Skip to content

State-Dependent Learning

The memory phenomenon in which material encoded in one internal physiological or psychological state is retrieved more reliably when the agent is again in a matching state — because the body's condition at learning is bound into the trace as part of the encoding context — diagnosed by a crossover interaction where matched states beat mismatched ones.

Core Idea

State-dependent learning is the phenomenon in cognitive psychology in which material encoded in one internal physiological or psychological state is retrieved more reliably when the agent is again in a matching internal state, and less reliably when the internal states at encoding and retrieval diverge. The internal state — pharmacological (intoxication level, caffeine, benzodiazepine), physiological (arousal, fatigue, body temperature), or affective (mood, anxiety level) — functions as part of the encoding context: the memory trace is bound not only to the content and the external situation but also to the body's condition at the time of learning. When retrieval occurs in a state that overlaps with the encoding state, that state-component of the context provides an additional retrieval cue, and recall improves; when there is a state mismatch, that cue is absent and recall degrades. The defining experimental signature is a crossover interaction in a 2×2 design of encoding-state × retrieval-state: both state-matched conditions (drug-encoded / drug-retrieved; sober-encoded / sober-retrieved) outperform both state-mismatched conditions (drug-encoded / sober-retrieved; sober-encoded / drug-retrieved), at effect sizes that cannot be explained by the impairing effect of the drug on encoding or retrieval considered alone. The effect is graded with degree of state overlap, content-neutral across verbal, motor, and episodic material, and distinct from mood-congruent recall — which concerns a match between mood and the emotional valence of the memorised content — because state-dependent learning does not require content to be mood-matching, only states to match. The clinical and training corollary is practical: material learned under conditions of high arousal, pharmacological state, or emotional context is best retrieved when those conditions are reproduced, which has implications for trauma-exposure therapy, athletic skill training, and examination preparation.

Structural Signature

Sig role-phrases:

  • the internal state — a pharmacological, physiological, or affective condition of the body (intoxication, arousal, fatigue, mood) at the time of learning
  • the encoding event — material learned while the agent is in that state, which binds the state into the memory trace as part of the encoding context, alongside content and external setting
  • the retrieval event — a later recall attempt occurring in a state that may or may not match the encoding state
  • the state-component cue — the bound state functioning as an additional retrieval cue, present under overlap and absent under mismatch
  • the match-not-level principle — recall governed by the overlap between encoding and retrieval states, not by how "good" or impairing either state is on its own
  • the graded benefit — partial state overlap yielding partial benefit, so recall reads as a continuous gradient over degree of overlap
  • the crossover signature — in a 2×2 of encoding-state × retrieval-state, both matched cells beating both mismatched cells (a mere main effect of state would refute the account)
  • the content-neutral scope — the effect appearing equally for verbal, motor, and episodic material, requiring only the states to match (distinguishing it from mood-congruent recall, which needs content valence to match)

What It Is Not

  • Not the state impairing memory. Poorer recall under intoxication, fatigue, or high arousal is not the drug or stress being "bad for the brain"; what governs recall is the match between encoding and retrieval states, because the body's condition at learning is bound into the trace. A mere main effect of state — uniformly worse recall under the drug regardless of encoding — would refute the account; the crossover interaction, where both matched cells beat both mismatched cells, is its signature.
  • Not state-dependent performance. State-dependent performance is a task simply being easier in some states regardless of where it was learned; this is about retrieval of material encoded in that state. The diagnostic: a performance benefit appears even for material not encoded in the state, whereas the learning effect is keyed to the encoding state specifically.
  • Not mood-congruent recall. Mood-congruent recall turns on a match between current mood and the emotional valence of the content (sad mood favoring sad material). State-dependent learning is content-neutral — it appears equally for verbal, motor, and episodic material and requires only the states to match, not the content to be mood-matching.
  • Not context-dependent memory. Context-dependent memory is cued by the external setting (the diving-on-land/in-water study); state-dependent learning is cued by the internal condition of the body. Same matching shape, different locus of the cue — internal physiological/affective state versus the surrounding environment.
  • Not ML distribution shift or a database read-consistency model. A model trained on one distribution and deployed on another shares the structural insight (encoding context ≠ retrieval context) but is distribution_shift/data_drift; a distributed database whose read succeeds only in a matching state is engineered consistency. Neither has an internal physiological state bound into a memory trace, a 2×2 crossover, or the pharmacology-and-exposure-therapy apparatus — the shared "match between store-time and read-time conditions" belongs to the appropriate parent, not to this named effect.

