Key answer
A reusable preference is a choice you want applied to a class of future tasks. Choices determined by one project, audience or temporary limit should retain that scope. One decision or satisfied response need not become a permanent rule.
Full explanation
Ask whether you would still want it for another task. “Use Simplified Chinese for everyday explanations” may transfer across tasks; “retain English product names in this project” is project-specific; “no formulas for this children's explanation” is a task exception. Budgets, dates and versions are updateable state.
Products may distinguish these scopes. Claude gives projects separate memory spaces. OpenAI recommends keeping continuously enforced team rules in explicit locations such as project documentation, with memory as an auxiliary recall layer. These support separating personal preferences, project rules and task exceptions. Claude project memory, OpenAI memory
State conditions and exceptions together:
Lasting preference: begin ordinary explanations with a concise conclusion and expand with evidence when needed. Project rule: retain references for every tutorial answer. This task's exception: show only conclusions in the demo; keep full evidence in handouts. Do not apply that exception to other tutorial pages.
Distinguish a preference change from a temporary exception. “Give detail by default from now on” replaces an old brevity preference; “explain this answer in detail” changes only this output. LongMemEval separately evaluates updates and temporal reasoning, reminding us to check correct replacement, not recall alone. Section 3.2
Periodically inspect memories/rules, remove obsolete entries and correct overgeneralizations. Writing in another person's voice, fictional settings and AI proposals should not become lasting facts or preferences about you.