Levels of Processing

Abstract

Levels of processing is a framework for memory in which how well an item is later remembered depends not on how long it is held in mind but on the depth to which it is analysed at encoding. Craik and Lockhart proposed that a stimulus can be processed at a shallow, sensory level, at an intermediate phonemic level, or at a deep, semantic level, and that deeper, more meaningful analysis leaves a more durable memory trace. Semantic orienting tasks reliably produce far better retention than structural or phonemic ones, even when the learner never intends to memorise. Later work qualified the framework: elaboration and distinctiveness, not depth alone, govern the benefit, and the value of any encoding depends on its match to the eventual retrieval task. Interactive demonstrations trace the depth gradient, the effect of elaboration, and the transfer-appropriate qualification.

Keywords: levels of processing, depth of processing, elaboration, semantic encoding, transfer-appropriate processing

Levels of processing holds that memory is a by-product of the operations performed on an item as it is perceived and understood, not the output of a dedicated storage system. On this account the same word can be encoded shallowly, by registering its visual shape, or deeply, by extracting its meaning, and the depth of that analysis, more than any deliberate effort to memorise, determines how well it is later recalled. The idea began as a critique of the modal store model, in which information was thought to enter long-term memory chiefly by being rehearsed in a short-term store, and grew into one of the most productive organising principles in the study of human memory (Craik & Lockhart, 1972).

What Levels of Processing Is

Levels of processing was introduced as an alternative to the prevailing account of memory, the modal or multi-store model, in which information passed from a sensory register into a limited short-term store and from there into long-term memory in proportion to how long it was rehearsed. Craik and Lockhart argued that the durability of a memory is better predicted by the kind of analysis an item receives than by the store it occupies or the time it spends there. They proposed that perception proceeds through a series of analyses, from shallow features such as lines and brightness, through the sound and name of a stimulus, to its meaning and its relations to other knowledge, and that this hierarchy of analysis constitutes a continuum of depth of processing (Craik & Lockhart, 1972).

The framework's central claim is that deeper, more semantic analysis leaves a more persistent trace. The evidence that made the idea compelling predates its formal statement. Hyde and Jenkins had shown that people who performed a meaning-based task on a list of words, without any instruction to memorise, recalled as many words as people who deliberately tried to learn them, and far more than people who performed a shallow task on the same words; the orienting task, not the intention to learn, governed retention (Hyde & Jenkins, 1969). This is the phenomenon of incidental learning: memory forms as a by-product of comprehension, so what determines it is the processing comprehension happens to require.

A corollary followed for rehearsal. If depth is what matters, then merely cycling an item in mind to hold it there, maintenance rehearsal, should do little to improve later recall, because it repeats a shallow operation rather than deepening it. Craik and Watkins confirmed exactly this: the length of time an item was held in a rehearsal buffer, when that rehearsal only maintained it, had no effect on later recall, dissociating sheer time in a short-term store from durable learning (Craik & Watkins, 1973). The distinction they drew between maintenance rehearsal, which sustains without deepening, and elaborative rehearsal, which enriches the analysis, became one of the framework's most-cited lessons.

The Depth Continuum

The sharpest test of the framework varies depth while holding everything else constant. Craik and Tulving gave people a series of words, each preceded by a question that forced a particular level of analysis: a structural question about the word's typography, a phonemic question about its sound, or a semantic question about its meaning or fit to a sentence. Because the questions could be answered without any intent to memorise, the design isolates the effect of depth itself. Recognition rose steeply across the three levels: words processed for meaning were remembered far better than words processed for sound, which in turn beat words processed for appearance (Craik & Tulving, 1975).

The demonstration reconstructs the depth gradient. Selecting a shallow, intermediate, or deep orienting question for a set of words yields a later-recognition estimate that climbs with depth, reproducing the ordered advantage Craik and Tulving reported. The lesson is not that people cannot memorise shallow features when told to, but that, for a fixed amount of engagement, meaning-based analysis leaves the most retrievable trace. Depth, in this original formulation, is the single variable that orders memory performance from worst to best.

