Encoding Specificity¶

Version: v1 · 2026-08-24 · History
Prime #: 828
Origin domain: Cognitive Science
Subdomain: memory and retrieval → Cognitive Science

Core Idea¶

Encoding specificity names the recurring structural pattern in which the retrievability of stored information depends not on the information's intrinsic properties but on the overlap between the features active at encoding and the features available at retrieval. The structural commitment is that storage is feature-bound: the features co-active when an item was laid down become part of the item's storage key, so a later cue retrieves the item to the extent that it reinstates those features, and not otherwise.

Four structural elements are jointly required: an item to be stored — a memory trace, an embedding, an indexed document, a tacit skill, a piece of organizational knowledge; an encoding context, the constellation of features (semantic neighbours, ambient state, modality, framing, language, task, location) co-active when the item was laid down; a retrieval cue with its own constellation of features, used later to attempt access; and a match function whose probability of retrieval rises with the overlap between encoding-context features and retrieval-cue features. The diagnostic signature is retrieval-by-context-reinstatement: identical items are differentially accessible from different retrieval contexts, even when the target is identical, because the path depends on context overlap. The deep structural insight is that content cannot be stored neutrally — the act of encoding always co-encodes the context, and the context becomes part of the key. This is sharper than generic content-addressable memory, which says retrieval is cue-driven without specifying which features bind the cue, and sharper than priming, which says recent activation lowers a threshold without specifying the encoding-context dependency.^[1]

How would you explain it like I'm…

Same Room, Same Memory

Have you ever forgotten something, but the moment you walked back into the room where you first learned it, it popped right back? That's because your brain saved the memory together with where and how you first got it. So a hint that brings back those same surroundings helps you remember much better than a hint that doesn't.

Clues That Match

Encoding specificity is the idea that whether you can remember something depends not just on the thing itself, but on how well your reminder matches the moment you first learned it. When you store a memory, the surrounding details — where you were, your mood, the smells, the language you were using — get tucked in alongside it, almost like part of its secret key. Later, a clue brings the memory back best when it brings back those same details. That's why studying for a test in a room like the test room can help. The very same memory can feel easy to reach from one situation and impossible from another, just because of how well the clues line up.

Context Is The Key

Encoding specificity is the pattern in which whether you can remember something depends not on the item's own importance but on the overlap between the features active when it was learned and the features available when you try to recall it. The core commitment is that storage is feature-bound: the features co-active at encoding become part of the item's storage key, so a cue retrieves the item to the extent that it reinstates those features, and not otherwise. That means identical items are differentially accessible from different contexts — the same target can be easy to reach from one situation and hard from another, because the path depends on context overlap. This is sharper than generic content-addressable memory, which says retrieval is cue-driven but doesn't specify which features bind the cue. It's also sharper than priming, which says recent activation lowers a threshold without pinning down the encoding-context dependency.

Encoding specificity names the recurring structural pattern in which the retrievability of stored information depends not on the information's intrinsic properties but on the overlap between the features active at encoding and the features available at retrieval. The structural commitment is that storage is feature-bound: the features co-active when an item was laid down become part of the item's storage key, so a later cue retrieves the item to the extent that it reinstates those features, and not otherwise. Four structural elements are jointly required: an item to be stored — a memory trace, an embedding, an indexed document, a tacit skill, a piece of organizational knowledge; an encoding context, the constellation of features (semantic neighbours, ambient state, modality, framing, language, task, location) co-active when the item was laid down; a retrieval cue with its own constellation of features, used later to attempt access; and a match function whose probability of retrieval rises with the overlap between encoding-context features and retrieval-cue features. The diagnostic signature is retrieval-by-context-reinstatement: identical items are differentially accessible from different retrieval contexts, even when the target is identical, because the path depends on context overlap. The deep structural insight is that content cannot be stored neutrally — the act of encoding always co-encodes the context, and the context becomes part of the key. This is sharper than generic content-addressable memory, which says retrieval is cue-driven without specifying which features bind the cue, and sharper than priming, which says recent activation lowers a threshold without specifying the encoding-context dependency.

