Not sure what they mean, if all the code and model data are the same, it's won't forget anything encoded in the model. Of course the actual post-training contex

Not sure what they mean, if all the code and model data are the same, it's won't forget anything encoded in the model. Of course the actual post-training context window is really short, it won't remember anything at all that falls out of the context window.