Clarification on carry variable used during training

[jren03]

Would you mind explaining the purpose of the carry variable used during training? From what I understand, it's a tuple of latent state and latent action.

From this section, it seems like prevlat is replaced with the context from the first sample of each batch. However, why does prevact or carry[1] not replaced in this case? Asked differently, if I were to alternate sampling from two different replay buffers during training, would I need need two different carry variables, or does a shared one suffice since it is getting replaced by context anyway?

Thank you in advance for the help!