What's this about?
LatentPress (47 upvotes) compresses conversational history into continuous memory tokens a frozen decoder reads directly — no text reconstruction, 4–16× compression, and it *beats uncompressed* evidence on LongMemEval.
What exactly is a continuous memory token?
Instead of summarizing history as text (lossy, verbose), history is packed into dense vectors the model consumes natively — memories stored in the model's internal language rather than English.
Why is this worth noting?
Compression that outperforms the original suggests distillation, not just shrinkage — the compressed form may strip distraction. Alongside Astra's notes-across-windows, it's the same race: agent memory beyond the context window, attacked from the representation side.