Structural Signature¶

the stored item — the encoding context co-active at storage — the retrieval cue with its own feature constellation — the match function over feature overlap — the context co-encoded into the storage key — the retrieval-by-context-reinstatement signature

A configuration exhibits encoding specificity when each of the following holds:

A stored item. Some item is laid down for later access — a memory trace, an embedding, an indexed document, a tacit skill, a piece of organisational knowledge.
An encoding context. A constellation of features — semantic neighbours, ambient state, modality, framing, language, task, location — is co-active at the moment the item is stored.
A context-bound key. The act of encoding co-encodes those features into the item's storage key; content cannot be stored neutrally. This is the load-bearing commitment.
A retrieval cue. A later cue arrives with its own constellation of features, used to attempt access.
A match function. The probability of retrieval rises with the overlap between encoding-context features and retrieval-cue features, and not with the item's intrinsic content.
A reinstatement signature. Identical items are differentially accessible from different retrieval contexts, because the access path depends on context overlap — so a strong semantic associate not co-encoded can be a worse cue than a weak one that was.

Composed, these make retrievability a function over two feature sets rather than a property of the item — separating presence, retrievability, and storage strength, and distinguishing the pattern from plain content-addressable memory (the cue matches encoding-context features, not intrinsic content), from priming (one reinstatement mechanism, not the principle), and from a genuinely absent item.

What It Is Not¶

Not associative_memory. Associative memory is the mechanism that retrieves by cue; encoding specificity is the principle that retrievability depends on overlap between encoding-context and retrieval-cue features — the use of that mechanism, with the context-as-key claim added.
Not priming. Priming is one reinstatement mechanism — recent activation lowering a threshold; encoding specificity is the broader principle that reinstatement of encoding-time features is what governs retrieval.
Not pattern_recognition. Pattern recognition matches input to a learned category; encoding specificity governs whether a stored item is reachable from a given cue, independent of any categorisation.
Not analogy. Analogy abstracts roles and maps structure across domains; encoding specificity matches feature overlap between encoding and retrieval, with no role abstraction.
Not interpretation. Interpretation derives meaning; encoding specificity concerns access — whether a cue reinstates the features that bind a stored item's key.
Not provenance. Provenance traces an item's origin chain; encoding specificity concerns retrievability from feature overlap, not the item's recorded history.
Common misclassification. Diagnosing a retrieval failure as a missing item (re-creating it) when it is merely unreachable from the current cue — or vice versa. The test is whether retrieval succeeds from a cue known to reinstate encoding features.

Broad Use¶

The canonical instance is memory research: strong semantic associates can fail as cues if a different associate was active at encoding; state-dependent and context-dependent recall and mood-congruent retrieval all show identical material differentially accessible across contexts.^[1] In information retrieval, a document indexed with one embedding model is retrievable only by queries embedded in the same space, and switching the query encoder while leaving the index unchanged silently breaks retrieval — the content is there, but the key has changed.^[2] In databases, a record stored under a composite key including a tenant identifier is unretrievable by a query that omits it, even when every other field matches. In organizational knowledge management, tacit knowledge encoded in the context of a specific project is often unretrievable in a later, different-context project even when written down, because the documentation does not reinstate the features that made it meaningful. In education, a procedure learned in one context frequently fails to transfer to a structurally identical problem in another, because retrieval cues do not reinstate encoded features — the transfer-of-learning literature reads as engineering against encoding specificity.^[3] In forensic interviewing, the cognitive interview deliberately reinstates the original context to improve recall, and at the infrastructural level a cache keyed on an incidental fingerprint retrieves only entries cached under the same fingerprint.^[4]

Clarity¶

The pattern separates three things otherwise confused: item presence (is it stored at all?), item retrievability (given that it is stored, can a given cue reach it?), and storage strength (how robust is it to interference?). Storage strength alone does not predict retrievability: a strongly stored item can be inaccessible from a non-overlapping cue, and a weakly stored item accessible from a strongly overlapping one — the tip-of-the-tongue state, where storage is intact but cues fail, is the clean case.^[5] It also exposes a counter-intuitive geometry: a strong semantic associate of the target may be a worse cue than a weak one if the weak one was co-encoded and the strong one was not. The cue-target strength matrix at retrieval is not the same as the abstract semantic-association matrix, and naming the pattern is what makes that distinction available rather than mysterious.

