Four ±1 patterns are stored. Load one into the query, then corrupt it (flip pixels by clicking, or hit the button). Press retrieve and one modern-Hopfield step — xnext = Σi softmax(β⟨ξi, x⟩)i · ξi — snaps the query back to the stored pattern it most resembles. With well-separated patterns the recovery is essentially one step: this is exactly content-addressable memory.