Working memory (WM) capacity, the number of items that can be simultaneously maintained in an active, accessible state, is one of the most fundamental constraints on human cognition. Since George Miller's (1956) influential "magical number seven" paper, mathematical models have sought to formalize the nature and origin of these capacity limits, with modern estimates converging on approximately 3-4 items for unstructured material.
Slot Models
Slot models (Luck & Vogel, 1997; Cowan, 2001; Zhang & Luck, 2008) propose that WM consists of a fixed number of discrete "slots," each capable of holding one item (or chunk). Cowan's K formula estimates the number of filled slots from change-detection performance:
where S is the set size, H is the hit rate, and F is the false alarm rate. This formula assumes that if an item is in a slot, change detection is perfect; otherwise, the observer guesses. Typical estimates yield K ≈ 3-4 for healthy young adults. The slot model predicts that increasing set size beyond K should produce a plateau in performance, as no additional items can be stored.
Resource Models
An alternative view holds that WM capacity reflects a continuous, divisible resource that is distributed across items (Bays & Husain, 2008; van den Berg et al., 2012). As more items are stored, each receives less resource, leading to noisier representations:
Resource models predict that recall precision should decline continuously with set size, with no sharp capacity limit. Empirically, both discrete (slot-like) and continuous (resource-like) patterns have been observed, leading to hybrid "slots + averaging" models where each slot holds a representation of variable precision.
Individual Differences and Cognitive Ability
WM capacity is one of the strongest predictors of higher-order cognitive abilities including fluid intelligence, reading comprehension, and reasoning. Engle and colleagues have shown that the correlation between WM capacity and intelligence is mediated by the ability to maintain goal-relevant information in the face of distraction, which can be formalized using computational models of attentional control. The mathematical relationship between K and general cognitive ability (g) has been estimated at r ≈ 0.5-0.7 across studies.
Neural Basis and Computational Constraints
Computational models suggest that the 3-4 item limit arises from fundamental constraints on neural oscillatory mechanisms. The theta-gamma coding framework (Lisman & Idiart, 1995) proposes that individual items are represented by gamma-frequency oscillatory cycles (~40 Hz) nested within theta-frequency cycles (~6 Hz). The number of gamma cycles per theta cycle (approximately 6-7) sets an upper bound on capacity. Accounting for noise and interference reduces the effective capacity to 3-4 items, matching behavioral estimates.
WM capacity is limited in terms of chunks, not raw features. Through chunking, experts can hold vastly more information in WM than novices. A chess master's "chunk" might contain an entire board configuration. This means that capacity models must be paired with a theory of chunk formation to make accurate predictions about performance in knowledge-rich domains.