Your ChatGPT thinks it knows you. The model has been quietly building a stack of "facts" about you for months: which projects you abandoned, what you said about your work in a one-off conversation, that random side hustle you researched in February. Every answer you've gotten in the past month was shaped by that stack.
Most of it is wrong. Stale, contradictory, or pulled from things you'd rather not be cited.
Why this week
OpenAI shipped Memory Sources this week, a panel that for the first time shows you exactly which past conversations and saved memories shaped any given answer. They also rolled GPT-5.5 Instant to all ChatGPT users on May 5. The new default model leans harder on stored memory than its predecessors did, partly to support the 52.5% drop in hallucinations they advertised. If the memory is wrong, the more-confident model is more-confidently wrong. Right moment to clean it up.
The 5-minute audit
Open chat.openai.com in your browser. Click your name in the bottom-left, then Settings, then Personalization, then Memory.
You'll see two sections: Saved Memories (things ChatGPT explicitly remembers) and Chat History (which past conversations it can reference). Most readers will be looking at Saved Memories for the first time. Expect 20 to 200 entries.
1. Scan for stale facts. Anything about projects, employers, or focus areas that are no longer current. ChatGPT may still think you work at a job you left six months ago, or be helping you with a side project you killed. Delete each one with the trash icon.
2. Scan for half-finished context. Memories that capture one specific conversation but read like permanent facts. ("User is writing a novel set in 1850s Boston" from that one creative-writing experiment.) Delete unless still active.
3. Scan for embarrassing one-offs. That time you asked about something private, sensitive, or just outside your usual work. If it would feel weird to have ChatGPT cite it in a business answer, delete it.
4. Add one good memory. Now that you've cleaned house, add one memory that genuinely helps: a brief paragraph on how you want ChatGPT to work with you. Tone, length, decision framework, what to push back on, what to assume. Click "Manage memories", then add it as a new entry.
Who this is for, who can skip
If you've used ChatGPT for more than a month and let memory stay on by default, this applies. Almost everyone reading this falls in that bucket.
You can skip if:
You turned memory off when you first set up ChatGPT (Settings, Personalization, Memory toggle). The audit is a no-op.
You only use ChatGPT casually for one-off questions. The panel is probably empty.
You exclusively use Claude or Gemini. Claude has its own Projects-based context, no equivalent global memory. Gemini's memory features are different. This guide is ChatGPT-specific.
GPT-5.5 hallucinates less, but only when its memory of you is accurate. Five minutes today, fewer confidently wrong answers next month.
Forward this to anyone who's been using ChatGPT for more than six months.
Hit reply with the weirdest thing you found in your memory. We'll feature the best ones (anonymized) in a future issue.