuber, 1968). Corkin extended the finding to other perceptual-motor tasks, including rotary pursuit and bimanual tracking, confirming that motor-skill acquisition was preserved despite the profound declarative deficit (Corkin, 1968).

Cohen and Squire generalised the dissociation beyond motor skill. Amnesic patients learned the cognitive skill of reading mirror-reversed word triplets at a normal rate and retained it for months, while remaining severely impaired at recognising the specific words they had read, the declarative record of the training (Cohen & Squire, 1980). The pattern, they argued, marks a fundamental division between a system for knowing how, spared in amnesia, and a system for knowing that, lost. Squire later placed procedural memory within a formal taxonomy: declarative memory (facts and events, dependent on the medial temporal lobe) versus nondeclarative memory, an umbrella spanning procedural skills and habits, priming, simple classical conditioning, and nonassociative learning, each with its own neural substrate (Squire & Zola, 1996; Squire, 2004).

Priming, the facilitation of processing a stimulus by prior exposure to it, is a distinct nondeclarative form that also survives amnesia. Graf and Schacter showed that amnesic patients completed word stems with recently studied words at above-baseline rates, implicit memory for new associations, despite failing explicit recognition of the same items (Graf & Schacter, 1985). Procedural memory is thus one branch of a wider nondeclarative family, unified by expression without awareness and by independence from the hippocampal system.

Figure 1

Procedural Memory Within the Taxonomy of Long-Term Memory

The Stages of Skill Acquisition

Skills are not learned all at once. Fitts and Posner's influential description, adopted across the skill-learning literature, divides acquisition into three phases: a cognitive stage, in which performance is slow, effortful, and heavily dependent on declarative instructions held in mind; an associative stage, in which errors are gradually eliminated and actions are chained; and an autonomous stage, in which the skill runs off rapidly, accurately, and with little attentional cost. Anderson formalised this progression in the ACT theory of skill acquisition as a shift from declarative to procedural knowledge, a process he termed knowledge compilation, in which slow interpretation of remembered rules is replaced by fast, direct production rules that fire automatically (Anderson, 1982). Willingham later proposed a neuropsychological theory that decomposes a motor skill further, separating its perceptual-motor, cognitive, and conscious components onto distinct control systems that can be learned and impaired independently (Willingham, 1998).

Improvement with practice is lawful. Across a wide range of tasks the time to perform a skill falls as a power function of the number of practice trials, the power law of practice: large early gains that diminish steadily, so that performance improves without ever quite stopping (Anderson, 1982). This regularity, and the transition from controlled to automatic processing it accompanies, are what the Worked Example below quantifies.

A signature of automaticity is that a well-learned skill can be expressed without the learner being able to articulate it. In the serial reaction time task, participants respond to a repeating spatial sequence and speed up on it relative to random trials, demonstrating that they have learned the sequence, while frequently being unable to report that any sequence existed (Nissen & Bullemer, 1987). The task became the standard laboratory assay of implicit sequence learning, though the degree to which its learning is truly free of awareness remains debated (Robertson, 2007).

The Basal Ganglia, Cerebellum, and Motor Cortex

Procedural memory is not stored where declarative memory is. Its critical structures are the basal ganglia (in particular the striatum), the cerebellum, and the primary motor cortex, a network largely separate from the hippocampus and medial temporal lobe. Knowlton, Mangels, and Squire demonstrated a neostriatal habit-learning system in humans directly: on a probabilistic classification task, amnesic patients with hippocampal damage learned the cue-outcome associations normally over early trials, whereas patients with Parkinson's disease, who have striatal dopamine loss, were impaired, the exact reverse of the pattern on a declarative memory test of the same material (Knowlton, Mangels, & Squire, 1996). Habit learning and declarative memory draw on dissociable brain systems.

Packard and Knowlton set this in a broader framework in which the basal ganglia support a stimulus-response habit system that operates in parallel with, and sometimes in competition with, the hippocampal system for flexible declarative memory (Packard & Knowlton, 2002). Poldrack and colleagues made the competition visible with neuroimaging: during classification learning, hippocampal and striatal activity were negatively correlated, activity in one predicting suppression in the other, so that the two systems can trade off in governing the same behaviour (Poldrack et al., 2001).

The motor components of skill are laid down partly in the cortex itself. Karni and colleagues used functional MRI to show that several weeks of practice on a finger-movement sequence enlarged the primary-motor-cortex representation activated by the trained sequence, evidence of practice-driven plasticity in the adult motor cortex (Karni et al., 1995). Doyon and Benali synthesised the human imaging evidence into a model in which motor-sequence and motor-adaptation learning depend on partly distinct cortico-striatal and cortico-cerebellar loops, and in which the balance of activity shifts across the course of learning from associative to sensorimotor territory as the skill consolidates (Doyon & Benali, 2005).

