Keep the test small. If you change your prompt, source material, and memory settings together, you will struggle to tell which change helped.
Find the memory controls your account has
Open Settings > Personalization > Memory. Controls vary by account and may include a memory summary or saved memories. You can also ask ChatGPT what it remembers. The summary is a high-level view and may omit information it can reference. OpenAI's Memory guide
Look for the actual mismatch: an old role, an audience you no longer write for, or a preference that was appropriate for one task. A note saying you write marketing copy may explain a sales tone; it does not prove that memory caused every unwanted phrase.
Correct one preference in plain language
Here is an original example request you can adapt:
My role has changed. I now prepare internal operational notes for
colleagues, not sales copy for prospective customers.
For those notes, prefer direct wording, specific actions and named
uncertainties. Do not assume I want promotional language.
Please correct any remembered preference that says my usual audience
is software buyers. Tell me what you changed, and identify anything
you could not change so I can inspect the available settings.
OpenAI documents correcting remembered information in chat and editing the summary or saved memory where those controls are available. Use the controls you can inspect rather than relying only on a reassuring reply. Correct remembered information
Avoid putting a temporary deadline into a lasting preference. “My next report is due Friday” belongs with the report. “For internal reports, state the decision needed before the background” is a reusable writing instruction.
Test with a task whose answer you can judge
Start a fresh ordinary chat and give it a brief, permitted fixture:
Draft a short internal note from these fictional facts:
The team will test a new equipment checkout process on Tuesday.
Two volunteers will take part. The shift lead must confirm the test
time. No results have been collected yet.
The note should tell colleagues what happens next and what is still
unconfirmed. Do not add benefits or results that these facts do not show.
You should be able to identify the trial, the unresolved time, and the person who must confirm it. A sentence promising faster checkouts fails that test because no result has been supplied. This fixture tests the draft's content; it does not establish exactly which internal context produced it.
If the old tone returns, inspect your current instructions and the material supplied with the request too. A pasted example full of promotional language can pull the draft in that direction. Revise the immediate request to name the audience and desired output.
Forgetting a preference and deleting its source are different actions
Turning Memory off does not delete past chats. Deleting a chat does not necessarily remove a separately saved memory. To remove remembered information, OpenAI says to remove it from memory controls and the relevant sources, which may include chats, Library files, or connected apps. Changes can take time. Memory removal guidance
For sensitive material, follow your organization's data rules and inspect the source locations before taking deletion actions. Do not treat “don't mention this” as proof that the underlying information has been deleted.
Keep a small approved writing reference
Save one short example that fits your current audience. When a draft drifts, compare its actual sentences with that reference: are actions clear, uncertainties named, and results supported? That gives you a practical correction even when personalization is difficult to diagnose.
Optional: review the finished note in NOS
If you already keep an HTML review copy of an internal note, you can read it in NOS and check the wording colleagues will see. Manage ChatGPT Memory in ChatGPT; the review copy does not change those settings.