Preprint proposes near-constant-memory long-context recall method

A preprint proposes reconstructing document facts from residual vectors in a language model's feed-forward-layer activations, rather than keeping the original document in context. The authors say the method maintains near-constant GPU memory use as context length rises and requires no additional training or weight fine-tuning. In their experiments, they report answering single-fact questions in two-million-token story contexts where prior methods failed. The paper frames the approach as a response to memory usage that otherwise grows with input length.