When you are unsure mid-conversation whether to restart, here is the judgment up front: if the topic has not changed, no wrong premise has crept in, and you still need the earlier context, keep going. If any one of those three flips, open a new chat. Do not cling to what you have already produced; pushing on with a wrong premise costs more time.
Signal one: the topic has switched tracks
Going from a weekly report to a resume to a child's essay topic in one thread makes the model carry over tone, role and constraints, and answers start to blend. A simple test: if the new question stands on its own without any history, it is a new track. Open a new chat and restate the needed background in a sentence or two, such as your role, the audience and the tone you want. That is far more reliable than letting the model fish through old messages.
Signal two: the thread is polluted by a wrong premise
If you gave a wrong number or date, or the model answered something wrong confidently early on, it tends to keep reasoning from that false premise. Do not keep correcting it in place: corrections also consume context, and the model sometimes apologizes yet slides back later. The cleaner move is a new chat, pasting a short note with the facts you have verified, and restarting from a clean base.
Signal three: you need a clean start, not memory
For brainstorming, a genuinely different second proposal, or asking the same question again to check stability, memory is a burden: the model reuses earlier conclusions and wording. A new chat gives an independent answer. Conversely, when revising the same draft across rounds or tracking one project, context is an asset; keep going and ask it to restate the current consensus at key points.
A handy move when unsure
Ask the current chat to summarize, in three sentences, the confirmed facts and the open questions. Copy that summary, then decide: if it looks clean and accurate, take it into a new chat; if even the summary is fuzzy or contains things you never said, this conversation is done.
