A frozen transformer learned that wombats produce cube shaped droppings and still knows after cold reload [R]

A transformer with a separate, isolated memory buffer. Backbone frozen. 300 gradient steps on the memory weights only: Query Prediction p "wombats produce cube-shaped" droppings 0.9997 "kaniva gets hot in" su (summer) 0.9998 "Lions Club president eats" V (Vegemite) 0.9990 Save, kill process, cold reload, query again. Same result. 20 unrelated facts encoded jointly: 20/20 correct, median p = 0.997. Two subjects encoded simultaneously with cross-contamination < 0.03. The mechanism: BDH ( Kosowski, arXi:2509.26507 ) computes a co-activation outer product at every token step and discards it.