Working memory

Working memory is the small amount of information a person holds in mind and works on at a given moment, such as a phone number being dialled or the partial result of a sum. Three models dominate the field: Baddeley and Hitch’s multicomponent model, with separate stores for speech and visual material under a central executive; Cowan’s embedded-processes model, in which working memory is the activated part of long-term memory; and Oberauer’s account, in which capacity is limited by interference between items rather than by a fixed number of slots (Saults & Cowan, 2007; Hart et al., 2025; Oberauer, 2013). Their authors disagree on almost every central question, including whether working memory is separate from long-term memory (Oberauer et al., 2018b). The best-known claim about working memory, that it holds about four items, is an estimate that depends on method, and whether capacity is a number of items at all is unresolved (Cowan, 2001; Morra et al., 2024; Ma et al., 2014).

History

Working memory grew out of the short-term store of the stage model of memory. Baddeley, Hitch and Allen trace their multicomponent model to that store and still defend a short-term memory separate from long-term memory against proposals that it is merely activated long-term memory (Baddeley et al., 2019). Cowan’s 2001 review proposed that short-term storage has a single central limit of about four chunks, at a time when the field held that the limit was one of time (how long items last) rather than of capacity (Cowan, 2001).

In 2018, theorists from the rival camps wrote a joint list of benchmarks, the findings any model of working memory should explain. They list what divides them: whether working memory is one system or several, whether items decay with time, whether capacity comes in slots or as a continuous resource, and whether short-term and long-term memory are separate (Oberauer et al., 2018a). They judge the benchmark set probably too large for any model to explain in the foreseeable future (Oberauer et al., 2018b). Logie and Vandierendonck criticised the benchmarks as too narrow and of doubtful use for building theory; the authors concede that the selection and its ratings need continuous updating (Oberauer et al., 2018a).

Core concepts

The multicomponent model

Baddeley’s model attributes working memory to a central executive, which controls attention, plus specialised stores: a phonological loop for speech-based material, a visuospatial sketchpad for visual and spatial material, and later an episodic buffer that binds information across them (Saults & Cowan, 2007). A 2024 review from this side states that keeping verbal and visuospatial storage domain-specific is a key feature of the model and the main point on which it differs from Cowan’s account (Hart et al., 2025).

The embedded-processes model

Cowan’s model treats working memory as the subset of long-term memory that is currently activated, with a limited-capacity focus of attention inside it (Hart et al., 2025). The focus holds about four chunks in normal adults, where a chunk is a unit of information grouped by prior knowledge, such as a familiar word (Morra et al., 2024). In this account the limit is single and central: other storage, such as sensory memory, is not limited in capacity (Cowan, 2001).

Interference-based accounts

Oberauer’s model has three embedded components. Cowan’s four-chunk focus becomes a broader “region of direct access”, and its limit comes from interference between the bindings that tie items to their contexts, not from a fixed number of slots (Oberauer, 2013). Oberauer argues that studies give different estimates for the focus of attention, such as Cowan’s four chunks and McElree’s single chunk, because retrieval is more parallel in some tasks and more serial in others, not because capacity differs (Oberauer, 2013).

Measuring capacity

Capacity is often measured with change detection: people see an array of coloured squares, then a second array, and say whether anything changed. Cowan’s estimate k multiplies the number of items shown by the difference between the hit rate and the false-alarm rate. As arrays grow, k rises to a ceiling of between three and four objects (Saults & Cowan, 2007).

Questions, controversies, and new developments

Is the limit four?

Cowan states that the four-chunk limit can be observed only under conditions that prevent rehearsal and prevent people from recoding items into larger chunks using long-term knowledge (Cowan, 2001). The number is therefore conditional on method. Pascual-Leone and Cowan disagree on whether four is an underestimate or seven an overestimate; the disagreement turns on what counts as a unit and on whether presentation times allow rehearsal and chunking (Morra et al., 2024). Morra and colleagues find that the estimate shifts by about one unit depending on what is counted, and use this to reconcile Cowan’s four with Pascual-Leone’s larger estimate (Morra et al., 2024).

Slots or resources?

Slot models treat working memory as all-or-none storage of a fixed number of items: an item is either held or not. Resource models drop the item limit and hold that a continuous resource is spread across all items, so that the quality of each representation determines performance (Ma et al., 2014). The question is unresolved. Resource models face their own open problems: resources may not be infinitely divisible, and their fit to natural scenes is untested (Ma et al., 2014).

Evidence has gone both ways:

• Comparing 32 model combinations, van den Berg, Awh and Ma found strong evidence against precision coming in fixed quanta and for continuous precision that varies from item to item and trial to trial (van den Berg et al., 2014).
• Adam, Vogel and Awh (2017) asked people to report every item in a display in turn. After the third or fourth report, errors looked like pure guesses, which they took as evidence of an item limit of about three to four (FLoRA Replication Atlas (FORRT), 2017).
• Schurgin, Wixted and Brady argue that once the nonlinear similarity of colours is taken into account, a single signal-detection model explains results taken to show a fixed capacity, and that slot predictions conflict with earlier data (Schurgin et al., 2019).

The failed replication of Adam and colleagues

Oberauer ran four experiments, including two exact replications of one of Adam and colleagues’ experiments, and found no guessing-like errors after the third or fourth response (Oberauer, 2022; FLoRA Replication Atlas (FORRT), 2017). Modelling each participant separately, Oberauer found a mixture of people who guessed and people who did not, which Oberauer reads as favouring strategic guessing over a built-in item limit (Oberauer, 2022).

Ngiam and colleagues, defending item limits, attribute Oberauer’s result to differences in the order in which people chose to report items (Ngiam et al., 2022). They report new evidence for item limits, but concede that the classic pattern of errors can also be fit by models in which memories are present but very low in fidelity, so that the item-limit reading is disputed rather than settled (Ngiam et al., 2022). They note that the standard mixture model suggested people store about three items, and that the debate over whether items ever fail to be stored continues (Ngiam et al., 2022). An independent analysis of Adam and colleagues’ data found that people group their reports by colour similarity, which violates the independence between items that standard models assume (Cuthbert et al., 2024).

Separate from long-term memory?

Whether working memory and long-term memory are separate systems, or one system in which working memory is a small activated region of long-term memory, is an open debate (Oberauer et al., 2018b). Baddeley and colleagues defend the separate store and reject the activation view, as a party to the dispute (Baddeley et al., 2019).

Broader connections

Working memory and attention are closely intertwined, but how they interact depends on the specific kind of each, which suggests that neither is a single, unitary construct (Awh et al., 2006). Working-memory load is the defining feature of deliberate, Type 2 thinking in dual-process theory (Evans & Stanovich, 2013). The modelling problems in the slot-versus-resource dispute, where different models fit the same error patterns, are an instance of the general limits discussed in Methods of cognitive psychology.

Further reading

• Cowan (2001): the case for a central limit of about four chunks, and the conditions under which it can be measured.
• Ma et al. (2014): an accessible review of slot and resource models.
• Oberauer et al. (2018b): the benchmark findings that models should explain, written by authors from the rival camps.
• Ngiam et al. (2022): the item-limit side’s reply to the failed replication.