I've been investigating this condensation theory, and at first I did what you suggested, but the problem becomes the size of the context. Either you limit the number of things condensed (reducing the usefulness of the long-term memory because you'll lose the ability to say things like "remember when X?"), or you let it grow until it breaks the model's limit.

I'm currently investigating a long-term storage system using embeddings, and having GPT output a "remember" command whenever it decides it should remember something. There's lots of work to be done to get it just right, but this is an incredibly exciting future, for sure.