Manages Complexity¶

The pattern compresses a heterogeneous set of retrieval phenomena — state-dependent recall, context-dependent recall, mood congruence, language-of-encoding effects, transfer failure, the cognitive interview, embedding-model brittleness, cache invalidation — into a single named shape with a portable three-question diagnostic: what features were co-active at encoding, which of them can be reinstated at retrieval, and where is the gap? It separates two scalars routinely conflated: the information content of an item, how much it carries, and the retrieval-context match, how reachable it is from a given vantage point. Most retrieval failures are the second kind, not the first; storing more does not help if the cue cannot reinstate encoding features. That separation is what turns a vague "we lost the knowledge" into a locatable gap between two feature sets.

Abstract Reasoning¶

Recognising the pattern supports inference about systems whose content seems lost but is in fact merely unreachable. The transfer problem: if a skill or fact does not transfer, the first hypothesis is that the encoding context is not being reinstated, not that the skill was never encoded — and the intervention space differs, reinstatement (replay the context) versus re-encoding (relearn under broader contexts). Indexer-query coherence: any system indexing information with a learned representation creates an implicit retrieval-key surface that queries must hit, so changing the indexer without updating the query path produces silent retrieval failure. Documentation that survives the project: writing a decision down is insufficient; it must be retrievable by cues that future readers will have, so the intervention is to encode the readers' future cues, not the encoder's own. Forensic retrieval: when memory fails, do not press for more recall under the new context — reinstate the original one and ask again. Cache miss versus genuine absence: if a system reports a miss when the data is plainly stored, the cue probably does not reinstate the composite key. These are structural inferences, true wherever storage is feature-bound.

Knowledge Transfer¶

Because the structural claim — retrievability is a function over the overlap between encoding-time and retrieval-time feature sets — is medium-neutral, the interventions transfer across substrates without re-derivation. The prediction that retrieval depends on encoding-context features transfers from memory research directly to embedding-based search, where the intervention vocabulary (reinstate context, expand cues to match encoding features, train query and document encoders jointly) reads as engineering against an encoding-specificity failure. It transfers to education, where varying the encoding contexts, teaching the retrieval cue alongside the content, and practising retrieval under diverse cues are encoding-specificity-aware moves against transfer failure. It transfers to organizational documentation, where the cure for knowledge becoming inaccessible after a project shift is to capture context features with the artifact. And the flow runs both ways: the engineering practice of composite-keying foregrounded the encoding-as-key insight that cognitive science articulated separately. The transferable intervention family is compact — index by the cues you will have, replay encoding features to retrieve, and separate "is it stored?" from "can I reach it?" — and these instructions ride substrate-unchanged. The transfer also carries its boundaries: a receiving domain must distinguish encoding specificity from plain content-addressable memory (the cue need not match intrinsic content, but encoding-context features), from priming (one mechanism for reinstatement rather than the principle that reinstatement is needed), and from a missing skill (a retrieval-cue mismatch rather than an absent capability). A practitioner who has diagnosed a retrieval failure as a key mismatch in one substrate arrives at the next already asking what was co-encoded, what the current cue reinstates, and where the gap lies — three questions that travel from human memory to vector store to runbook to classroom without translation.

