Establish whether memory content improves performance beyond constant input

Establish whether pooled document memory provides a performance advantage over a memory module trained with the same constant input for every document under target-only training.

Background

The constant-input control uses the same memory architecture but receives no document-specific facts, allowing the study to distinguish benefits of a learned input from benefits attributable to document content. Although pooled memory outperforms the local adapter under target-only training, its advantage over the constant-input model is not resolved by the reported confidence interval.