The Magical Number Four

George Miller, Nelson Cowan, and Why Working Memory Is Controlled Attention

In 1956, the Harvard psychologist George Miller published the most famous paper in the history of cognitive science, opening with the playful complaint that he was being persecuted by an integer: seven, plus or minus two. Reviewing decades of psychophysical and memory experiments, Miller pointed out that two completely different cognitive limits had been blurred together in the popular imagination. When human observers judge a single raw sensory attribute such as acoustic pitch, loudness, or visual position along a line, their channel capacity breaks down around 2.6 bits of information—roughly six distinguishable categories. Immediate memory span holds around seven digits, letters, or words only because human subjects refuse to store raw bits passively. Instead, they rapidly recode incoming strings into familiar multi-item packages that Miller named chunks.

In a classic demonstration at MIT and Harvard, Sidney Smith trained himself to recode long strings of binary zeros and ones on the fly into octal (three-bit) and hexadecimal (four-bit) chunks. That recoding immediately expanded his recall span from nine binary digits to forty consecutive binary digits without growing a single new synapse in his short-term buffer. For forty years after Miller's paper, textbooks treated short-term memory as a passive seven-slot wooden shelf where information sat waiting to be used. Then, in 2001, the University of Missouri psychologist Nelson Cowan synthesized dozens of behavioral and electrophysiological experiments to ask what happens to human capacity when researchers strictly block subjects from silently repeating words under their breath or grouping letters into familiar acronyms like FBI and CIA.

Across every visual and auditory paradigm where rehearsal and chunking were prevented—such as flashing an array of colored squares briefly on a screen or presenting four spoken voices simultaneously—the true capacity of the human focus of attention converged not on seven, but on a strict ceiling of three to four independent items. In Steven Luck and Edward Vogel's 1997 visual change-detection experiments at the University of Iowa, human observers tracked colored squares, oriented bars, and multi-feature objects with near-perfect accuracy up to four items. At five items, performance hit a vertical cliff where parietal brain potentials flatlined completely.

Even more revealing than the four-slot ceiling was the discovery that what psychologists call working memory is not a separate storage box at all. It is simply executive attention turned inward to sustain mental representations while keeping interference out. Throughout the late 1990s and 2000s, Georgia Institute of Technology psychologist Randall Engle tested thousands of adults on complex working-memory span tasks, asking them to solve arithmetic equations while holding unrelated words in mind. He then compared high-span and low-span individuals on elementary attention tests such as the antisaccade task, where a volunteer must resist looking at a sudden flashing light on the left and immediately look to the opposite side of the screen.

If working memory were merely the size of a passive storage bin, someone with a large memory span should have no special advantage in resisting a flashing light. Instead, Engle proved that high-working-memory individuals excelled at every test of attentional control because working-memory capacity is governed by the prefrontal cortex's ability to keep task goals active and block distraction from entering the four-slot workspace. In 2005, Edward Vogel, Maro Machizawa, and Steven Luck confirmed Engle's mechanism directly using scalp electrophysiology. When asked to remember two red bars while ignoring two blue distractor bars, individuals with low working-memory scores showed just as much parietal storage activity as high-capacity observers. Their executive gate had failed and let irrelevant background distractors flood their four slots.