Working Memory Explained: Baddeley's Model and Capacity Limits

The Brain's Workspace Has a Size Limit

A surgeon performing a complex procedure holds the patient's anatomy, the current step, potential complications, and incoming instrument cues simultaneously — a cognitive demand that taxes working memory to near its limit. When distracted or fatigued, errors emerge not from lack of knowledge but from information dropping out of this temporary holding system. Working memory is not memory in the archival sense; it is the cognitive workspace where active processing occurs. Its capacity is smaller than most people believe, its architecture is more specialized than a single "mental RAM" metaphor suggests, and its limitations account for a striking proportion of real-world cognitive failures.

Baddeley's Four-Component Model

Alan Baddeley and Graham Hitch introduced the first multi-component working memory model in 1974, replacing the single-unit short-term memory concept from Atkinson and Shiffrin's modal model. Baddeley refined the architecture over subsequent decades, adding a fourth component in 2000. The model remains the most empirically supported framework in cognitive psychology.

The Phonological Loop stores and rehearses verbal and auditory information. It has two sub-components: the phonological store (holds speech-based information for approximately 2 seconds before decay) and the articulatory rehearsal process (subvocal repetition that refreshes items in the phonological store). Evidence for the phonological loop comes from the word length effect (more short words than long words can be held in working memory — because articulatory rehearsal takes longer for longer words) and the phonological similarity effect (sequences of phonologically similar letters like B, C, D, G are harder to recall in order than dissimilar letters like K, W, R, Y).

The Visuospatial Sketchpad maintains and manipulates visual and spatial information. It supports tasks like mentally rotating objects, navigating by map, and tracking the positions of moving objects. Separate neural substrates serve visual content (object identity, color, shape — ventral stream) and spatial information (location, movement — dorsal stream). The sketchpad is disrupted by concurrent tasks involving eye movements or spatial tracking but not by verbal tasks, confirming modality-specific storage separate from the phonological loop.

The Central Executive is the attention control system of working memory — a limited-capacity supervisor that coordinates the two slave systems, focuses and divides attention, updates working memory content, shifts between cognitive strategies, and links working memory to long-term memory. Baddeley drew from Norman and Shallice's Supervisory Attentional System (SAS) model. Damage to the frontal lobes — particularly the dorsolateral prefrontal cortex — produces deficits consistent with a compromised central executive: difficulty task-switching, impaired inhibition of irrelevant information, and inability to coordinate dual-task performance.

The Episodic Buffer (added by Baddeley in 2000) is a temporary storage system that integrates information across the phonological loop, visuospatial sketchpad, and long-term memory into coherent multimodal episodes. It explains how coherent sentences (which require integration of verbal content with semantic long-term memory) can be held in working memory beyond the capacity of the phonological store alone. The episodic buffer requires the central executive to access and is limited to approximately 4 chunks of integrated information.

Cowan's 4-Chunk Capacity: Correcting Miller

George Miller's famous 1956 paper, "The Magical Number Seven, Plus or Minus Two," established 7 ± 2 as the capacity of short-term memory, a figure that became one of the most cited statistics in psychology. Nelson Cowan's 2001 reanalysis, published in Behavioral and Brain Sciences, challenged this figure systematically. Cowan distinguished between the span of apprehension (everything visible or audible in immediate awareness) and the focus of attention — the genuinely active working memory system.

When chunking and rehearsal are controlled for — using rapidly presented items, articulatory suppression to prevent subvocal rehearsal, and items participants cannot form long-term memory associations with — working memory capacity converges on approximately 4 chunks (range 3–5 across studies). The "7" figure largely reflected rehearsal strategies and long-term memory contributions artificially inflating apparent capacity. Cowan's estimate aligns with independent evidence from visual change detection tasks (Luck & Vogel 1997: ~4 visual objects held simultaneously) and from developmental and neurological data showing consistent convergence near 4.

Dual-Task Interference

When two tasks require the same working memory component simultaneously, performance on both degrades. This is dual-task interference. Talking on a mobile phone while driving impairs driving performance even with a hands-free device — not because hands are occupied, but because both tasks compete for the central executive and phonological loop. The impairment is comparable to driving at the legal blood alcohol limit (Strayer & Johnston 2001, University of Utah).

• Two tasks using different components (e.g., verbal shadowing + spatial tracking) can be performed simultaneously with less interference than two verbal tasks
• The bottleneck is the central executive: when two tasks both require attention control, one must wait for the other (the psychological refractory period)
• Expert performance reduces dual-task costs by automating component processes that novices must perform under executive control
• Working memory capacity predicts dual-task performance: higher-capacity individuals show less cross-task interference

N-Back Training: The Debate

The n-back task — monitoring a sequence of items and responding when the current item matches the one presented n steps earlier — became the basis of commercial "brain training" claims after Susanne Jaeggi and colleagues published a 2008 Science paper showing that dual n-back training improved fluid intelligence in healthy adults. The claim that working memory training could improve general cognitive ability (rather than just n-back task performance) generated enormous public interest and a commercial industry.

Subsequent meta-analyses have not supported transfer to general fluid intelligence or real-world cognitive performance. Melby-Lervåg, Redick, and Hulme's 2016 meta-analysis (Perspectives on Psychological Science, analyzing 87 studies) found that working memory training produced robust near-transfer effects (improved performance on other working memory tasks) but no reliable far-transfer to fluid intelligence, reading, math, or attention in healthy participants. The commercial brain training industry's claims — that n-back training improves driving, academic performance, or executive function in everyday life — are not supported by the current evidence base for healthy adults. Clinical applications (ADHD, traumatic brain injury rehabilitation) show more promising but still inconsistent results.