Scope of Application

State-dependent learning lives across the memory and learning subfields of cognitive psychology; its reach is within that domain — a memory trace bound to the agent's internal physiological or psychological state, diagnosed by the 2×2 crossover. The cross-substrate look-alikes (ML distribution shift, database read-consistency) route to different primes (distribution_shift, encoding-specificity composition), not to this named effect.

  • Pharmacology of memory — the strongest empirical case: alcohol, cannabis, nicotine, caffeine, and benzodiazepine state-dependence in human and animal memory tests.
  • Affective and arousal-dependent retrieval — mood states, test anxiety, and sports-performance arousal show matched-state retrieval advantages, motivating "study how you'll be tested."
  • Physiological-state research — fatigue and body-temperature states as encoding context that conditions later retrieval.
  • Animal-learning paradigms — an extensive rodent literature on state-dependent acquisition and extinction of conditioned responses.
  • Clinical and training practice — arousal reinstatement in trauma-exposure therapy and "train like you fight" skill regimes apply the state-match principle directly; its sibling context-dependent memory (Godden and Baddeley's diving study) extends it to external context under Tulving's encoding-specificity principle.

Clarity

Naming state-dependent learning relocates the explanation of an erratic recall pattern from the level of an internal state to the match between two states. Without the concept, the natural reading of poorer recall under intoxication, fatigue, or high arousal is that the state simply impairs memory — that the drug or the stress is bad for the brain. The label commits the analyst instead to an encoding-retrieval-matching account: the body's condition at learning is bound into the memory trace as part of the encoding context, so what governs recall is whether the retrieval state overlaps the encoding state, not how "good" either state is on its own. That converts a vague sense that "conditions affect memory" into a sharp, falsifiable prediction — a crossover interaction in a 2×2 of encoding-state × retrieval-state, in which both matched cells beat both mismatched cells, at an effect that the impairing action of the state alone cannot produce. A main effect of state would refute the account; the crossover is its signature.

The concept also draws lines among neighbours that are otherwise easy to run together. It separates state-dependent learning from state-dependent performance, where a task is simply easier in some states regardless of where it was learned; from context-dependent memory, where the matching cue is the external setting rather than the internal condition; and, most importantly, from mood-congruent recall, which turns on a match between current mood and the emotional valence of the content. The distinguishing test is that state-dependent learning is content-neutral and requires only the states to match, not the material to be mood-matching. Pinning that down tells a clinician or trainer exactly which lever is operating, and makes the practical move precise — reproduce the encoding state at retrieval — rather than leaving "study how you'll be tested" as folklore.

Manages Complexity

Across the pharmacological, physiological, and affective memory literatures the raw data is a mass of recall results that look mutually inconsistent: material learned under alcohol recalled poorly when sober but well when intoxicated; high-arousal encoding hurting some test conditions and helping others; caffeine, fatigue, body temperature, and mood each producing their own scatter of better-and-worse recall across studies. Faced with any one result, a researcher would otherwise need a substance-specific or state-specific account — is this drug bad for encoding, that arousal bad for retrieval, this mood impairing in general? State-dependent learning compresses the entire sprawl onto a single principle the Core Idea and Clarity establish: recall is governed by the match between encoding state and retrieval state, because the body's condition at learning is bound into the trace as part of the encoding context. The whole inconsistent-looking pattern resolves once the analyst stops asking how good or bad each state is on its own and instead tracks one quantity — the degree of overlap between the two states — and reads recall off it.