Examples¶

Formal/abstract¶

Tulving and Thomson's cued-recall demonstration is the prime's canonical experimental instance, and it isolates the match function with a clean counter-intuitive result.^[1] The stored item is a target word, say BLACK.^[1] At encoding, BLACK is studied alongside a weak associate cue — the word train — so train becomes part of the encoding context co-active at storage, and is co-encoded into the storage key.^[1] At test, the experimenter offers a retrieval cue that is a strong semantic associate of the target — white, since black-white is among the strongest associations in the language.^[1] Intuition says the strong associate should be the better cue. The prime predicts, and the data confirm, the opposite: recall is higher from the weak co-encoded cue train than from the strong but non-co-encoded cue white, because retrievability is a match function over feature overlap, not a function of the item's intrinsic semantic neighbours.^[1] This is the retrieval-by-context-reinstatement signature in its sharpest form — the identical target word is differentially accessible purely as a function of which cue reinstates the encoding context. It also cleanly separates presence (BLACK is stored), retrievability (reachable from train, not from white), and storage strength, exactly the distinctions the prime makes available.

Mapped back: BLACK is the stored item, train is the co-encoded context bound into the key, white versus train are competing retrieval cues, and the weak-cue-beats-strong-cue result is the feature-overlap match function overriding intrinsic semantic strength.

Applied/industry¶

A vector-search retrieval pipeline instantiates the same prime in an information-retrieval substrate, where the failure mode is silent and the fix is exact. The stored items are documents, each indexed as an embedding produced by a particular encoder model; the encoding context is that model's learned representation space, co-encoded into the storage key — the document's vector position.^[2] A retrieval cue is a user query, embedded into a vector and matched against the index by cosine similarity — the match function over feature overlap.^[2] The prime predicts a specific, easily-missed failure: if an engineer upgrades the query encoder while leaving the document index built by the old encoder, the two feature spaces no longer overlap, and retrieval silently degrades even though every document is still present. This is the prime's clean separation of presence from retrievability — the content is there, the key has changed — and its indexer-query coherence inference names the fix directly: re-embed the corpus with the new encoder, or train query and document encoders jointly, so the cue reinstates the encoding-time features. A structurally identical applied instance is organisational knowledge management, where a decision documented in the rich context of one project is unretrievable in a later project because the write-up fails to reinstate the cues a future reader will actually have — fixed by indexing on the readers' future cues, not the author's.

Mapped back: Document embeddings are the stored items, the encoder's space is the co-encoded key, the embedded query is the retrieval cue, cosine similarity is the match function, and the mismatched-encoder silent failure is a retrievability gap with presence intact.

Structural Tensions¶

T1 — Specific Encoding versus Transferable Retrieval (sign/opposed goals). The prime says the encoding context becomes part of the key, which makes retrieval reliable from matching contexts but brittle across them. Specificity buys within-context accessibility at the cost of cross-context transfer — and the two are in direct tension. Failure mode: encoding richly to a single context (one project, one task, one cue set) and finding the knowledge unretrievable everywhere else, the transfer failure the prime names. Diagnostic: ask whether the item will be retrieved from the same context it was encoded in or a different one; if different, narrow encoding is the liability and varied-context encoding is needed.

T2 — Reinstatement versus Re-encoding (intervention/boundary). When an item fails to transfer, the prime offers two opposed remedies — reinstate the original encoding context (replay the cues) or re-encode under broader contexts (relearn). They pull apart: reinstatement preserves the narrow key, re-encoding rebuilds it. Choosing wrong wastes the effort. Failure mode: laboriously reinstating an original context (reconstructing the project setting) when the durable fix was to re-encode against the cues future users will actually have. Diagnostic: ask whether the original retrieval context will recur; if it will, reinstate; if it will not, reinstatement is a dead-end and re-encoding is required.