Two Systems, Two Learning Curves

The interactive demonstration below makes the amnesic dissociation concrete. It contrasts a declarative measure (recognising the specific items encountered during training) with a procedural measure (speed or accuracy of the skill itself) across successive practice sessions, for a simulated healthy learner and a simulated amnesic learner. The skill curve rises for both; the recognition curve rises only for the intact learner. The two measures moving apart is the dissociation that defines procedural memory.

Learning a Hidden Sequence

The third demonstration reconstructs the serial reaction time task. A target appears at one of four positions and the learner responds; unknown to them, the positions follow a fixed repeating sequence for most blocks, with an occasional random block inserted. Reaction time falls across the repeating blocks as the sequence is learned, then jumps back up on the random block, the rebound revealing that what was learned was the sequence specifically, not merely general speed-up. This is procedural sequence knowledge expressed in performance without necessarily reaching awareness.

Forms of Procedural Learning

Procedural memory is not a single skill but a family of learning types that share expression through performance and independence from the declarative system. Table 1 summarises the principal forms and their characteristic tasks and substrates.

Worked Example: The Power Law of Practice

Consider a learner whose time to complete a skill on trial N follows the power law of practice, T(N) = a × N^-b, with an initial single-trial time of a = 2000 ms and a learning-rate exponent b = 0.3. The formula captures the diminishing-returns shape of nearly all skill curves (Anderson, 1982).

On the first trial, T(1) = 2000 × 1^-0.3 = 2000 ms. On the tenth trial, T(10) = 2000 × 10^-0.3. Since 10^-0.3 = 0.501, T(10) = 2000 × 0.501 = 1002 ms, an improvement of 998 ms, nearly halving the time. On the hundredth trial, T(100) = 2000 × 100^-0.3; since 100^-0.3 = 0.251, T(100) = 2000 × 0.251 = 502 ms.

The pattern is the essence of skill learning: the improvement from trial 1 to trial 10 (998 ms) is far larger than the improvement from trial 10 to trial 100 (500 ms), even though the second interval spans ninety trials and the first spans nine. Each tenfold increase in practice yields the same proportional gain, a constant factor of 10^-0.3 = 0.501, so performance keeps improving but ever more slowly. Plotted on log-log axes, T(N) becomes a straight line of slope -b, which is how the power law is usually identified. The first demonstration above lets the exponent b be varied to see how the rate of skill acquisition changes.

The power-law form is not beyond dispute. Heathcote, Brown, and Mewhort argued that the apparent power law is largely an artefact of averaging over many learners: fit to individual practice curves, an exponential function, in which performance improves by a constant proportion per trial rather than a constant proportion per tenfold increase in practice, describes the data better (Heathcote, Brown, & Mewhort, 2000). The functional form of the learning curve therefore remains contested, even though the qualitative fact of large, diminishing gains is not.

Discussion

Procedural memory reframed what memory is. Before the amnesia evidence, memory was often treated as a single faculty that damage would weaken uniformly. The demonstration that H.M. and other amnesic patients learned skills at a normal rate while forming no memory of the learning forced a division into multiple systems with separate neural bases, of which the declarative-nondeclarative split is the broadest (Squire, 2004). Procedural memory sits on the nondeclarative side: acquired gradually, expressed in action, inaccessible to introspection, and spared by the very lesions that devastate memory for facts and events.

The parallel-systems view carries a functional logic. A flexible declarative system that can be updated in a single trial is well suited to remembering unique episodes; a slow procedural system that changes only with repetition is well suited to extracting the stable regularities on which reliable skill depends. That the two can compete (Poldrack et al., 2001) suggests they are not merely different filing cabinets but genuinely alternative strategies for controlling behaviour, with the balance between them shifting as a task moves from novel to routine. The clinical mirror image, spared skill learning in hippocampal amnesia and impaired habit learning in basal-ganglia disorders such as Parkinson's disease, remains the strongest evidence that the systems are anatomically real rather than a convenience of description (Knowlton, Mangels, & Squire, 1996).

