Memory

Memory is the capacity to keep information over time and use it later. Cognitive psychology has described it in two ways that still coexist: as a sequence of stores that information passes through, from brief sensory traces to a lasting long-term store, and as a set of separate systems, some conscious and some not, each supported by different brain structures (Spielman et al., 2020; Squire & Zola, 1996). The systems view remains the field’s dominant framework, but two disputes run through it (Sherman et al., 2023). The first is whether conscious and nonconscious memory are truly separate systems or one memory measured in different ways. The second is whether depth of processing, rather than the store information sits in, decides what is remembered. Several well-known memory effects have also been tested in large replications, with mixed results.

History

The stage model credited to Atkinson and Shiffrin (1968) divides memory into sensory, short-term and long-term stores, and introductory textbooks still teach memory this way (Spielman et al., 2020). About fifty years after it was proposed, its defenders describe it as the most cited model of its period (Baddeley et al., 2019). They also acknowledge two weaknesses. The model underestimates the complex encoding strategies that people use, which levels-of-processing studies brought to light. And patients described by Shallice and Warrington had impaired short-term memory but could still learn new long-term material, which contradicts the model’s claim that information reaches long-term memory by passing through the short-term store (Baddeley et al., 2019). Cowan offered another reading of these patients; Baddeley and colleagues reply that there is no evidence for it.

The short-term store itself became contested. Some researchers propose that it is not a separate store at all but the currently activated part of long-term memory. Baddeley and colleagues reject this and defend a separate short-term store (Baddeley et al., 2019); they are a party to the dispute, which the Working memory article covers in detail.

Long-term memory was in turn divided into systems. Squire and Zola’s taxonomy separates a declarative system, which holds facts and events that can be consciously recalled and depends on the medial temporal lobe, from nondeclarative learning that operates without awareness (Squire & Zola, 1996). Textbooks give the same split under the names explicit memory (consciously recalled) and implicit memory (not part of consciousness) (Spielman et al., 2020).

Core concepts

Stores and stages

A store is a hypothetical place where information is held for a characteristic length of time. In the stage model, sensory memory holds a brief copy of what the senses register, short-term memory holds a small amount of information for a few seconds, and long-term memory holds information indefinitely (Spielman et al., 2020). Short-term memory is now usually studied as working memory, the system that holds and manipulates information for the task at hand.

Memory systems

A memory system, in this tradition, is a form of memory with its own rules of operation and its own neural basis. Declarative memory is conscious and depends on the medial temporal lobe; nondeclarative memory covers learning that proceeds without awareness (Squire & Zola, 1996). The standard view holds that the two are supported by different brain systems (Squire & Dede, 2015).

Implicit memory is defined by what it lacks. Schacter and colleagues describe it as a change in performance caused by earlier information or practice, without conscious recollection of that earlier episode (Schacter et al., 1993). Its two main forms are priming, where an earlier exposure makes a later response to the same or a related item faster or more likely, and skill learning. Reber notes that learning outside declarative memory is characterised only by this absence of awareness, and Reber proposes that implicit memory is general plasticity of the cortex rather than one separate system (Reber, 2013).

The multiple-systems framework is still the field’s dominant guide to memory research. Sherman and colleagues argue that it needs revision, because recent work challenges its one-to-one mapping of brain structures to memory types (Sherman et al., 2023). Duss and colleagues, for example, report that unconscious encoding of relations between items depends on the hippocampus, a structure the standard taxonomy assigns to conscious memory; they propose that the kind of processing, not consciousness, divides memory systems (Duss et al., 2014).

Learning within memory

Textbooks and reviews treat learning as part of memory rather than as a separate faculty: the types of long-term memory are themselves the types of learning, split into declarative and nondeclarative, or explicit and implicit (Reber, 2013; Schacter et al., 1993). Learning by association, as in classical and operant conditioning, has its own models and is covered in Associative learning.

Levels of processing

The levels-of-processing framework, proposed by Craik and Lockhart in 1972, holds that how well something is remembered depends on how deeply it was processed when first met: thinking about a word’s meaning leaves a better memory than noticing its sound or its typeface. Craik’s own retrospective surveys the framework’s legacy and its common criticisms, including what role short-term memory plays and how depth can be measured (Craik, 2002).