T3 — Presence versus Retrievability (scopal/diagnostic-confusion). The prime's sharp separation — an item can be stored yet unreachable — is its core insight, but it is also a diagnostic hazard: a true absence (the item was never encoded) and a retrieval-cue mismatch present identically as "we can't find it." Failure mode: assuming a retrievability gap (and hunting for the right cue) when the item genuinely was never stored, or assuming absence and re-creating an item that was merely unreachable. Diagnostic: test retrieval from a cue known to reinstate encoding features; success means it was a retrievability gap, failure across all plausible cues points toward genuine absence.

T4 — Strong Associate versus Co-encoded Cue (measurement/counter-intuition). The prime's signature result is that a weak co-encoded cue can beat a strong semantic associate that was not co-encoded — retrieval tracks feature overlap, not intrinsic association. This inverts the natural cue-selection heuristic. Failure mode: choosing retrieval cues by their semantic strength to the target (the obvious associate) rather than by what was actually co-active at encoding, so the "best" cue systematically misses. Diagnostic: ask what was co-active at encoding, not what is most strongly associated in the abstract; the retrieval-time cue-target matrix is not the semantic-association matrix.

T5 — Index by Encoder's Cues versus Index by Retriever's Cues (scopal/perspective). The prime warns that documentation must be retrievable by cues future readers will have, not the cues the author had — yet the author naturally encodes against their own context. The tension is between whose feature set the key is built from. Failure mode: capturing rich author-side context (jargon, project-specific framing) that no future retriever will reinstate, producing an artefact that is present but unreachable by its intended audience. Diagnostic: ask whose cues the storage key is built from; if it is the encoder's and not the anticipated retriever's, the item is keyed for the wrong vantage point.

T6 — Indexer–Query Coherence versus Independent Evolution (temporal/coupling). The prime notes that any learned-representation index creates a retrieval-key surface the query must hit — so changing the indexer without updating the query path silently breaks retrieval. The tension is that index and query encoders are often maintained independently and drift apart over time. Failure mode: upgrading the query encoder (or the index) in isolation, leaving two non-overlapping feature spaces, so retrieval degrades while every item remains present. Diagnostic: ask whether the encoding and retrieval representations are kept coherent; independently versioned encoders are a latent silent-failure waiting for the next one-sided upgrade.

Structural–Framed Character¶

Encoding specificity sits just structural of the middle on the structural–framed spectrum: the core — retrievability set by the overlap between features active at encoding and features available at retrieval, because context is co-encoded into the storage key — is a genuine relational rule, but it was minted in memory psychology and is bound to systems that store and retrieve, which gives it a mild residual frame.

Evaluative weight reads fully structural: feature-overlap-governs-retrieval carries no approval or disapproval — high overlap is helpful for recall and harmful for context-bound forgetting, value-neutral until specified. The three diagnostics that nudge it toward framed sit at the half-mark. Human-practice-bound (0.5): the rule applies to memory-bearing systems — human recall, an information-retrieval index, a vector embedding store, a database — so it presumes something that stores items with a key, rather than running in indifferent physical media; but those substrates are not all human, which keeps it at 0.5 rather than fully framed. Vocabulary travels (0.5): "encoding," "retrieval cue," "context reinstatement," "the storage key" carry a cognitive-psychology home lexicon that IR/search and embeddings adopt by translation. Institutional origin (0.5): cognitive science supplies the concept rather than a bare formal relation. Import-vs-recognise (0.5): invoking it half-imports the memory-trace framing alongside the bare feature-overlap structure.

The honest reading is that the relational rule is real and substrate-spanning — that is what keeps it on the structural side of the boundary — while its binding to storage-and-retrieval systems, its psychology vocabulary, and its disciplinary origin give it a half-measure of frame on three diagnostics against fully neutral evaluative load. The result is an aggregate just structural of centre, matching the assigned mixed-structural grade.

