3M-param transformer learns via fast-weight memory at inference

The fast-weight memory bank lets the model write and read vectors as weights during forward pass, requiring no test-time training. The method is fully reproducible and published as a preprint by an independent researcher.