The compression collapses to essentially one parameter with a sharp diagnostic shape and a few qualifying conditions, which is what lets outcomes be predicted rather than re-discovered per substance. The governing parameter is state overlap, and because the effect is graded, partial overlap predicts partial benefit, so the analyst reads a continuous recall gradient off how closely retrieval reproduces encoding. The signature that tells the analyst this mechanism (rather than a state-impairment mechanism) is operating is the crossover interaction the Clarity foregrounds: in the 2×2 of encoding-state × retrieval-state both matched cells beat both mismatched cells, and a mere main effect of state would refute the account — so the model is falsifiable and the read-off unambiguous. Two further commitments keep the parameter set small by ruling out competing explanations: content-neutrality (verbal, motor, episodic material all behave the same, so the analyst need not track material type), and the boundary lines the Clarity draws — internal state, not external setting (versus context-dependent memory), and state-match, not content-valence match (versus mood-congruent recall) — which route any given finding to this mechanism only when the states, not the contents, do the matching. So instead of a substance-by-substance, state-by-state inventory of memory effects, the analyst carries one match principle indexed by a single overlap parameter, a crossover diagnostic, and a content-neutral scope, and from them predicts both the direction of a recall effect and the practical move (reproduce the encoding state at retrieval) — the move from a tangle of state-specific results to one encoding-retrieval-matching law.

Abstract Reasoning

State-dependent learning licenses a tight family of inferences in the memory literature, each flowing from the encoding-retrieval-matching account rather than a state-impairment one.

Diagnostic — reason from a recall pattern to match, not level. The characteristic observable is recall that varies with internal state — poorer when intoxicated, fatigued, or highly aroused. The framework runs that variation backward to its true governing variable: rather than inferring that the state impaired memory, infer that the state at encoding was bound into the trace and that recall depends on how far the retrieval state overlaps it. The decisive diagnostic is the crossover interaction in a 2×2 of encoding-state × retrieval-state: if both matched cells (drug-encoded/drug-retrieved and sober-encoded/sober-retrieved) beat both mismatched cells, infer the matching mechanism; a mere main effect of state — uniformly worse recall under the drug regardless of encoding — would refute it and point to simple impairment. So the shape of the 2×2, not the average level, tells the analyst which mechanism is operating, and the model is falsifiable on exactly that shape. A further diagnostic uses gradedness: a continuous recall gradient that tracks degree of state overlap infers the matching mechanism, since partial overlap yields partial benefit, whereas an all-or-none pattern would not.

Interventionist — reason from the match principle to a manipulation with a predicted direction. Because recall is governed by state overlap, the single lever is the retrieval state's resemblance to the encoding state, carrying a directional prediction. Reproduce the encoding state at retrieval (test under the arousal, pharmacological, or physiological condition in which material was learned): predicted to raise recall toward the matched-cell level. Conversely, a state mismatch between study and test is predicted to depress recall by exactly the cue the matching state would have supplied — so studying caffeinated and testing sober predicts a decrement attributable to mismatch, not to ignorance. The practical corollary is "reproduce the encoding state," and it makes "study how you'll be tested" precise: train a skill under the arousal it will be performed under, sit an exam in conditions resembling study, or in exposure therapy reinstate aspects of the high-arousal encoding state to access the material. The interventionist move is a reading — measure how closely retrieval reproduces encoding and predict recall rises with that overlap; a recall gain that appears only when the state is reinstated (and not when content cues alone are added) confirms it was state-match, not content, doing the work.

Boundary-drawing — reason about which matching effect is in play. The framework draws sharp lines that route a finding to this mechanism only under specific conditions. It separates state-dependent learning from state-dependent performance, where a task is simply easier in some states regardless of where it was learned — diagnosable because performance benefit appears even for material not encoded in that state. It separates internal-state matching from context-dependent memory, where the matching cue is the external setting rather than the body's condition. And, most importantly, it separates state-match from mood-congruent recall, which turns on a match between current mood and the emotional valence of the content: the distinguishing test is content-neutrality — state-dependent learning appears equally for verbal, motor, and episodic material and requires only the states to match, not the material to be mood-matching. The boundary inference runs: if the benefit requires the content to share the mood, it is mood-congruence; if it follows the external setting, it is context-dependence; if it appears regardless of where material was encoded, it is performance, not learning — and only a content-neutral, internal-state, encoding-conditioned benefit is state-dependent learning proper.

