It sounds like a great idea, but I have reservations:

1. Primarily: I don't trust LLMs to output 100% correct information - I will always second-guess whatever I read there 2. The devil is in the details - condensation naturally has to exclude some of the data, and I'm quite sure I don't trust an LLM to figure out what's important; better yet, there might be something not really all that important in there which really aids in knowledge retention (eg a funny story), which is stripped from the output

It sounds like a neat idea if you're willing to ingest invalid, incomplete knowledge in the hope that it will be sufficient. I guess time will tell.