Substrate Independence¶

Encoding specificity is a strongly substrate-independent prime — composite 4 / 5 on the substrate-independence scale. Its domain breadth is wide (4 / 5): retrieval succeeds to the degree the retrieval context reinstates the encoding context, a feature-overlap rule that recurs across human memory, information retrieval and search, embeddings (query-document overlap in vector space), databases (index-aligned access), knowledge management, and education (transfer succeeding when cues match the learning context). Its structural abstraction is high (4 / 5): the relational rule is real and substrate-spanning, stated as a match between stored and probe features without commitment to any medium. What holds it to a 4 is that it is bound to storage-and-retrieval systems, carries psychology vocabulary, and has a disciplinary origin (transfer evidence 4 / 5): the cross-domain recurrence is concrete and documented, but each domain adopts the encoding/retrieval lexicon rather than already owning it, giving it a half-measure of frame on three diagnostics against fully neutral evaluative load.

Composite substrate independence — 4 / 5
Domain breadth — 4 / 5
Structural abstraction — 4 / 5
Transfer evidence — 4 / 5

Relationships to Other Abstractions¶

Current abstraction Encoding Specificity Prime

Parents (1) — more general patterns this builds on

Encoding Specificity presupposes Associative Memory Prime

'associative memory is the mechanism that retrieves by cue; encoding specificity is the PRINCIPLE that retrievability depends on overlap between encoding-context and retrieval-cue features' — it governs/uses that mechanism, adding the context-as-key claim.

Children (6) — more specific cases that build on this

Context-Dependent Memory Domain-specific is a kind of Encoding Specificity

Context-dependent memory is the human-memory specialization of encoding specificity in which incidental environmental, physiological, or cognitive context features become part of the trace's retrieval key.
Memory color effect Domain-specific is a kind of Encoding Specificity

The proposed strict upward parent is prime:encoding_specificity.
Mood-dependent memory Domain-specific is a kind of Encoding Specificity

The proposed strict upward parent is prime:encoding_specificity.

▸ Show 3 more

State-Dependent Learning Domain-specific is a kind of Encoding Specificity
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 memory Domain-specific is a kind of Encoding Specificity
The proposed strict upward parent is `prime:encoding_specificity`.
Transfer-Appropriate Processing Domain-specific is a kind of Encoding Specificity
**Encoding Specificity — instantiates and specializes.** TAP restricts encoding-retrieval overlap to the cognitive operations and information products recruited at study and test.

Neighborhood in Abstraction Space¶

Encoding Specificity sits in a moderately populated region (46^th percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.

Family — Encoding, Retrieval & Substrate Mapping (11 primes)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-10

Not to Be Confused With¶

The most important contrast is with associative_memory, because the two are related as principle to mechanism. Associative memory is the content-addressable retrieval mechanism — a cue evokes a stored item by some matching operation. Encoding specificity is the sharper claim about that mechanism: retrievability depends not on the item's intrinsic content but on the overlap between the features co-active at encoding and the features present in the retrieval cue, because the encoding context is co-encoded into the storage key. Plain content-addressable memory says "retrieval is cue-driven" without specifying which features bind the cue; encoding specificity says the binding features are the encoding-time ones, which yields the counter-intuitive result that a weak co-encoded cue can beat a strong semantic associate that was not co-encoded. The distinction is load-bearing: a designer who holds only associative memory will pick cues by their intrinsic association to the target, whereas encoding specificity directs them to pick cues that reinstate what was co-active at storage — a systematically different and more reliable choice.

A second confusion is with priming. Priming is genuinely a piece of the same territory — recent activation of a representation lowers its retrieval threshold — but it is one mechanism of reinstatement, not the principle that reinstatement is what matters. Encoding specificity is the general claim that retrieval succeeds to the extent the cue reinstates encoding-context features, of which priming-driven activation is a single instance alongside context reinstatement, state dependence, and mood congruence. Treating encoding specificity as priming narrows a broad structural principle to one of its mechanisms, and leads a practitioner to reach only for activation-based interventions (recency, repetition) when the deeper fix is to match the retrieval context to the encoding context by whatever means — replaying the setting, indexing on the retriever's future cues, or co-encoding the cues the audience will actually have.

