Attractor

A dent in the landscape that nearby states roll into — the shape a remembered pattern makes.

Like you're twelve

Roll a marble across a lumpy mattress. Wherever you release it, it ends up in one of the dips. Release it slightly differently and it still ends up in the same dip. The dips are not written down anywhere; they are just the shape the mattress has, and the shape came from what has been lying on it.

Remembering works like that. Your query is the marble. The pattern you get back is the dip.

Precisely

An attractor is a stable state of a dynamical system: a point that a whole neighbourhood of nearby states converges to. In an associative memory, stored patterns are the attractors, and their basins are the region of inputs that settle onto them.

This is what makes recall different from lookup. A lookup returns the nearest stored row; a settle returns the state the input converges to, which can be a stored pattern or the shape a family of stored patterns agree on. How far the input had to travel to get there is the fidelity of the answer.

Related

Hopfield network · Associative memory · Fidelity · Cleanup memory