Verified against ChatGPT · 2026-07-05
Decide what ChatGPT remembers about you instead of letting it drift
A periodic audit prompt that makes ChatGPT show every memory addition and deletion before saving it, so long-term memory stays accurate instead of quietly accumulating stale facts.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Show me every change to memory before you save any of it. Nothing gets written silently in this pass. WHAT HAS CHANGED SINCE MEMORY WAS LAST UPDATED Left my job at Acme Corp, now a freelance consultant working mostly with early-stage startups WHAT I NO LONGER WANT REMEMBERED, IF ANYTHING Forget that I was training for a marathon, I stopped in the spring Do the following, in order. 1. List every fact currently in memory about me that Left my job at Acme Corp, now a freelance consultant working mostly with early-stage startups makes outdated, and mark each one OUTDATED — REMOVE, with the reason. 2. List every new fact worth remembering long-term from Left my job at Acme Corp, now a freelance consultant working mostly with early-stage startups, marked NEW — ADD, phrased as a single durable fact about me, my work, or my preferences, not a summary of this conversation. 3. Remove anything named under things I no longer want remembered, and confirm explicitly that it has been forgotten, not just left unmentioned going forward. 4. Show the complete resulting list of what memory will contain after this update, and wait for me to confirm before saving anything. Do not add a fact about the specific topic of this conversation unless it is genuinely durable. A one-off question I asked today is not memory-worthy just because we discussed it.
Customize the highlighted detailsoptional — the prompt above already works
Why this works
ChatGPT writes to Memory both automatically, by inferring durable facts from what you say, and explicitly, when you say something like remember that, and both paths can be reviewed and edited under Settings > Personalization > Manage Memory, which is what makes an audit like this possible rather than theoretical. The propose-then-confirm structure matters because memory that updates silently in the background is memory you cannot debug when it starts giving stale-context answers weeks later, quietly assuming a job you left or a project you finished; forcing an explicit remove-and-add list before saving turns an invisible background process into something you actually approve line by line. Separating outdated removal from new addition into two distinct steps also prevents a common failure mode of unscoped memory edits, where a model asked to update memory adds the new fact but leaves the contradicted old one sitting alongside it, since nothing prompted it to check for the conflict. The closing constraint against saving one-off conversational details, rather than durable facts, is what keeps memory from slowly filling with the questions you asked instead of the things actually worth ChatGPT knowing about you going forward.
What you get back
OUTDATED — REMOVE: "Works at Acme Corp" — superseded by the new freelance-consultant status. NEW — ADD: "Freelance consultant working primarily with early-stage startups." Forgotten as requested: "Training for a marathon." Resulting memory list: [role: freelance consultant, focus: early-stage startup clients, ... ] — confirm before saving?
Verified against
ChatGPT GPT-5.1 (Memory) · 2026-07-05
Changelog
- 2026-07-05 — Initial publish, verified against ChatGPT GPT-5.1 Memory settings.
Building this for real?
This is a free starting point. If you'd rather have what Scult builds built and running for your business, that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