Finally, encoding specificity is distinct from analogy, with which it shares the language of "matching" but differs in what is matched and how. Analogy abstracts the roles in a source structure and maps them onto a target in a different domain — it transfers relational structure, deliberately ignoring surface features. Encoding specificity matches feature overlap between an encoding context and a retrieval cue, and is exquisitely sensitive to exactly the surface and contextual features analogy abstracts away — which is why a strong semantic (analogical) associate can fail as a retrieval cue while a weak but co-encoded one succeeds. The two pull in opposite directions: analogy's power is its indifference to co-occurrence, whereas encoding specificity's whole content is that co-occurrence at encoding governs access. Confusing them leads to expecting structurally-similar cues to retrieve when only context-overlapping cues will.

Solution Archetypes¶

Solution archetypes in the catalog that build on this prime — directly (this prime is a source ingredient) or as a related prime.

Built directly on this prime (3)

Also a related prime in 6 archetypes

Context-Keyed Representation Switching: Maintain several context-specific representations on one substrate, activate the right one from validated context cues, isolate inactive maps from interference, and preserve them for reliable re-entry.
Fast–Slow Store Coupling: Keep a volatile fast store and a durable integrated store coupled by governed transfer so the system gets immediate access without losing long-term coherence.
Nearest-Exemplar Response Reuse: Use the closest remembered or stored case as the model for the present response, while making similarity, adaptation, confidence, and exception boundaries explicit.
Offline Replay Consolidation: Replay captured experience traces in a protected offline window so the rerun, not the live event alone, writes durable memory, skill, policy, or model structure.
Post-Encoding Trace Stabilization: Protect a newly encoded trace long enough for it to stabilize, integrate, and survive later interference rather than relying on immediate recall.
Retrieval-Cued Revision: Reactivate a stored pattern in a bounded update window, pair it with a corrective difference, and re-stabilize the revised version instead of trying to overwrite it while it is dormant.

References¶

[1] Tulving, Endel, and Donald M. Thomson. "Encoding Specificity and Retrieval Processes in Episodic Memory". Psychological Review, vol. 80, no. 5 (1973): 352–373. Originating paper of the encoding specificity principle, including the cued-recall demonstration that a weak co-encoded cue can outperform a strong non-co-encoded semantic associate. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g

[2] Karpukhin, Vladimir, Barlas Oğuz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. "Dense Passage Retrieval for Open-Domain Question Answering". Proceedings of EMNLP 2020: 6769–6781. Dual-encoder dense retrieval: documents are indexed as embeddings from a passage encoder and queries embedded by a separate encoder into the same space, with retrieval by dot-product/cosine similarity — supporting that the encoder's learned space is the storage key, that similarity is the match function, and that a query encoder mismatched to the index space breaks retrieval (requiring re-embedding or jointly trained encoders). registry ↩a ↩b ↩c

[3] Barnett, Susan M., and Stephen J. Ceci. "When and Where Do We Apply What We Learn? A Taxonomy for Far Transfer". Psychological Bulletin, vol. 128, no. 4 (2002): 612–637. Reviews the transfer-of-learning literature showing procedures learned in one context routinely fail to transfer to structurally identical problems in another. registry ↩

[4] Geiselman, R. Edward, Ronald P. Fisher, David P. MacKinnon, and Heidi L. Holland. "Eyewitness Memory Enhancement in the Police Interview: Cognitive Retrieval Mnemonics Versus Hypnosis". Journal of Applied Psychology, vol. 70, no. 2 (1985): 401–412. Demonstrates the cognitive interview's deliberate reinstatement of original context to improve eyewitness recall. registry ↩

[5] Brown, Roger, and David McNeill. "The 'Tip of the Tongue' Phenomenon". Journal of Verbal Learning and Verbal Behavior, vol. 5, no. 4 (1966): 325–337. The canonical study of the tip-of-the-tongue state, the clean case where storage is intact but retrieval cues fail. registry ↩