Intelligent vector storage - recursive loop

[xhlsa]

Could there be a recursive loop added where the LLM can occasionally resort it's memory cloud after reviewing the conversation to keep key concepts that are the most relevant in the current topic?

I could see this later going to a more refined system and space for the AI to formulate the thought.

Basically, giving the LLM permission to recompile the vector storage but have it print the refinement at first to see what parameters should be tweaked. That will likely be a huge task: getting the LLM to truly intelligently keep a short term memory.