Elaboration and Distinctiveness

Depth alone soon proved too blunt an account. Craik and Tulving themselves found that a positive answer to a semantic question produced better memory than a negative one, and that a word that fit a rich, complex sentence frame was remembered better than the same word in a simple frame, even though both were processed at the semantic level. What varied was not depth but elaboration: the extent to which the item was connected to other information and integrated into a richer encoded structure (Craik & Tulving, 1975).

A second refinement was distinctiveness, the degree to which an encoding sets an item apart from its neighbours and from competing memories. An elaborate encoding helps most when it is also distinctive, furnishing a trace that stands out and so affords a more diagnostic retrieval cue. The two ideas together explain cases that pure depth cannot, such as a shallow but highly unusual encoding outperforming a deep but generic one, and they reframe the framework: what deep processing usually buys is an elaborate, distinctive trace, and it is those properties, not a mere position on a depth scale, that drive the benefit (Craik & Tulving, 1975; Lockhart & Craik, 1990). The demonstration illustrates the point by holding the level of processing fixed at semantic while varying how elaborately each word is embedded, so that the memory estimate rises with elaboration even though depth does not change.

The Self-Reference Effect

One encoding operation reliably outperforms even ordinary semantic processing: relating an item to oneself. Rogers, Kuiper, and Kirker asked people to judge whether trait adjectives described them, and found that words encoded with reference to the self were recalled better than words encoded for meaning, structure, or sound, a robust advantage they named the self-reference effect (Rogers et al., 1977). Within the levels framework the self acts as an especially rich and well-organised knowledge structure, so referring an item to it is an unusually elaborate and distinctive form of semantic encoding.

The effect has held up under quantitative scrutiny. Symons and Johnson's meta-analysis of the self-reference literature confirmed a reliable memory advantage for self-referent encoding over both semantic and other-referent encoding, and attributed it to the self's power to promote elaboration and organisation at encoding (Symons & Johnson, 1997). The self-reference effect thus sits naturally at the deep end of the processing continuum while also pointing beyond it, toward the special mnemonic status of self-relevant information that later work would pursue in its own right.

Transfer-Appropriate Processing

The most consequential challenge to a strict depth ordering was the demonstration that no encoding is best in the abstract. Morris, Bransford, and Franks had people encode words either semantically or by attending to their rhyme, then tested memory in one of two ways: a standard recognition test, or a rhyme-recognition test in which the task was to identify words that rhymed with studied items. Semantic encoding won on the standard test, as levels of processing predicts, but rhyme encoding won on the rhyme test. The value of an encoding operation was not fixed by its depth but by how well it matched the operations the test would later require, a principle they called transfer-appropriate processing (Morris et al., 1977).

Table 1 lays out the crossover that makes the point. Reading down the columns, deep semantic encoding is superior only when the test demands semantic information; when the test demands phonemic information, the shallower rhyme encoding transfers better. The finding does not overturn the levels framework so much as bound it: deep processing is advantageous for the kinds of retrieval that dominate everyday memory and most laboratory tests, but it is not universally superior, and depth is no guarantee of better memory when the test rewards surface information instead. A related demonstration by Fisher and Craik showed the complementary point from the retrieval side: recall was best when the retrieval cue reinstated the same qualitative relation that had been encoded, so encoding and retrieval operations interact rather than acting independently (Fisher & Craik, 1977). This dependence is the encoding specificity principle, Tulving and Thomson's demonstration that a retrieval cue aids recall only to the degree that its relation to the target was encoded at study, so what is remembered is jointly determined by the encoding and the retrieval environment rather than by depth in isolation (Tulving & Thomson, 1973).

Neural Correlates of Depth

If depth of processing is a real dimension of encoding, deeper analysis should have identifiable neural signatures, and it does. Using positron emission tomography, Kapur and colleagues found that a semantic encoding task, judging whether words denoted living things, produced greater activity in the left inferior prefrontal cortex than a shallow task judging the presence of a letter, and that this difference tracked the superior memory the semantic task produced, giving the depth effect a concrete neuroanatomical correlate (Kapur et al., 1994). Deep encoding, on this evidence, is not a metaphor but a distinct pattern of prefrontal engagement.

