Working Memory Models: From Baddeley’s Multicomponent Theory to Integrative Cognitive and Neural Frameworks

Baddeley’s multicomponent framework remains a central reference point, but subsequent work has elaborated, challenged, and partially integrated it with attention-based, interference-based, individual-differences, and neurocognitive models of working memory (Adams et al., 2018; Baddeley, 2012; Logie et al., 2022).

Baddeley’s multicomponent model and revisions

Baddeley and Hitch’s original model proposed that short-term maintenance for complex cognition is supported by a central executive plus two specialised “slave systems”: the phonological loop and the visuospatial sketchpad (Baddeley, 2012). Dual-task experiments showed that people could maintain verbal material while performing visuospatial tasks with relatively modest interference, but performance dropped sharply when both tasks drew on the same subsystem, supporting the idea of partially independent verbal and visuospatial stores under a shared attentional controller (Baddeley, 2012).

The phonological loop was further decomposed into a temporary phonological store and an articulatory rehearsal process, explaining phenomena such as phonological similarity effects, word-length effects, and the benefits of subvocal rehearsal for span (Baddeley, 2012). The visuospatial sketchpad was treated as a counterpart for visual and spatial information, with later suggestions that even this system may have separable visual and spatial components, based on dissociations in neuropsychological and imaging data (Baddeley, 2012).

To explain how information from different modalities and from long-term memory could be combined into integrated representations, Baddeley (2000) introduced the episodic buffer as an additional component. The episodic buffer is conceived as a limited-capacity, multimodal store that supports binding of visual, verbal, and semantic information into coherent “episodes” and provides a link between the slave systems and long-term memory (Baddeley, 2000, 2012). In Baddeley’s own later reflections, the multicomponent model is presented as a flexible framework rather than a fixed boxology, with ongoing efforts to refine the nature of the components and their neural underpinnings (Baddeley, 2012).

Embedded-processes and attention-based approaches

Cowan’s embedded-processes model reconceptualised working memory as the currently activated subset of long-term memory, within which a small focus of attention selects a few items for direct manipulation (Cowan, 1999, 2010). In this framework, activation decays rapidly without attention, but activated representations are not confined to a separate short-term store; instead, they are ordinary long-term memory traces that are temporarily heightened in accessibility (Cowan, 2010). Empirical estimates of capacity in conditions that limit rehearsal and chunking suggest that the focus of attention holds about three to four chunks, aligning with the “magical number four” rather than seven (Cowan, 2010).

Oberauer extended the embedded-processes idea by distinguishing three levels of accessibility: activated long-term memory, a “region of direct access,” and the single-item focus of attention (Oberauer, 2002, 2013). In his account, many representations can be activated, a small subset in the region of direct access can be used for current task operations, and one selected item occupies the focus of attention and is immediately available for the next cognitive operation, such as updating in complex span or running span tasks (Oberauer, 2002, 2013). This fine-grained structure is supported by behavioural paradigms that differentiate costs of switching between items in the focus versus bringing new items from activated long-term memory into direct access, and by time-course data in probe-recognition and retro-cue studies (Adams et al., 2018; Oberauer, 2013).

Adams et al. (2018) argue that attention-based theories such as Cowan’s and Oberauer’s can be located on continua of modularity and attentional reliance, with both emphasising the central role of executive attention and goal maintenance rather than specialised structural stores. From this perspective, many classic “storage” phenomena (e.g., span limits, interference) are reinterpreted as consequences of limitations in attentional selection, goal maintenance, and control over activation (Adams et al., 2018).

Capacity, interference, and resource accounts

Interference-based and resource-based models were developed to explain fine-grained patterns of performance in working memory tasks, especially those involving continuous report or complex binding demands (Ma et al., 2014; Oberauer, 2009). Oberauer’s interference-based formal models show how feature overwriting and similarity-driven interference can produce capacity-like limits without assuming a hard item cap: as more items share features, retrieval becomes noisier and errors increase, even if the system does not run out of discrete “slots” (Oberauer, 2009).

Resource models, inspired by signal-detection and population-coding frameworks, treat working memory as a flexible but finite representational resource that can be continuously divided among stored items (Ma et al., 2014). In these accounts, adding items reduces the precision of each representation, leading to broader error distributions and increased confusability, a pattern that fits data from visual working memory tasks using continuous colour wheels or orientation reports (Luck & Vogel, 2013; Ma et al., 2014). These models challenge strict fixed-slot capacity views by showing that graded resource allocation can fit empirical data as well or better than models positing a constant item limit (Luck & Vogel, 2013; Schneegans & Bays, 2016).