The original idea of a fixed series of levels, from perceptual to semantic, was abandoned after Craik and Tulving (1975), but levels of processing survives as a broad framework (Baddeley & Hitch, 2017). Its effects depend on the material. In Baddeley and Hitch’s survey they are modest for visual stimuli, with a mean effect size d of 0.5 (d expresses a difference between conditions in standard deviations), but range from 0.6 to 1.4 for verbal stimuli (Baddeley & Hitch, 2017).

Transfer-appropriate processing is the main rival. It holds that memory depends on the match between how information was processed at study and what the test requires. Morris, Bransford and Franks (1977) showed that shallow coding of a word’s sound beat deep coding of its meaning when the test asked for rhymes (Baddeley & Hitch, 2017). Roediger and Gallo treat levels of processing as a special case of transfer-appropriate processing, for meaning-based tests of memory for words (Baddeley & Hitch, 2017). Lockhart finds little support for a simple version of that idea and suggests only a modified version can hold (Lockhart, 2002).

Questions, controversies, and new developments

One memory or several?

Whether conscious and nonconscious memory are separate systems is contested. The case for separation rests on dissociations: conditions in which one form of memory is impaired while the other is spared. Amnesic patients, whose declarative memory is severely damaged, showed intact priming for words, pseudowords and nonwords after a single exposure while their recognition was impaired (Hamann & Squire, 1997). Patient E.P. showed intact priming on a word-stem completion task but recognised the same items only at chance (Stark & Squire, 2000). In five memory-impaired patients, conceptual priming was fully intact while the two most severely amnesic patients recognised items at chance (Levy et al., 2004). In another group of severely amnesic patients, fluency and priming were present and recognition was at chance; the authors calculate that fluency raises accurate recognition only to 52 percent (Conroy et al., 2005). These researchers read their results as showing that priming does not depend on declarative memory and is not the source of the feeling of familiarity. The standard view that the two kinds of memory rest on different brain systems is set out by Squire and Dede (Squire & Dede, 2015).

Berry, Shanks and Henson argue that dissociations like these do not necessarily imply distinct systems (Berry et al., 2008). Their single-system model assumes one memory signal that drives both priming and recognition, with different amounts of noise in each measure, and it accounts for key dissociations once used as evidence for independent systems (Berry et al., 2008). In a model comparison, the single-system model predicted results in advance that an independent multiple-systems model did not, although a more flexible multiple-systems model, which lets the systems depend on each other, could accommodate them (Berry et al., 2012). Applied to amnesic patients, the single-system model predicted the pattern of recognition and priming and was favoured over two multiple-systems models (Berry et al., 2014). Newell and Shanks add that unconscious influences in general are overstated because tests of awareness are inadequate (Newell & Shanks, 2014).

The two sides thus read the same kind of patient data in opposite ways, and the question is unresolved.

Retrieval practice and forgetting

Testing yourself on material, rather than restudying it, improves later memory; this is the testing effect. Rowland’s 2014 meta-analysis found larger benefits when the practice test required recall rather than recognition, and concluded that mechanisms other than those usually proposed should be considered (Rowland, 2014). Retrieving some items can also make related items harder to recall later, an effect called retrieval-induced forgetting. A 2014 meta-analysis mostly supported accounts that explain it by active inhibition of the competing items, but found evidence against them that varied with how forgetting was measured (Murayama et al., 2014).

Suppression-induced forgetting is the claim that deliberately trying not to think of a memory weakens it. It is tested with the Think/No-Think procedure, in which people learn word pairs and are then told to recall some answers and suppress others.

Replication record

A gap in effect size between classroom testing and its laboratory extension stands out. Boustani and Shanks conclude that test-potentiated new learning with standard designs rests on weak evidence and call for preregistered experiments (Boustani & Shanks, 2022). Wiechert and colleagues say their null result may reflect the online format, but they still question the robustness of suppression-induced forgetting (Wiechert et al., 2023). Replication in psychology more broadly is covered in Replicability and generality.

Broader connections

Memory and attention depend on each other: attention selects what gets encoded, and memory guides what is attended next (Chun & Turk-Browne, 2007). Turner and Sloutsky (2024) describe this as a cycle in which memory establishes what is relevant and attention encodes what becomes memory. The short-term end of memory is treated in Working memory, and learning by association in Associative learning.

Further reading

• Baddeley et al. (2019): the history of the stage model, told by defenders of a separate short-term store.
• Sherman et al. (2023): the case that the multiple-systems taxonomy needs revision.
• Berry et al. (2008): the single-system model of implicit and explicit memory.
• Yang et al. (2021b): the largest meta-analysis of the testing effect in classrooms.