Functional MRI sharpened the picture by relating encoding activity to whether an item would actually be remembered. Otten, Henson, and Rugg compared the neural correlates of depth across and within tasks and identified regions, again including the left inferior prefrontal cortex, whose activity at study predicted later memory success, the subsequent memory effect. Crucially, deeper semantic processing recruited these regions more strongly, linking the behavioural depth advantage to the same encoding activity that forecasts remembering item by item (Otten et al., 2001). The neuroimaging thus converges on the framework's core claim: meaning-based encoding engages a characteristic left-prefrontal and medial-temporal circuit, and the degree of that engagement is what deeper processing buys.

Worked Example

The depth gradient can be made quantitative with a simple incidental-learning study in the style of Craik and Tulving (Craik & Tulving, 1975). Suppose 48 words are studied, evenly split so that 16 are processed with a structural orienting question, 16 with a phonemic question, and 16 with a semantic question, and that the later recall probabilities these levels typically produce are 0.15, 0.35, and 0.70.

- Words recalled per level. Multiplying each probability by 16 gives 0.15 × 16 = 2.4 words from structural encoding, 0.35 × 16 = 5.6 from phonemic, and 0.70 × 16 = 11.2 from semantic. - Total and overall rate. The sum is 2.4 + 5.6 + 11.2 = 19.2 words recalled of 48, an overall rate of 19.2 / 48 = 40%. The single number hides a steep gradient underneath it. - The size of the depth effect. Semantic encoding yields 11.2 / 2.4 = 4.67 times as many recalled words as structural encoding of the very same list. Had all 48 words been encoded semantically, expected recall would be 0.70 × 48 = 33.6 words, or 70%, against 0.15 × 48 = 7.2 words, or 15%, for all-structural encoding.

The arithmetic makes concrete what the framework asserts qualitatively: with study time, list, and intention held constant, shifting the level of analysis from appearance to meaning multiplies retention several-fold. The first demonstration runs exactly this computation, letting the orienting level set the per-word probability and reading off the resulting recall.

Discussion

Levels of processing reoriented the study of memory by relocating the determinants of retention from the properties of hypothetical stores to the operations performed at encoding. Its lasting contributions are the demonstration that meaning-based processing produces durable memory almost automatically, the dissociation of maintenance from elaborative rehearsal, and a vocabulary, depth, elaboration, distinctiveness, that remains standard. Yet the framework's original form drew a well-aimed critique. Baddeley argued that depth was never independently measured: the only evidence that a task was deep was that it produced good memory, and the only evidence that it produced good memory was that it was deep, so the central term threatened to be circular and the framework unfalsifiable without an independent index of depth (Baddeley, 1978). One response to this measurement worry was methodological rather than definitional. Jacoby's process-dissociation procedure separated the consciously controlled, intentional contribution to a memory judgment from the automatic, familiarity-based one, letting researchers ask not merely whether a manipulation such as depth improved overall memory but which component of memory it acted on — a finer-grained target than the single recall or recognition score on which the circularity charge rested (Jacoby, 1991).

The reply, developed over the following decades, was to abandon depth as a literal metric and keep it as a useful description of encoding operations. Lockhart and Craik's retrospective conceded the circularity of a strict depth scale and recast the framework around elaboration, distinctiveness, and the encoding-retrieval interaction that transfer-appropriate processing had made unavoidable (Lockhart & Craik, 1990). Craik's later assessments treat levels of processing not as a falsified theory nor a settled one but as a durable orienting attitude toward memory, one that correctly insists that remembering is a consequence of the mental activity an event provokes rather than of its consignment to a store (Craik, 2002; Craik, 2020). What survives is not a ruler of depth but a principle: memory is the residue of thought.

Current Directions

The framework's most active descendant is the study of self-referential encoding, which has grown from a levels-of-processing footnote into a research programme of its own. Building on the self-reference effect, Humphreys and Sui proposed a self-attention network in which self-relevant information is prioritised by a distinct fronto-parietal circuit, arguing that the memory advantage for self-referent material reflects an attentional bias toward the self rather than depth of semantic processing alone (Humphreys & Sui, 2016). The proposal reframes an old levels-of-processing result as one facet of a broader prioritisation of self-relevant stimuli across cognition (Cunningham & Turk, 2017).

