CONCEPT
Content-Addressable Memory
The capacity to recover a complete stored pattern from a partial or corrupted fragment of its content—without knowing its storage address—which John Hopfield modeled as descent through an energy landscape, and which is mathematically identical to the attention mechanism powering every modern transformer.
Ask yourself how you remember a song: not by consulting a numbered address, but because a few notes drift by and the whole thing assembles itself. This is content-addressable memory—recall driven by what a thing
is rather than by where it sits—and for most of the twentieth century it had no mathematics adequate to it.
John Hopfield's 1982 network supplied the mathematics: store each pattern as a valley in a high-dimensional energy landscape, and recall it by releasing a partial fragment as a starting point, letting the system descend to the nearest valley floor. The complete pattern emerges not by lookup but by
relaxation—the fragment falls into the basin of the stored memory and the memory reconstitutes itself. A 2020 result made the lineage between this mechanism and modern AI explicit and exact: the attention update at the heart of every
transformer—take a query, compare it against keys, return a