Current Directions

Much recent work concerns how a newly practised skill is stabilised after practice ends, the problem of consolidation. Doyon and colleagues have refined the account of motor-sequence learning into fast and slow stages served by shifting cortico-striatal and cortico-cerebellar contributions, and have highlighted unresolved questions about how sleep and wakefulness each contribute (Doyon et al., 2018). Krakauer and colleagues have pressed a conceptual distinction that the older literature blurred: motor adaptation, the recalibration of movement to a perturbation, is not the same as the acquisition of genuinely new skill, and the two have different time courses and neural signatures (Krakauer et al., 2019).

The timescale of consolidation has been revised sharply downward. Bonstrup and colleagues reported that much of the gain in an explicit motor-sequence task accrues during the short rest periods between practice blocks rather than during practice itself, a rapid form of offline consolidation occurring within minutes (Bonstrup et al., 2019). Buch and colleagues linked this waking consolidation to hippocampo-neocortical neural replay, implicating the declarative system in the stabilisation of a procedural skill and complicating any strict separation of the two (Buch et al., 2021). Sleep remains central to the longer-term picture: Klinzing, Niethard, and Born reviewed the mechanisms by which sleep supports systems-level consolidation, including the active reactivation of newly encoded memories (Klinzing, Niethard, & Born, 2019). At the representational level, Yokoi and Diedrichsen used pattern analysis to show that practised action sequences are encoded hierarchically in the human neocortex, with separable representations of individual movements and of the chunks that bind them (Yokoi & Diedrichsen, 2019). Common mechanisms may span domains: Censor, Sagi, and Cohen argued that perceptual and motor learning share consolidation and reconsolidation processes rather than being wholly distinct (Censor, Sagi, & Cohen, 2012).

Common Misconceptions

Procedural memory is just muscle memory.

Motor skill is only one branch. Procedural memory also covers perceptual skills such as mirror-reading, cognitive skills, sequence learning, and stimulus-response habits, none of which are stored in muscle. The term muscle memory is a popular label for what is in fact learning in the brain's motor and striatal circuits (Squire & Zola, 1996).

Procedural memory is a lesser or more primitive form of memory.

It is not a degraded version of declarative memory but a distinct system with its own advantages: it is durable, resistant to forgetting, and supports the automatic, high-speed performance that deliberate recollection could never achieve (Cohen & Squire, 1980).

Skills are learned consciously and only later become automatic memories.

Some skills are acquired with no awareness of what is being learned at all. In the serial reaction time task, learners speed up on a hidden sequence they cannot report, showing that procedural learning does not require, and is not merely the residue of, conscious knowledge (Nissen & Bullemer, 1987).

Amnesia wipes out all memory equally.

Dense amnesia from medial-temporal damage spares procedural learning almost entirely. Patients who cannot remember a training session at all still improve on the trained skill across sessions, the founding observation of the field (Milner, Corkin, & Teuber, 1968).

Glossary

ACT theory.

Anderson's theory of cognitive architecture in which skill acquisition is a shift from declarative rules to compiled procedural production rules.

Associative stage.

The middle phase of skill acquisition, in which errors are gradually eliminated and component actions are chained together.

Automaticity.

The end-state of skill learning in which performance is fast, accurate, and requires little attention or conscious control.

Basal ganglia.

A group of subcortical nuclei, including the striatum, central to habit learning and the acquisition of motor skills.

Cerebellum.

A hindbrain structure critical for motor adaptation and the timing and coordination of skilled movement.

Consolidation.

The process by which a newly acquired memory or skill is stabilised over time, including during rest and sleep.

Declarative memory.

The memory system for consciously accessible facts and events; memory for knowing that, dependent on the medial temporal lobe.

Habit learning.

The gradual, incremental acquisition of stimulus-response associations, supported by the neostriatum and largely automatic.

Knowledge compilation.

In ACT theory, the conversion of slow, interpreted declarative rules into fast, direct procedural production rules through practice.

Motor adaptation.

The recalibration of movement in response to a perturbation, such as a visuomotor rotation; distinct from acquiring a new skill.

Nondeclarative memory.

The umbrella category of memory expressed through performance without conscious recollection, spanning procedural skills, priming, and conditioning.

Power law of practice.

The regularity that the time to perform a skill decreases as a power function of the number of practice trials.

Priming.

A nondeclarative effect in which prior exposure to a stimulus facilitates its later processing, independent of conscious memory for the exposure.

Procedural memory.

The nondeclarative system that stores skills, habits, and learned sequences, expressed through improved performance; memory for knowing how.

Serial reaction time task.

A laboratory task in which participants respond to a repeating spatial sequence and speed up on it, often without awareness that a sequence exists.

Striatum.

The principal input structure of the basal ganglia, central to habit formation and reinforcement-driven learning.

