Drag the orange query. Line thickness = the score s_i = β⟨x, ξ_i⟩ turned into a softmax weight; each memory's dot grows with its weight. The teal star is the next state x_next = Σ_i w_i ξ_i — the weighted barycenter, always inside the convex hull of the memories. Press step to move the query there (one attention update); play iterates until it settles into a single memory. Raise β and the very first step lands much closer to one prototype.