Persistent context is an assumption applied to every future conversation. Review it like any other assumption.
Do now: Open your saved instructions and memory today and read them as if you had not written them.
Never store anything you would not want retrieved in an unrelated conversation.Saved instructions and memory make these tools far more useful and introduce a failure mode that is invisible by design: you stop seeing the context, so you stop noticing when it goes wrong.
What belongs, and what does not
Good candidates are stable and impersonal: your role and sector, the formats you prefer, your language and spelling conventions, the level of explanation that suits you, and standing constraints such as 'always show your uncertainty' or 'never invent citations'.
Poor candidates are anything sensitive, anything about other people, and anything likely to change soon. Client names, financial details, health information, and specifics of a project ending next month all belong in the individual conversation where they are needed, not in permanent storage.
The test is simple: would you be comfortable if this appeared in an unrelated conversation six months from now, possibly in front of someone else? If not, keep it out.
If you would not want it surfacing in an unrelated conversation, it does not belong in memory.
The failures stored context causes
Staleness. A stored fact that has changed keeps being applied. Your role, your organisation, your project, your priorities — all of these move, and none of them announce the move to a saved instruction.
Bleed. Context from one project seeps into another. A tone set for internal notes appears in a client proposal; an assumption from one client's situation shapes advice about a different one. This is subtle and easy to miss precisely because the output remains fluent.
Over-fitting. Heavy instructions can narrow output more than you intended. If you have specified format, length, tone, structure and vocabulary, you may find you cannot get a genuinely different perspective when you need one.
Accumulation. Instructions added over a year contradict each other, and the behaviour becomes hard to predict. Nobody rewrote them; they just piled up.
Review on a schedule, and keep separations
Put a recurring reminder — quarterly is reasonable — to read your saved instructions and memory as a stranger would. Remove anything stale, anything sensitive, and anything you no longer recognise as your intention.
Use separate projects or spaces for genuinely separate work, especially where clients or confidentiality are involved. Shared context between two clients is a professional problem regardless of whether anything is technically disclosed.
When output starts feeling subtly wrong in a way you cannot place, check the persistent context first. That is a far more common cause than a change in the model, and it is the first thing worth eliminating.
When output feels subtly off, read your saved instructions before blaming the tool.