A simple curriculum turns recurrent recall from chance to perfect
Fixed-state models do not hit a hard capacity wall in associative recall. In matched experiments, a short convolution explains roughly +0.5 recall, while a distance curriculum pushes performance from 0.021 to 1.000 without changing the architecture. The catch is training stability: measured-accuracy gating raises long-context lock-in from 0/6 to 6/6 at length 512, suggesting interference and optimization, not recurrence itself, are the real bottlenecks.