I wanted to suggest the UI feature of being able to fork a pre-existing conversational thread, along with the pre-existing context and prior attachments. This can save time if I intend to run a bunch of queries from a certain knowledge base. It would also be more lightweight than creating a custom GPT, much like an ad hoc request. And as it stands today, this ability to fork would allow me to use more arbitrary models (rather than a custom GPT which seems to be fixed to a single model). Compared to a global memory for all conversations, this forking ability would be pinpointed to certain use cases.
So the intended/suggested workflow, if implemented, would be a button to push that forks the existing conversation. An enhancement to this would be the ability to retroactively fork the conversation from an earlier point in time (this one might be more expensive in terms of resource use).
The implementation of this could also leverage copy-on-write technologies to save on resources. If the internal state of the conversational thread can be stored as inert data (until the next request comes in), that state could be cloned conceivably without requiring lots of GPU computation.