The Impossibility of a Perfect Memory Token

To get a truly continuous personal agent, we’d love a model with infinite memory. A tempting idea is to fold long histories into a single perfect memory token without changing any future logits. Under softmax attention, however, this is mathematically impossible.