Predictive / order-of-events. The mechanism fixes a sequence — material is encoded while the agent is in some internal state → that state is bound into the memory trace as part of the encoding context, alongside content and external setting → at retrieval the current state either overlaps the encoding state (supplying the state-component of the cue) or does not → recall is enhanced under overlap and degraded under mismatch, graded by degree of overlap. This ordering licenses predictions: that recall is highest when retrieval reinstates the encoding state and falls continuously as the states diverge (a gradient, not a cliff); that the effect is content-neutral, so the same matched-versus-mismatched advantage should appear whatever the material type, letting the analyst predict the direction of a recall effect without tracking content; and that the advantage is symmetric across states — both the drug-encoded and the sober-encoded material recall best in their own matching state — which is exactly the crossover that distinguishes this mechanism, at an effect size the impairing action of the state alone cannot produce.

Knowledge Transfer

Within the memory and learning literature state-dependent learning transfers as mechanism, with the encoding-retrieval state-match account and its crossover signature as the portable core. The same mechanism runs across the pharmacological case (alcohol, cannabis, nicotine, caffeine, benzodiazepine state-dependence in human and animal tests), the affective and arousal cases (mood, test anxiety, sports performance), the physiological cases (fatigue, temperature), and the extensive rodent acquisition-and-extinction paradigms — so the single overlap parameter, the 2×2 crossover diagnostic, the gradedness, and the content-neutral scope carry across all of them. Its structural sibling context-dependent memory (Godden and Baddeley's diving study) shares the same shape with the matching cue external rather than internal, and both sit under Tulving's encoding-specificity principle (with transfer-appropriate processing alongside), so the within-domain family is unified by one matching mechanism. The interventions port without translation too — reproduce the encoding state at retrieval, making "study how you'll be tested," "train like you fight," and arousal-reinstatement in exposure therapy precise rather than folkloric. The within-domain transfer is the state-match mechanism itself moving across the pharmacological, affective, physiological, and animal-learning settings, and into clinical and training practice.

Beyond human cognition the situation is mixed, and the seed marks it cleanly. The genuinely portable cross-substrate residue is context-cued retrieval / encoding-specificity — retrieval succeeds to the degree the retrieval context overlaps the encoding context — and that residue is carried by the parent composition encoding_and_decoding (the encoding-retrieval machinery), context (the broader background dependency), and transfer_of_learning (cross-context application, of which state-mismatch is a failure mode). Those parents travel. But the cross-substrate look-alikes are explicitly metaphor, not mechanism, and route to different primes: a distributed database's read-consistency model, where a read succeeds only if the system is in a matching state, shares the surface but is engineered consistency, not a memory trace binding a physiological state; and the famous ML failure of a model trained on one distribution and deployed on another shares the structural insight (encoding context ≠ retrieval context, so performance degrades) but is better expressed as distribution_shift / data_drift, the substrate-general parent for that case. Neither requires importing state-dependent learning, and calling them "state-dependent" would carry over the internal-physiological-state, 2×2-crossover, pharmacology-and-EMDR apparatus that has no referent in a database or a training pipeline — analogy at the level of "match between store-time and read-time conditions," mechanism only at the level of the appropriate parent. So the honest move is layered: within memory research the state-match mechanism and its crossover diagnostics travel across substances, states, species, and clinical practice; the abstract "retrieval depends on encoding-retrieval context overlap" lesson belongs to the encoding_and_decoding+context+transfer_of_learning (encoding-specificity) composition; a distribution mismatch in ML is distribution_shift/data_drift, not state-dependent learning; and "state-dependent learning," as named — internal-state-conditioned, content-neutral, crossover-diagnosed — is reserved for the cognitive case (see Structural Core vs. Domain Accent).

