Working memory: the number design quotes and the memory it supports

A design rule is often a number in disguise. The number most often quoted in interface work is about seven. Miller (1956) reviewed absolute-judgement and immediate-memory experiments. He concluded that the unaided observer is severely limited in what can be received, processed and remembered, but that recoding, further stimulus dimensions and mnemonic devices break that bottleneck. The figure entered practice as a fixed boundary, although what Miller measured was a span that shifted with the material and with the number of dimensions in use.

This unit argues that the capacity limit is real but small and variable between people, and that its design consequence is a shift of load onto the environment rather than a smaller number on a screen. The units before it asked what attention selects and how it splits; this one asks what the selection holds. The strongest counter is that the limit may not be a count: a resource account makes it a fixed quantity distributed across items, so one number misleads (Ma et al., 2014), and the benefit of offloading is assumed rather than measured (Scaife and Rogers, 1996). The steps are these: the number shrank as measurement tightened; the limit falls on integrated objects; it is not storage alone but also the control of attention; and memory is shaped by the statistics of the environment.

The number design quotes shrank as the measurement tightened

Cowan (2001) begins from Miller. The seven was meant, in Miller's own framing, as a rough estimate and a rhetorical device rather than a real capacity limit. He then imposes boundary conditions, which is the step that matters: the limit is observable when overload holds chunks to single items, when recoding into larger chunks is blocked, and in performance discontinuities. Under those conditions Cowan finds a single central limit averaging about four chunks, where others had put it at three to five.

Cowan (2010) restated the result as a central store of three to five meaningful items for young adults. Cowan (2022) then reanalysed a word-list study, arguing that the limit covers the binding of an item to its trial as well as to a serial position, so the unit the limit applies to is still being specified. A rule that quotes seven, or four, without those conditions misreports the evidence it rests on.

Work the number back to what was measured. Miller (1956) ran two kinds of experiment. In absolute judgement a listener names the value of a single stimulus on one dimension. The number of reliably distinguished categories settles near seven. In immediate memory a list is recalled, and the span again lands near seven for unrelated items. The two coincide by accident of material, and recoding breaks one but not the other. The case shows why the number is a coincidence of two spans rather than a capacity, and not that either span is fixed across tasks.

A limit on objects, and a limit that varies between people

Luck and Vogel (1997) established what the limit falls on. Observers retained only about four colours or four orientations at a time, yet they retained both the colour and the orientation of four objects, and objects defined by a conjunction of four features were retained as well as single-feature objects. Sixteen features were carried when distributed across four objects, so the limit is on integrated objects rather than on features. That is the empirical form of the claim that a chunk is something the user has made, and not something the display supplies.

Vogel and Machizawa (2004) recorded an electrophysiological index of maintained items. Its amplitude tracked the number of objects held, plateauing at a limit set by each person, with capacity ranging from about 1.5 to about 5 objects. The average of four summarises a distribution rather than any particular user.

The limit is not storage alone

Baddeley and Hitch (1974) gave the architecture behind the number. A concurrent task showed that a limited short-term store is required for reasoning, comprehension and learning, and they proposed a system with separable parts. Baddeley (2003) reviewed the model and its revisions, including the episodic buffer added to bind information across modalities.

Engle (2002) presses on what the number is a proxy for. On his account capacity is separable from short-term memory and represents a domain-free limitation in the control of attention. Individual differences show up on antisaccade, Stroop and dichotic-listening tasks, and they matter most where interference retrieves a response tendency that conflicts with the current task. If part of the limit is control, an interface can spend capacity by inviting the wrong response, and not only by presenting more items.

Work the control limit on a task that holds nothing. In an antisaccade task a dot appears to one side, and the instruction is to look away from it. No item has to be remembered, yet performance differs between people. The case shows a limb of the limit that an item count misses. It does not show that storage and control are independent, since the same span tasks measure both.

