At the 1952 dedication of the Institute for Advanced Study computer, AI was framed as a machine‑amplified von Neumann: immense speed, breadth and symbolic memory. When an AI solves a hard math problem we often credit new intuition, but the real boost may simply be a vastly larger working memory.

Human working memory is tiny. We can hold only a handful of unfamiliar elements at once, relying on scratch paper, notation and chunking to extend it. Even experts compress information into single conceptual objects, yet the biological limit remains.

Research shows this bottleneck matters. Alloway and Passolunghi (2011) found working‑memory measures added unique variance to math skill beyond verbal ability. A 2010 longitudinal study found early working memory predicted literacy and numeracy better than IQ. Blankenship et al. (2015) and Friso‑van den Bos et al. (2013) reached similar conclusions.

A language model’s context window acts as a gigantic external notebook, storing the entire problem, equations, and its own reasoning. This scale advantage lets it keep dozens or hundreds of symbols in view, turning symbolic math into a largely externalized, stable process—though imperfect retrieval remains an issue.