Examples

Canonical

The defining human demonstration is Donald Goodwin and colleagues' 1969 Science study "Alcohol and recall: state-dependent effects in man." Male subjects learned material — avoidance tasks, word associations, and rote lists — either sober or after a substantial dose of alcohol, then attempted recall a day later either sober or intoxicated, arranging the full 2×2 of encoding-state × retrieval-state. On several measures the two state-matched cells (learned-drunk/tested-drunk and learned-sober/tested-sober) yielded better recall than the two mismatched cells, including the striking result that material acquired while intoxicated was recovered better when subjects were again intoxicated than when they had sobered up. Crucially the pattern was a crossover, not a uniform alcohol decrement: alcohol did not simply damage memory, since drunk-encoded material was retrieved best in the drunk state — the fingerprint that recall tracked the match between states rather than the quality of either state alone.

Mapped back: Alcohol intoxication is the internal state; learning the word lists under it is the encoding event that binds the state into the trace. The 2×2 dissociation is exactly the crossover signature — both matched cells beating both mismatched cells — and the fact that intoxicated encoding was best recovered intoxicated instantiates the match-not-level principle: recall follows state overlap, not how impairing alcohol is. That the effect held across avoidance, association, and rote-list material shows the content-neutral scope.

Applied / In Practice

High-stakes occupational training operationalizes the state-match principle through stress-exposure training, developed for military, aviation, and emergency-response performers (Driskell and Johnston). Rather than teaching a procedure in a calm classroom and hoping it survives combat, a cockpit emergency, or a resuscitation, trainers deliberately reproduce the physiological arousal of the real event during acquisition — time pressure, noise, physical exertion, simulated threat — so that the skill is encoded in the same aroused state in which it must later be executed. The rationale is precisely that a checklist rehearsed under calm conditions is retrieved in a state that no longer overlaps the high-arousal moment of performance, degrading recall exactly when it matters most. Reinstating the encoding arousal at the point of performance restores the state-component of the retrieval cue, which is why "train like you fight" is treated as a design principle rather than mere toughening.

Mapped back: Physiological arousal is the internal state; drilling the procedure under simulated stress is the encoding event, and the real emergency is the retrieval event whose arousal must match. Building training to reproduce that arousal supplies the state-component cue at performance and follows the match-not-level principle — the goal is overlap, not a calmer or "better" state. Because the method applies equally to motor procedures, verbal checklists, and decision drills, it exercises the content-neutral scope.

Structural Tensions

T1: Match principle versus the state's own main effect (a pure crossover over a messy reality). The account stakes its identity on match-not-level: recall is governed by state overlap, and a mere main effect of state — uniformly worse recall under the drug regardless of encoding — would refute it. The clean crossover is the theory's falsifiable signature. But real states rarely behave so obligingly: alcohol, high arousal, and fatigue genuinely do impair encoding and retrieval on their own, so an empirical dataset typically carries a main effect and an interaction at once. Isolating the matching component then requires the crossover to survive on top of the impairment, not to appear in its absence. The tension is that the model's cleanest diagnostic assumes a purity the data seldom offer, and an analyst who reads only the average level will attribute to impairment what is actually match. Diagnostic: Is the matched-state advantage a crossover surviving the state's own impairing main effect, or is a simple main effect being misread as a matching one?

T2: Content-neutral generality versus surface confusability (the breadth that unifies also blurs). Content-neutrality is the effect's signature virtue: because it appears equally for verbal, motor, and episodic material and requires only the states to match, it distinguishes cleanly from mood-congruent recall and reaches across substances, arousal, fatigue, and species with one overlap parameter. Yet the same generality means that at the surface every "conditions affected recall" observation looks like a candidate for it — including the ones that are really context-dependent memory (external cue), state-dependent performance (benefit even for material not encoded in the state), or content-valence matching. The feature that gives the mechanism its wide reach is the feature that forces the analyst to run the discriminating 2×2 before claiming it, because generality removes the content-level tell that would otherwise sort the cases. Diagnostic: Does the benefit survive when content cues are held constant and only the internal state is reinstated, or could an external-context or performance effect produce the same surface pattern?