The limit is a fit to the world

Work the environmental check. Anderson and Schooler (1991) asked why memory has the shape it does, and answered that it is adapted to environmental regularities. They counted how often specific items recur in newspapers, in speech to children and in a person's email, and compared those statistics with the probability that the memory is retrievable. The two follow the same relationships to frequency, recency and pattern of exposure, so the shape of memory tracks the shape of the world it serves. A screen that displays an item therefore removes a demand the memory system was not adapted to carry, while a screen that requires recall imposes one. The case shows why recognition is cheap and recall is not, and not that any single display will benefit.

The design consequence is offloading, not trimming

Kirsh and Maglio (1994) showed that people offload deliberately. Analysing Tetris play, they distinguished pragmatic actions, which move the player closer to a goal, from epistemic actions, which change the world to uncover information hidden or hard to compute in the head. A standard model could not account for players' performance without that second function.

Zhang and Norman (1994) showed that the external part of a representation is not a neutral surface. Across four experiments on the Tower of Hanoi, changing how the problem was represented externally changed the representational properties of the task. Difficulty therefore belongs to internal and external representation together.

The benefit is not automatic. Test that the shift helps, because the evidence supports the mechanism rather than the outcome.

Contested question: slots, a shared resource, or attention control?

Three positions are live. The discrete-limit position holds that working memory holds a small number of objects or chunks, four in the ordinary case, with a sharp boundary (Luck and Vogel, 1997; Cowan, 2001). The resource position holds that it is a limited resource distributed flexibly across the items maintained. On that view quality rather than quantity of representation sets performance, and performance declines smoothly with load (Ma, Husain and Bays, 2014). The attention-control position holds that the measured capacity is a domain-free limit on the control of attention. That is why it predicts performance across tasks that share no material (Engle, 2002).

The evidence that would settle it is a dissociation between quantity and precision. A discrete limit predicts a step, with a few objects held at full fidelity and further items dropped. A resource account predicts a smooth decline in the precision of every item as set size rises, with no discontinuity, and Ma and colleagues argue their delayed-estimation data take that shape. The discriminating study varies set size and measures precision separately from guessing. It fits both model classes to the same data, and reports the fit and the free parameters rather than only the better-fitting curve.

While the dispute is open, a designer can act on what all three positions allow. Reduce the number of items a user must hold at once, and show each item's value rather than requiring recall, which matters most for a reader with low literacy. Keep the required precision low wherever an item must be held, because every account agrees that precision and count trade off.

The positions may not be rivals. A discrete cap on a flexible allocation would fit both. They also differ in the strength of their evidence. The discrete and resource accounts each fit the same delayed-estimation data, so that fit does not choose between them, while the attention-control account rests on correlations across tasks and is the weakest.

Where the argument is heading

Two moves are visible in recent work. The first specifies the unit of the limit rather than overturning the number: Cowan (2022) adds the binding of an item to its trial to the binding of an item to its position. The second moves measurement from how many to how precisely, Ma and colleagues (2014) treating the limit as an allocation of a resource whose distribution sets the fidelity of each item. For design that replaces a count with a question: not how many fields to place on a form, but how precisely the user must hold each item.