That prioritisation appears to operate even before long-term encoding. Yin and colleagues showed that self-associated stimuli are automatically favoured in working memory, competing more successfully for its limited capacity than stimuli associated with others, evidence that the mnemonic privilege of self-relevant information begins at the earliest stages of processing rather than only at retrieval (Yin et al., 2019). Alongside this, contemporary reviews continue to fold the depth effect into an account of memory as an activity of mind and brain, in which encoding, attention, and the self jointly determine what is retained (Craik, 2020). The through-line is a shift from asking how deep an encoding is to asking what makes some encodings automatically prioritised, with self-reference the clearest current case.

Common Misconceptions

There are exactly three fixed levels of processing.

Structural, phonemic, and semantic are convenient illustrative points, not a rigid three-stage architecture. Craik and Lockhart described depth as a continuum of analysis, and later statements treated the levels as a graded description of encoding operations rather than discrete boxes (Craik & Lockhart, 1972; Lockhart & Craik, 1990).

Deeper processing always produces better memory.

Depth helps only when the retrieval task rewards the information that deep encoding captures. When a test demands phonemic information, shallower rhyme-based encoding transfers better, so the best encoding depends on the match between study and test, not on depth alone (Morris et al., 1977).

Repeating something over and over is what fixes it in memory.

Maintenance rehearsal that merely holds an item in mind does little for long-term recall; it is elaborative processing that connects the item to meaning which builds a durable trace. Sheer time in a rehearsal buffer, without deepening, does not improve later memory (Craik & Watkins, 1973).

Glossary

Deep processing.

Analysis of a stimulus for its meaning and its relation to existing knowledge, the level of encoding that produces the most durable memory trace.

Depth of processing.

The dimension along which encoding operations range from shallow sensory analysis to deep semantic analysis; the central explanatory variable of the framework.

Distinctiveness.

The degree to which an encoded trace stands apart from competing memories, affording a more diagnostic retrieval cue and enhancing recall.

Elaboration.

The extent to which an item is connected to other information at encoding, enriching the trace; a richer, more elaborate encoding is better remembered than a spare one at the same depth.

Elaborative rehearsal.

Rehearsal that deepens or enriches the analysis of an item by relating it to meaning, as opposed to merely holding it in mind; it builds durable memory.

Encoding specificity.

The principle that a retrieval cue is effective only to the extent that its relation to the target was encoded at study, so memory depends jointly on encoding and retrieval conditions.

Encoding.

The set of operations that transform a perceived stimulus into a memory trace; in this framework the nature of those operations, not their duration, sets retention.

Incidental learning.

Memory that forms without any intention to learn, as a by-product of processing a stimulus for some other purpose; used to isolate the effect of encoding depth.

Levels of processing.

The framework holding that memory durability depends on the depth to which an item is analysed at encoding rather than on time in a short-term store.

Maintenance rehearsal.

Rehearsal that holds an item in mind by repeating it without deepening its analysis; it sustains availability briefly but does little for long-term recall.

Orienting task.

A question or judgment that directs processing to a particular level, such as appearance, sound, or meaning, and thereby controls the depth of encoding.

Phonemic processing.

Intermediate-level analysis of a stimulus for its sound or pronunciation, deeper than structural analysis but shallower than semantic analysis.

Process dissociation.

A procedure that decomposes performance on a memory test into a consciously controlled, intentional component and an automatic, familiarity-based one, so the separate influence of a variable such as encoding depth on each can be estimated.

Self-reference effect.

The superior memory for information encoded in relation to oneself, understood as an unusually elaborate and organised form of semantic encoding.

Semantic processing.

Deep analysis of a stimulus for its meaning and associations, the level of encoding that yields the strongest and most durable memory.

Shallow processing.

Analysis confined to surface features such as a stimulus's appearance or sound, producing a fragile trace that is quickly lost.

Structural processing.

The shallowest level of analysis, concerned with the physical or visual form of a stimulus, such as whether a word is printed in capital letters.

Transfer-appropriate processing.

The principle that the benefit of an encoding operation depends on how well it matches the operations required at retrieval, so no encoding is best for every test.

