Attention is a lookup. To look something up you must have kept it. That is the KV cache, and it is why serving a long-context model costs what it costs.
The usual responses are to compress the cache, evict from it, or approximate the lookup. All three concede the premise: that the model must, somewhere, hold every token it has seen.
Sponor's PSSR does not hold them. State is a field, updated in place. Past tokens are not stored; they are heard. Their influence persists in the field's phase, and decays the way influence actually decays.
This is either the right idea or an elaborate way of forgetting. We are far enough in to say which failure modes are real and not yet far enough to say the thing works.
loading=lazy
fetchpriority=auto
/>