The sources this unit rests on

• Anderson, J. R. and Schooler, L. J. (1991). Reflections of the environment in memory. Psychological Science, 2(6), 396-408. doi:10.1111/j.1467-9280.1991.tb00174.x Compared the availability of memories for specific items against the statistics of newspapers, parental speech and electronic mail. Both showed the same relationships to frequency, recency and pattern of exposure, which grounds the unit's claim that memory is adapted to the environment.
• Baddeley, A. D. and Hitch, G. (1974). Working memory. Psychology of Learning and Motivation, 8, 47-89. doi:10.1016/s0079-7421(08)60452-1 Used a concurrent task to show that a limited short-term store is required for reasoning, comprehension and learning, and proposed a multicomponent system. It is the architecture behind the capacity claims, making working memory a system rather than one box.
• Baddeley, A. (2003). Working memory: looking back and looking forward. Nature Reviews Neuroscience, 4(10), 829-839. doi:10.1038/nrn1201 Reviews the model and its revisions, including the episodic buffer and the binding problem it addresses. The review states what the model explains and leaves open, which is why the unit treats the architecture as a working account.
• Cowan, N. (2001). The magical number 4 in short-term memory: a reconsideration of mental storage capacity. Behavioural and Brain Sciences, 24(1), 87-114. doi:10.1017/s0140525x01003922 Argues that Miller's seven was a rough estimate and a rhetorical device, imposes boundary conditions that block recoding, and finds a single central limit averaging about four chunks. It is the unit's evidence that the quoted number shrank once the measurement was tightened.
• Cowan, N. (2010). The magical mystery four: how is working memory capacity limited, and why? Current Directions in Psychological Science, 19(1), 51-57. doi:10.1177/0963721409359277 Restates the limit for a general readership as a central store of three to five meaningful items for young adults. It shows the figure is stable enough to teach while its explanation remains contested.
• Cowan, N. (2022). Item-position binding capacity limits and word limits in working memory: a reanalysis. Journal of Cognition, 5(1), article 3. doi:10.5334/joc.193 Reanalyses a word-list study that had attributed capacity to item-position binding alone, arguing for a word-item limit where an item must be bound to its trial. It shows the unit of the limit is still being specified.
• Engle, R. W. (2002). Working memory capacity as executive attention. Current Directions in Psychological Science, 11(1), 19-23. doi:10.1111/1467-8721.00160 Argues that capacity is separable from short-term memory and represents a domain-free limitation in the control of attention, supported by antisaccade, Stroop and dichotic-listening results. It licenses the unit's claim that an interface can spend capacity through interference rather than item count.
• Kirsh, D. and Maglio, P. (1994). On distinguishing epistemic from pragmatic action. Cognitive Science, 18(4), 513-549. doi:10.1016/0364-0213(94)90007-8 Analysed Tetris play and showed that players rotate and translate pieces to uncover information rather than to advance towards a goal, and that a standard model could not explain their performance without this second function of action. It is the evidence that offloading is a strategy users already pursue.
• Luck, S. J. and Vogel, E. K. (1997). The capacity of visual working memory for features and conjunctions. Nature, 390(6657), 279-281. doi:10.1038/36846 Found that observers retain only about four colours or orientations, yet retain both features of four objects and hold sixteen features across four objects. It fixes the limit on integrated objects rather than features.
• Ma, W. J., Husain, M. and Bays, P. M. (2014). Changing concepts of working memory. Nature Neuroscience, 17(3), 347-356. doi:10.1038/nn.3655 Reviews behavioural and neural evidence that working memory is better described as a limited resource distributed flexibly among items, so that quality rather than quantity determines performance. It is the rival position in the unit's contested question.
• Miller, G. A. (1956). The magical number seven, plus or minus two: some limits on our capacity for processing information. Psychological Review, 63(2), 81-97. doi:10.1037/h0043158 Reviewed absolute-judgement and immediate-memory experiments, concluding that the unaided observer is severely limited in what can be received, processed and remembered, while recoding and mnemonics break the bottleneck. It is the origin of the number design quotes.
• Scaife, M. and Rogers, Y. (1996). External cognition: how do graphical representations work? International Journal of Human-Computer Studies, 45(2), 185-213. doi:10.1006/ijhc.1996.0048 Critiques the literature on graphical representations through four representative studies and concludes that little is known about the cognitive value of any of them. It is the unit's caution that the offloading benefit must be demonstrated rather than promised.
• Vogel, E. K. and Machizawa, M. G. (2004). Neural activity predicts individual differences in visual working memory capacity. Nature, 428(6984), 748-751. doi:10.1038/nature02447 Recorded an electrophysiological index of maintained items that plateaued at a limit set by each person, with capacity ranging from about 1.5 to about 5 objects. It makes the average of four a summary of a distribution rather than a property of every user.
• Zhang, J. and Norman, D. A. (1994). Representations in distributed cognitive tasks. Cognitive Science, 18(1), 87-122. doi:10.1207/s15516709cog1801_3 Ran four experiments on the Tower of Hanoi showing that changing the external representation changes the representational properties of the task and hence its difficulty. It is the evidence that the design consequence of the limit is to move load into the display.