Key Researchers

Alan D. Baddeley (b. 1934). Emeritus Professor of Psychology at the University of York; his 1978 critique The Trouble With Levels argued that depth of processing was defined circularly and pressed the field toward independent, operational measures of encoding, shaping how the framework was subsequently defended and revised. ORCID - Faculty Page - Google Scholar - Wikipedia - Wikidata

Fergus I. M. Craik (b. 1935). University Professor Emeritus at the University of Toronto and Senior Scientist Emeritus at the Rotman Research Institute, Baycrest; he co-originated the levels-of-processing framework with Lockhart in 1972 and, with Tulving, demonstrated that semantic encoding tasks yield far better retention than shallow structural or phonemic ones. ORCID - Faculty Page - Google Scholar - Wikipedia - Wikidata

Larry L. Jacoby (1944-2024). Late Professor of Psychology at Washington University in St. Louis; through his work on transfer-appropriate processing and process dissociation he showed that the mnemonic value of an encoding operation depends on its match to the retrieval task, qualifying any strict depth ordering. Faculty Page - Google Scholar - Wikipedia - Wikidata

Robert S. Lockhart (dates unavailable). Professor Emeritus in the Department of Psychology at the University of Toronto; he co-authored the 1972 framework and the 1990 retrospective with Craik, reframing levels of processing as a description of encoding operations rather than a literal set of stages. Faculty Page

Michael D. Rugg (b. 1954). Distinguished Professor at the Center for Vital Longevity, University of Texas at Dallas; with Otten and Henson he used functional MRI to identify the neural correlates of depth-of-processing effects, linking deeper semantic encoding to left inferior prefrontal and medial-temporal activity that predicts later memory. ORCID - Faculty Page - Google Scholar - Wikipedia - Wikidata

Endel Tulving (1927-2023). Late University Professor Emeritus at the University of Toronto and Senior Scientist at the Rotman Research Institute; he collaborated with Craik on the 1975 depth-of-processing experiments and, through encoding specificity, established that retention depends on the compatibility between encoding and retrieval rather than depth alone. Faculty Page - Google Scholar - Wikipedia - Wikidata

Frequently Asked Questions

What is the levels-of-processing theory in simple terms?
It is the idea that how well an item is remembered depends on how deeply it is thought about when first encountered. Analysing a word for its meaning leaves a stronger memory than noticing how it looks or sounds, so memory is a by-product of the kind of mental work a person does, not of how long the item is held in mind (Craik & Lockhart, 1972).

Who proposed the levels-of-processing framework?
Fergus Craik and Robert Lockhart introduced it in 1972 as an alternative to the multi-store model of memory, and Craik developed it further with Endel Tulving in a 1975 series of experiments that varied the depth of encoding while holding other factors constant (Craik & Lockhart, 1972; Craik & Tulving, 1975).

What are the levels of processing?
The framework describes a continuum from shallow structural processing, which attends to a stimulus's physical appearance, through intermediate phonemic processing of its sound, to deep semantic processing of its meaning. Deeper analysis produces more durable memory, though the three levels are illustrative points on a continuum rather than fixed stages (Craik & Tulving, 1975).

What is the difference between maintenance and elaborative rehearsal?
Maintenance rehearsal keeps an item in mind by simple repetition without deepening its analysis, and does little for long-term memory. Elaborative rehearsal connects the item to meaning and existing knowledge, and it is this deeper processing that builds a lasting trace (Craik & Watkins, 1973).

Is deeper processing always better for memory?
No. Deep semantic encoding is best only when the memory test rewards the meaning-based information it captures. When a test instead requires surface information such as rhyme, shallower encoding that matches the test transfers better, a principle called transfer-appropriate processing (Morris et al., 1977).

What is the self-reference effect?
It is the finding that information encoded in relation to oneself is remembered better than information encoded for general meaning or surface features. Referring an item to the self is treated as an especially elaborate and organised form of deep processing, and meta-analysis confirms its reliability (Rogers et al., 1977; Symons & Johnson, 1997).

What was the main criticism of levels of processing?
Alan Baddeley argued that depth was never measured independently of the memory it was meant to explain: a task was called deep because it produced good memory, and good memory was explained by depth, making the account circular without a separate index of processing depth (Baddeley, 1978).

Does brain imaging support levels of processing?
Yes, in part. Positron emission tomography and functional MRI show that semantic encoding tasks engage the left inferior prefrontal cortex more than shallow tasks, and that the strength of this engagement predicts which items are later remembered, giving the depth effect a neural correlate (Kapur et al., 1994; Otten et al., 2001).

References