Large language models keep some facts in huge lookup tables beside the main network. Squeezing those tables first erases facts about rare entities. The usual fix, giving more precision to frequent rows, makes things worse.
Each retrieved row is scaled by a gate value from the context, computable from a single-bit approximation. So rows are stored in layers of single-bit residual codes; deeper layers are read only for lookups the gate says matter.
This beats frequency-based and random choices in every comparison, cuts memory 7.8 times, and loses 0.2 points at the higher bit budget tested. Better fact retention at a fraction of the memory, no retraining.