Baddeley's working memory model proposes that short-term information is not held in a single store but managed by four components: the phonological loop for verbal material, the visuospatial sketchpad for images and locations, the central executive that allocates attention, and the episodic buffer that binds features into integrated chunks. Proposed by Alan Baddeley and Graham Hitch in 1974 and revised in 2000, the model replaced the older "short-term store" view. Its core claim is that separate limited-capacity subsystems run in parallel under attentional control, which explains why you can rehearse a phone number while tracking a moving cursor.
The architectural problem the model solved
Before 1974, memory was often pictured as a chain of boxes ending in a single short-term store. That picture failed a simple prediction. If all short-term retention used one store, then doing two short-term tasks at once should collapse performance to near zero. Baddeley and Hitch tested this by asking participants to hold strings of digits while reasoning or reading. Holding three to six digits slowed performance but did not abolish it. In plain terms: the tasks interfered only partially, which implied separable stores rather than one.
The model's answer was modular. Each slave system has its own capacity limit and its own code. The central executive is a controller, not a storage bin, and it supervises the others. The 1974 paper estimated the phonological store at roughly two seconds of spoken material, a limit that later work tied to subvocal rehearsal speed. The key mechanism is that concurrent tasks compete only when they demand the same code, so verbal-plus-verbal collides while verbal-plus-spatial runs fairly cleanly.
Mechanism: how the four components actually operate
The phonological loop has two parts: a temporary acoustic store that decays in about 2 seconds, and an articulatory rehearsal process that refreshes it by silently speaking the items. The word-length effect follows directly. Memory span drops from about 7 items for one-syllable words to about 4.5 for five-syllable words, because longer words take longer to rehearse within the decay window. The phonological similarity effect is the second signature: similar-sounding items (B, C, D, G) are recalled worse than dissimilar ones, since their codes blur in the acoustic store.
The visuospatial sketchpad holds visual appearance and spatial location, often split into a passive visual cache and an active inner scribe. Dual-task studies show that tracking a moving target interferes with spatial retention but leaves verbal span largely intact. Its capacity is usually reported as around 4 objects rather than 7, though estimates vary with task design.
The central executive is the least specified component, and Baddeley has acknowledged this openly. It handles task switching, selective attention, and inhibition of irrelevant material. Evidence links it to frontal lobe function: patients with frontal damage often show preserved span but poor dual-task coordination and frequent perseveration.
The episodic buffer, added in 2000, binds information across modalities and links it to long-term memory. It was introduced because the three original components could not explain how 12- to 15-word sentences are recalled despite exceeding loop capacity. The buffer holds roughly 4 chunks of integrated material, which matches Cowan's widely cited estimate of a 4-item focus of attention.
Quantified capacity: the numbers side by side
The most useful comparison is between the model's components and competing capacity estimates. Miller's classic 1956 figure of 7 ± 2 items describes chunk-based span in a single stream. Cowan's 2001 review argued for about 4 chunks when rehearsal and chunking are controlled. Baddeley's subsystems sit between and below these figures depending on code.
| Component | Typical capacity | Primary code | Decay or limit | Interference signature |
|---|---|---|---|---|
| Phonological loop | 2 seconds of speech; ~7 one-syllable items, ~4.5 five-syllable | Acoustic-verbal | ~2 s trace decay, refreshed by rehearsal | Phonological similarity, word length |
| Visuospatial sketchpad | ~4 objects or locations | Visual-spatial | Decay plus displacement | Spatial tracking, visual similarity |
| Central executive | Not a storage limit; capacity reflected in dual-task cost | Amodal control | Attentional, not decay-based | Dual-task decrements, switching cost |
| Episodic buffer | ~4 integrated chunks | Multimodal binding | Binding failure under load | Cross-modal binding errors |
The reconciliation matters: 7 ± 2 and 4 are not contradictory. Miller counted chunks already formed, while Cowan counted chunks that can be held without rehearsal. The loop's 2-second figure is a decay constant, not an item count, and the 4.5-versus-7 word span gap is the same capacity expressed in different rehearsal times.
Timeline: from single store to four components
| Year | Development | Key claim or finding |
|---|---|---|
| 1968 | Atkinson and Shiffrin multi-store model | One short-term store with a fixed slot capacity |
| 1974 | Baddeley and Hitch dual-task studies | Partial interference implies separate subsystems |
| 1975 | Word-length effect documented | Span scales inversely with spoken duration |
| 1986 | Baddeley's "Working Memory" book | Articulatory loop and sketchpad formalized |
| 1992 | Reading span and frontal patients | Central executive tied to attentional control |
| 2000 | Episodic buffer proposed | Binding explains sentence recall beyond loop capacity |
| 2012 | Time-based resource-sharing accounts | Decay and attention jointly explain forgetting |
| 2020s | Neural and network models | Component boundaries remain debated, buffer role strengthened |
The trajectory shows a pattern. Each addition answered a specific anomaly rather than expanding for its own sake. The buffer is the clearest case: it was added because no existing component could store bound, cross-modal content.
Where the model strains: edge cases and professional caveats
The strongest challenge comes from interference-based accounts. Oberauer and colleagues argue that forgetting is driven by similarity-based interference and overwriting, not passive decay. Under that view, the 2-second trace is not a timer but a byproduct of distractors displacing items. The debate is unresolved, and it matters because decay implies rehearsal can rescue any item, while interference implies some items are simply lost.
A second caveat concerns the central executive. Critics note that a homunculus-like controller risks explaining too little, because it relocates the hard problem of attention rather than solving it. Baddeley's response has been to break the executive into separable functions, but the component remains a placeholder more than a mechanism.
A third edge case is expertise. Chess players recall board positions at near-perfect rates when the positions are legal, far above the 4-chunk limit. Chunking through long-term memory bypasses the buffer's apparent cap, which shows the model describes novices better than experts. In plain terms: capacity limits are real but porous once knowledge is organized.
Misconceptions worth correcting
The most common error is treating working memory as a synonym for short-term memory. Short-term memory is retention over seconds; working memory is retention plus manipulation, which is why the executive is central. A second error is reading the model as a literal brain map. The components are functional descriptions, and neuroimaging shows distributed networks rather than four discrete regions. A third error is assuming the 7 ± 2 span is a hard biological constant. It depends on chunk size, rehearsal strategy, and material, and falls to about 4 when those are controlled, per Cowan's 2001 review.
What to actually do with the model
The model has practical uses in education and clinical assessment, though none are dramatic.
- Present verbal instructions in short, one- to two-syllable chunks and pause for rehearsal, staying within the loop's 2-second window.
- Avoid pairing a heavy verbal task with another verbal task; split modalities when possible, such as speaking while showing a diagram.
- Keep working sets to about 4 unfamiliar items before building chunks, rather than assuming the old 7-item rule.
- Use spatial and visual supports for tasks that strain the sketchpad, and verbal mnemonics for sequences.
- In cognitive assessment, treat dual-task cost as a more sensitive executive measure than raw span.
Limitations and open questions
The model is a framework, not a finished theory. Its slave systems are well supported, but the executive and buffer remain under-specified. The decay-versus-interference dispute is genuinely open, and neural evidence does not cleanly map components onto regions. The strongest current reading is that Baddeley's architecture captures the functional division of labor correctly while leaving the mechanisms of control and forgetting to be refined. For a knowledge baseline, it remains the most useful scaffold in cognitive psychology, provided its limits are stated as plainly as its strengths.