The running representation a transformer carries from layer to layer, which each block adds to rather than replaces.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
A transformer does not hand a fresh answer from one layer to the next. It carries a single vector per token, the residual stream, and every block reads that vector, computes a small adjustment, and adds the adjustment back. Nothing is overwritten.
That additive structure is what makes very deep stacks trainable, and it is also why interpretability research treats the stream as a shared workspace: different blocks can write features into it that a much later block reads out.