Current conversation context, saved history and long-term memory serve different purposes. The conversation contains the active discussion; history preserves the work; memory holds relevant information that may help a later task. You should be able to see and correct what is being reused.
Separate a preference from a fact
“I prefer short labels” is a drawing preference. “This project uses a queue” is a project fact. Neither should become an unsupported assumption about every future diagram. A remembered item needs a relevant scope and a source that you can inspect.
Use explicit current instructions for requirements that must apply to this task. A new instruction can override a prior preference without forcing every previous diagram to change.
Correct or exclude information
Review remembered items when the agent reports an update or when an old fact appears in a new conversation. Correct information that has changed, delete information that should no longer be reused and exclude a source conversation when its context should remain local.
Deleting a visible memory item should not leave the same fact active in an outdated summary. Keep the source and the derived memory aligned when reviewing your information.
Respect project boundaries
A project-only scope keeps its material separate from account-wide work and other projects. Tightening that boundary should also apply to prior summaries and retrieval results, not only new messages.
Named brand colors are managed separately from automatic memory. Requesting a specific palette does not give the agent permission to reuse unrelated client conversations.
Make it about your work
Edit this starting prompt, then send it when you are ready.