T3: The prescription versus the state it prescribes (reproduce the encoding condition, whatever it was). The mechanism yields a clean, actionable corollary — reproduce the encoding state at retrieval — that makes "study how you'll be tested," "train like you fight," and arousal reinstatement in exposure therapy precise design principles rather than folklore. But the lever it identifies is state overlap as such, indifferent to whether the state is one anyone should want to reinstate: the prescription can recommend reproducing intoxication, high trauma arousal, or acute anxiety because that is where the material was encoded. The move that maximizes recall may require re-entering a pharmacological or affective condition that is itself harmful, costly, or ethically fraught. The tension is that the mechanism optimizes for match, not welfare, and the state it points at is sometimes exactly the one the practitioner is trying to avoid. Diagnostic: Does reproducing the encoding state buy retrieval at an acceptable cost, or is the state whose reinstatement the mechanism prescribes itself the thing you must not reproduce?

T4: Graded prediction versus categorical measurement (a continuum read through a binary). The mechanism predicts a continuous recall gradient: partial state overlap yields partial benefit, so recall should fall smoothly as encoding and retrieval states diverge, and that gradedness is one of the diagnostics that separates the matching account from an all-or-none story. But internal states are usually operationalized categorically — drunk versus sober, aroused versus calm — because degree of overlap between two physiological or affective conditions is genuinely hard to quantify. A binary design cannot see the predicted gradient; it collapses the very continuum the theory rests on into two cells, so a graded mechanism and an all-or-none one become empirically indistinguishable. The tension is that the theory's richest prediction lives at a resolution the standard measurement rarely supplies. Diagnostic: Is state overlap measured finely enough to reveal the predicted continuous gradient, or only as a binary that cannot distinguish graded matching from an all-or-none effect?

T5: Autonomy versus reduction (its own named effect or the cognitive instance of encoding-specificity). "State-dependent learning" is a canonically studied cognitive effect with heavy proprietary cargo — the internal-physiological-state locus, the 2×2 crossover diagnostic, the pharmacology-and-exposure-therapy apparatus — and within memory research the state-match mechanism travels intact across substances, states, species, and clinical practice. But the structure that recurs across substrates is not "state-dependent learning" abstracted; it is the parent composition it instantiates — context-cued retrieval / encoding-specificity (encoding_and_decoding + context + transfer_of_learning) — and the cross-substrate look-alikes route to different parents: an ML model failing off its training distribution is distribution_shift, not state-dependence, and a database's read-consistency is engineered, not a trace binding a physiological state. Calling those "state-dependent" would import an internal-state apparatus with no referent there. Diagnostic: Resolve toward encoding-specificity / context-cued retrieval (or distribution_shift for the ML case) when carrying the lesson beyond cognition; toward the named effect only for an internal-state, content-neutral, crossover-diagnosed memory result.

Structural–Framed Character

State-dependent learning sits toward the structural end of the spectrum but stops short of the pole — best read as mixed-structural, in the family of STDP, spatial updating, and isostasy: a genuine, evaluatively-neutral natural mechanism wearing heavy domain vocabulary. Its evaluative_weight is nil: recall tracking the match between encoding and retrieval states is neither good nor bad, and the effect is content-neutral and value-free — the entry pointedly relocates the phenomenon from a state being "bad for memory" to a neutral matching account. It is not human_practice_bound: the effect runs automatically in memory, below any observer, and is documented across an extensive rodent literature as well as in humans — remove every psychologist and the drug-encoded trace is still recalled best in the drug state. Its institutional_origin is none: the mechanism is a discovered fact of how memory binds state into the encoding context (Goodwin 1969), not a stipulation. And cross-substrate reuse, at the level of the underlying pattern, is recognition rather than import: the state-match mechanism is recognized intact across substances, physiological and affective states, and species, and its abstract residue recurs as genuine mechanism under the encoding-specificity parents — while the ML and database look-alikes are marked as metaphor routing to different primes.

What keeps it off the structural pole is vocab_travels: the effect's distinctive cargo — the internal-physiological-state locus, the 2×2 crossover diagnostic, the pharmacology, and the exposure-therapy/"train like you fight" apparatus — is irreducibly cognitive-and-pharmacological and does not float free; calling an ML model's distribution failure "state-dependent" would import an internal-state apparatus with no referent there. The portable skeleton is context-cued retrieval / encoding-specificity — retrieval succeeds to the degree the retrieval context overlaps the encoding context — a composition of the parent primes encoding_and_decoding, context, and transfer_of_learning (of which state-mismatch is a failure mode), with its external-cue sibling being context-dependent memory under the same Tulving principle. That skeleton is what state-dependent learning instantiates by fixing the overlapping context to the body's internal state, and it is what travels; the crossover-and-pharmacology machinery stays home. Its character: structural in skeleton — a real, evaluatively-neutral, recognized-across-species encoding-retrieval matching mechanism — but stated in cognitive-pharmacological vocabulary that pins it to the memory substrate, leaving it mixed-structural rather than a free-floating prime.

