I think it’s possible this might just require reworking the transformer architecture a bit to give it an inherent awareness of the contents of its attention. Kind of a working memory workspace of sorts.

I.e. Instead of limiting autoregression to the token level, you introduce a persistent compressed global workspace latent memory vector that is fed back into the self-attention mechanism at every layer or every token step, allowing the network to attend to its own prior attentional states before computing the next token. Obviously that’s going to involve some compression steps.

Trouble is… I think the architecture there is much simpler a tweak than figuring out how to train it.

…that’s likely to just destabilize training for not much if any gain at first. You’re probably gonna have to resort to some really clever (and currently missing) tricks to figure out how to train the network to actually use that feature.