I have to agree, but I think some people find the joy in finding their solutions in code rather than from code…meaning that for some it is the puzzle solving that is their interest. That can be rather hard to find an analog for, especially if coding is the skill they are good at.

The lack of AI mental model between conversations or even within the conversation is a definite frustration. “No, I didn’t just decide to entirely change everything I’ve planned based on one ambiguous word…”

Your use of a separate session for q&a is actually something I find interesting in the new Granite Switch architecture’s KV Cache Normalization, where the kv cache is essentially frozen while the model performs certain other tasks in order to prevent the mental ‘conversation’ from getting polluted. However, I’m also assuming what you mean by “helps grounding” is that it helps keep the agent grounded in the context by not introducing a bunch of external information, which is certainly beneficial.