Structural Core vs. Domain Accent

This section decides why state-dependent learning is a domain-specific abstraction and not a prime, and carries the case for its domain-specificity — so it is worth being exact about what could lift and what cannot.

What is skeletal (could lift toward a cross-domain prime). Strip the pharmacology and the memory apparatus away and a thin relational structure survives: a store is written together with the conditions holding at write-time, and a later read succeeds to the degree the read-time conditions overlap those write-time conditions — governed by the match between the two, not by the quality of either. The portable pieces are abstract — an encoding event that binds context into the stored trace, a retrieval event whose context may or may not overlap, an additional cue supplied only under overlap, and a graded benefit tracking degree of overlap. That is context-cued retrieval, or encoding-specificity. The skeleton is genuinely substrate-portable — it recurs with the matching context external rather than internal (context-dependent memory, the diving study) and factors into the parent composition encoding_and_decoding (the encode-retrieve machinery), context (the background dependency), and transfer_of_learning (cross-context application, of which state-mismatch is a failure mode) — but it is the core the entry shares, not what makes it distinctively state-dependent learning.

What is domain-bound. What makes the concept state-dependent learning in particular is cognitive-and-pharmacological furniture. The locus is specifically the internal physiological or psychological state of the body — intoxication, arousal, fatigue, mood — bound into a memory trace; the diagnostic is the 2×2 encoding-state × retrieval-state crossover with its content-neutral scope; the empirical cases are the pharmacology-of-memory literature (Goodwin's alcohol study, the rodent acquisition-and-extinction paradigms); and the applied apparatus is clinical and training practice (arousal reinstatement in exposure therapy, stress-exposure "train like you fight" regimes). The decisive test: remove the internal-physiological-state locus and the memory trace and it stops being state-dependent learning — if the matching cue is the external setting it is context-dependent memory; if the benefit appears even for material not encoded in the state it is state-dependent performance; if it needs the content's valence to match the mood it is mood-congruent recall. The crossover-and-pharmacology machinery is exactly what pins the concept to the cognitive substrate.

Why this does not clear the prime bar. A prime is a relational structure whose vocabulary travels and whose transfer is recognition of the same mechanism, not analogy. State-dependent learning's transfer is bimodal, and the entry marks the boundary sharply. Within memory research the state-match mechanism travels intact as mechanism — the single overlap parameter, the 2×2 crossover diagnostic, the gradedness, and the content-neutral scope carry across substances (alcohol, cannabis, nicotine, caffeine, benzodiazepines), across affective and physiological states, across species, and into clinical and training practice, because each supplies an internal state bound into an encoding trace. Beyond human cognition the cross-substrate look-alikes are explicitly metaphor, not mechanism, and they route to different parents: an ML model failing off its training distribution shares the insight (encoding context ≠ retrieval context) but is distribution_shift/data_drift, and a distributed database's read-consistency is engineered consistency, not a physiological state bound into a trace — calling either "state-dependent" would import an internal-state, crossover, exposure-therapy apparatus with no referent there. When the bare structural lesson is needed cross-substrate, it is already carried, in more general form, by the parents the entry instantiates: encoding_and_decoding supplies the encode-with-context / retrieve-by-overlap machinery, context supplies the background dependency, and transfer_of_learning supplies the cross-context-application frame of which state-mismatch is the failure mode. The cross-domain reach belongs to those parents (and, for the ML case, to distribution_shift); the named effect carries cognitive-pharmacological baggage that should stay home.

