Inference APIs allow users to send input and receive output, facilitating conversation archiving and replaying. However, challenges such as prompt cache storage on external GPUs, differing tokenization between models, and intentional sampling unreproducibility complicate session portability.