Key Researchers

Suzanne Corkin (1937-2016). Massachusetts Institute of Technology; studied patient H.M. for nearly half a century and documented that his motor-skill learning improved across days while he retained no memory of the training, the founding demonstration that procedural learning is preserved in medial-temporal amnesia.

Faculty Page - Wikipedia

Julien Doyon. McGill University and the Montreal Neurological Institute; mapped the cortico-striatal and cortico-cerebellar circuits that support motor-sequence learning and how their contributions shift across the fast and slow stages of skill acquisition and consolidation.

Faculty Page - ORCID - Google Scholar

Barbara J. Knowlton. University of California, Los Angeles; demonstrated a neostriatal habit-learning system in humans, dissociable from declarative memory, through probabilistic classification learning that amnesic patients acquire normally but Parkinson's patients cannot.

Faculty Page - ORCID

John W. Krakauer (b. 1967). Johns Hopkins University; sharpened the distinction between the reduction of movement error (adaptation) and the genuine acquisition of new skill, and characterised the time course and neural basis of practice-driven improvement.

Faculty Page - ORCID - Google Scholar - Wikipedia

Brenda Milner (b. 1918). Montreal Neurological Institute, McGill University; showed that patient H.M. learned the mirror-drawing task normally across days despite having no memory of ever having done it, the first evidence that skill learning is a separate memory system from conscious recollection.

Faculty Page - Wikipedia

Russell A. Poldrack (b. 1967). Stanford University; used neuroimaging to show that the hippocampal declarative system and the striatal procedural system can compete during learning, with activity in one predicting suppression in the other.

Faculty Page - ORCID - Google Scholar - Wikipedia

Larry R. Squire (b. 1941). University of California San Diego and the VA San Diego Healthcare System; coined the declarative-nondeclarative taxonomy that gives procedural memory its formal place as a form of nondeclarative memory, and drew the knowing-how versus knowing-that dissociation from amnesic skill learning.

Faculty Page - ORCID - Google Scholar - Wikipedia

Daniel B. Willingham (b. 1961). University of Virginia; proposed an influential neuropsychological theory of motor-skill learning that separates the perceptual-motor, cognitive, and conscious components of a skill onto distinct control systems.

Faculty Page - Google Scholar - Wikipedia

Frequently Asked Questions

What is procedural memory in simple terms?
Procedural memory is the memory for how to do things: skills and habits such as riding a bicycle, typing, or reading. It is expressed by doing the skill better rather than by consciously recalling anything, which is why it is called memory for knowing how (Cohen & Squire, 1980).

How is procedural memory different from declarative memory?
Declarative memory stores facts and events that can be consciously brought to mind and stated, memory for knowing that; procedural memory stores skills expressed through performance, memory for knowing how. The two depend on different brain systems and can be independently damaged (Squire & Zola, 1996).

Why can amnesic patients still learn new skills?
Because procedural memory does not depend on the hippocampus and medial temporal lobe, the structures destroyed in dense amnesia. Patient H.M. improved on mirror-drawing across days while having no memory of the practice, showing that skill learning is a separate, spared system (Milner, Corkin, & Teuber, 1968).

Which parts of the brain support procedural memory?
Chiefly the basal ganglia (especially the striatum), the cerebellum, and the primary motor cortex. Damage to the striatum, as in Parkinson's disease, impairs habit learning while sparing declarative memory, the opposite of the amnesic pattern (Knowlton, Mangels, & Squire, 1996).

Can skills be learned without awareness?
Yes. In the serial reaction time task, people speed up on a repeating sequence they cannot consciously report, demonstrating that procedural learning can occur with little or no awareness of what has been learned (Nissen & Bullemer, 1987).

What is the power law of practice?
It is the finding that the time to perform a skill falls as a power function of the number of practice trials, producing large early gains that steadily diminish. Each tenfold increase in practice yields the same proportional improvement (Anderson, 1982).

Does sleep help procedural memory?
Sleep supports the consolidation that stabilises newly learned skills, through mechanisms including the reactivation of recently encoded memories. Recent work also finds rapid consolidation during brief waking rest periods between practice blocks (Klinzing, Niethard, & Born, 2019; Bonstrup et al., 2019).

Is motor adaptation the same as skill learning?
No. Motor adaptation is the recalibration of movement to a perturbation, such as a visual rotation, and is largely cerebellar; acquiring a genuinely new skill is a different process with a different time course and neural basis. Conflating them has obscured the study of both (Krakauer et al., 2019).

References