Relationships to Other Abstractions

Local relationship map for State-Dependent LearningParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.State-DependentLearningDOMAINPrime abstraction: Crossover Interaction — is a decomposition ofCrossoverInteractionPRIMEPrime abstraction: Encoding Specificity — is a kind ofEncodingSpecificityPRIME

Current abstraction State-Dependent Learning Domain-specific

Parents (2) — more general patterns this builds on

  • State-Dependent Learning is a kind of Encoding Specificity Prime

    State-dependent learning is encoding specificity specialized to an internal physiological or psychological state serving as the bound and later reinstated retrieval cue.

  • State-Dependent Learning is a decomposition of Crossover Interaction Prime

    State-Dependent Learning is the memory form of a crossover interaction because each retrieval state helps material encoded in the matching state and hurts relative recall of material encoded in the other state.

Hierarchy paths (6) — routes to 5 parentless roots

Not to Be Confused With

  • Context-dependent memory. The structural sibling in which the reinstated cue is the external setting — Godden and Baddeley's divers who learned on land or underwater and recalled best in the matching environment. Same matching shape, but the bound cue is the surrounding situation, not the body's internal condition. Tell: is the cue that must match a feature of the outside world (room, sounds, location — context-dependent memory), or a physiological/affective condition of the body (intoxication, arousal, mood — state-dependent learning)?

  • Mood-congruent recall. A distinct effect where retrieval favors material whose emotional valence matches the current mood — a sad mood surfacing sad memories. It turns on a content–mood match, whereas state-dependent learning is content-neutral and requires only the states to align, appearing equally for neutral verbal, motor, and episodic material. Tell: does the benefit require the content to carry a matching emotional valence (mood-congruent recall), or does it appear for any material as long as the internal states match (state-dependent learning)?

  • State-dependent performance. A pure contrast: a task simply being easier in some internal states regardless of where it was learned — the state helps execution, not retrieval of state-encoded material. State-dependent learning is keyed specifically to the encoding state, so its benefit does not extend to material learned elsewhere. Tell: does the state-related advantage appear even for material not encoded in that state (performance), or only for material actually learned in the matching state (learning)?

  • Encoding-specificity principle (Tulving). The broad super-type stating that a cue aids retrieval to the degree it was encoded with the target — the general law under which both state- and context-dependent memory fall (part vs. whole). State-dependent learning is the special case that fixes the encoded cue to the body's internal physiological/affective state. Tell: is the claim the general one that retrieval tracks encoding–retrieval cue overlap of any kind (encoding-specificity), or the specific one that the overlapping cue is an internal bodily state (state-dependent learning)?

  • Transfer-appropriate processing. A neighbor under the same Tulving umbrella holding that memory is best when the cognitive operations at retrieval match those at encoding (e.g., semantic vs. phonological processing). It matches the type of mental processing, not a physiological or affective state or an external setting. Tell: is what must match the kind of processing performed on the material (transfer-appropriate processing), or the internal bodily state the agent was in (state-dependent learning)?

  • Distribution shift / data drift (the ML look-alike). The machine-learning failure where a model trained on one data distribution degrades when deployed on another — a genuine "encoding context ≠ retrieval context" resemblance, but routing to the prime distribution_shift, with no internal physiological state, no memory trace, and no 2×2 crossover. Calling it "state-dependent" would import a pharmacology-and-exposure-therapy apparatus that has no referent in a training pipeline. Tell: is the mismatch between a model's training and deployment data distributions (distribution shift), or between an agent's internal states at learning and recall (state-dependent learning)?

  • The parent composition it instantiates (encoding_and_decoding, context, transfer_of_learning). The substrate-neutral primes whose composition — context-cued retrieval, where a read succeeds by overlap with write-time conditions — is what genuinely travels cross-substrate. State-dependent learning is the cognitive instance that fixes the overlapping context to the body's internal state and carries the crossover-and-pharmacology machinery the parents do not. Tell: is the bare "retrieval depends on encoding–retrieval context overlap" lesson being carried into another substrate (the parent composition), or is it the internal-state, content-neutral, crossover-diagnosed memory effect (state-dependent learning)? (Treated more fully in earlier sections.)

Neighborhood in Abstraction Space

State-Dependent Learning sits in a moderately populated region (54th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Memory Encoding & Retrieval Effects (22 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-07-12