Current replay buffer uses memory as the storage option.
This approach is limited by the size of available memory and is problematic if the dataset is large, as in the case of offline learning.
Providing a disk storage option might be one useful feature for this case.