Comment by dist-epoch

> latent space representation of the conversation

and how would you load that back into the model? they are token-in, token-out, plus the KV-cache which is derived from token-in