A copy layer for language models that finds earlier occurrences of the current context with learned hash tables and copies what came next — with memory linear in sequence length.