Ericsson and Kintsch’s (1995) long-term working memory theory complements these capacity accounts by focusing on skilled performance in domains such as chess, mental calculation, and reading. They argue that experts encode incoming information into structured retrieval cues stored in long-term memory, allowing rapid access that functionally extends working memory capacity for familiar material without altering basic core limits (Ericsson & Kintsch, 1995). This framework predicts that apparent working memory capacity can be task- and domain-dependent, increasing with the development of domain-specific retrieval structures and strategies (Adams et al., 2018; Ericsson & Kintsch, 1995).

Individual differences and controlled-attention models

Engle and colleagues reframed working memory capacity as an index of executive attention—the ability to maintain task-relevant information and goals in an active, accessible state in the face of interference and distraction (Engle, 2002; Kane & Engle, 2002). In this controlled-attention view, complex span tasks (e.g., operation span, reading span) are demanding not just because they require storage but because they test the ability to prevent goal neglect, resist proactive interference, and recover from lapses when attention is captured by salient but irrelevant information (Engle, 2002).

This approach helps explain why working memory measures correlate strongly with fluid intelligence, reasoning, and high-level comprehension: individuals with better executive attention can maintain task goals and relevant representations despite distraction and interference, leading to better performance across a wide range of cognitively demanding tasks (Kane & Engle, 2002; Unsworth & Engle, 2007). Just and Carpenter’s (1992) capacity theory similarly conceptualised working memory as a limited mental resource supporting both storage and processing during language comprehension, suggesting that individual differences in capacity determine how many elements can be simultaneously activated and integrated (Just & Carpenter, 1992).

Recent commentaries highlight that individual-differences approaches have forced more precise task analysis, revealing that many “working memory” tasks draw on overlapping but not identical processes such as processing speed, inhibition, shifting, and long-term retrieval (Adams et al., 2018; Miller-Cotto & Gordon, 2024). This has spurred interest in decomposing variance in working memory tasks into separable components (e.g., primary memory, secondary memory search, attention control), rather than treating capacity as a single monolithic construct (Unsworth & Engle, 2007).

Cognitive architectures and neurocognitive implementations

Cognitive architectures such as ACT‑R and Soar embed working memory as part of broader, mechanistic theories of the mind in which perception, action, memory, and decision-making are integrated (Anderson et al., 2004; Newell, 1990). In ACT‑R, for instance, working memory corresponds to a set of specialised buffers (e.g., goal buffer, retrieval buffer) that temporarily hold chunks of information, interfacing between procedural knowledge (production rules) and long-term declarative memory (Anderson et al., 2004). Soar, in turn, represents working memory as a dynamic set of symbolic elements that describe the current state, which production rules inspect and modify to generate behaviour (Newell, 1990).

From a neurocognitive perspective, computational models have linked working memory to persistent activity and attractor dynamics in prefrontal and parietal networks, as well as to more transient, activity-silent mechanisms based on short-term synaptic plasticity (Funahashi, 2017; Riley & Constantinidis, 2016). Electrophysiological and imaging evidence shows that representations can be maintained either through sustained firing in fronto-parietal networks or through patterns that can be reactivated by cues, suggesting multiple neural mechanisms underpinning what is behaviourally labelled “working memory” (Riley & Constantinidis, 2016).

Logie et al. (2022) argue that these neurocognitive findings can be reconciled with psychologically oriented models by treating working memory as a collection of domain-specific temporary stores and control processes that interact dynamically, rather than a single, centralised buffer (Logie et al., 2022). In this integrated picture, domain-specific representational systems, attentional control, interference processes, and long-term knowledge all contribute to short-term cognitive performance, and different theoretical traditions can be seen as emphasising different levels of explanation or components of this larger system (Adams et al., 2018; Logie et al., 2022).

Integrative perspectives and current directions

Contemporary integrative accounts suggest that many apparent conflicts between models—multicomponent vs embedded-processes, slots vs resources, storage vs attention—stem from differences in research questions, task paradigms, and levels of analysis rather than direct empirical incompatibility (Adams et al., 2018; Logie et al., 2022). For example, a multicomponent architecture with domain-specific buffers can be understood as a functional description at one level, while interference-based resource allocation and attention-focused control models offer more mechanistic or computational accounts at other levels (Baddeley, 2012; Logie et al., 2022).

Recent work emphasises integrating theories by mapping where they sit along key continua such as modularity (domain-general vs highly modular), degree of reliance on attention, and whether the theory is aimed at group-level performance or individual differences (Adams et al., 2018). Logie et al. (2022) propose that progress will be faster if researchers explicitly acknowledge these levels of explanation, use converging methods (behavioural experiments, neuroimaging, computational modelling), and design tasks and analyses that can differentiate between competing mechanisms rather than only between verbal labels for models (Logie et al., 2022). On this view, working memory is best conceptualised as a family of interacting systems and processes that together support the temporary maintenance and manipulation of information, rather than a single, unitary “box.” (Baddeley, 2012; Logie et al., 2022).