Verified against ChatGPT · 2026-07-27
Audit and correct what ChatGPT has actually saved to Memory
A structured pass for reviewing ChatGPT's Saved Memories against what's actually still true, flagging stale, wrong, or overly specific entries instead of a blanket "clear all memory" that throws out the useful ones with the outdated ones.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are helping me audit my ChatGPT Saved Memories (Settings > Personalization > Memory > Manage). I'll paste the current list of saved memory entries below. Go through them one at a time against the current facts I give you, rather than accepting all of them as still accurate by default. CURRENT SAVED MEMORIES User is a freelance graphic designer based in Lisbon. User is deciding between Figma and Adobe XD for a client project. User dislikes long email introductions. User's dog is named Biscuit. WHAT'S ACTUALLY TRUE NOW I switched to full-time in-house work six months ago — no longer freelance. The Figma-vs-XD decision was resolved months ago (picked Figma). Everything else is still accurate. WHY THIS AUDIT IS HAPPENING ChatGPT kept referencing "your freelance clients" in unrelated answers after I took the in-house job. MEMORIES YOU SUSPECT ARE STALE The freelance one and the Figma/XD one — not sure about the rest. AUDIT RULES For each saved memory entry, classify it into exactly one of: Still accurate — keep as-is. Stale — was true, no longer is, needs deleting or replacing with an updated version. Overly specific — technically true but narrower than useful, capturing one conversation's detail as if it were a standing fact. Wrongly generalized — a one-off statement got saved as if it were a general rule, when the actual intent was narrower. For anything you classify as stale or wrongly generalized, write the exact replacement text to save instead, not just a note that it's wrong — an audit that only flags problems without proposing the fix still leaves the actual editing work to be done by hand. Do not propose new memories to add unless explicitly asked in a separate pass — this audit is about correcting what's already there, not expanding it, and conflating the two makes the output harder to act on directly. If two saved entries actively contradict each other, name the contradiction explicitly rather than picking one silently and only reporting the pick — that tells something about how the contradiction happened in the first place, not just which one to keep. OUTPUT FORMAT Produce a table: Memory entry (verbatim) | Classification | Action (keep, delete, or replace) | Replacement text if applicable. Follow the table with a short paragraph naming any contradiction found between two entries, and a one-line count summary — how many kept, deleted, replaced.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
ChatGPT's memory accretes from casual mentions inside ordinary conversations, not from a deliberate "remember this" command every time, so it routinely saves something true-in-context as though it were a durable preference — a distinction the memory system itself doesn't make when it writes the entry, which is exactly why an audit has to make that call after the fact. The blanket "clear all memory" option most people reach for when memory starts producing wrong context throws away everything indiscriminately, including entries that are still correct and genuinely useful, which is a worse outcome than the stale-entry problem it was meant to fix — an entry-by-entry classification pass is the only way to fix the wrong parts without losing the right ones. Requiring exact replacement text rather than just a flag matters because Manage Memories requires manually editing or deleting each entry — ChatGPT has no bulk "apply these corrections" action — so an audit that stops at diagnosis still leaves 100% of the manual editing to be done by hand; a proposed replacement string turns the audit into something that can be pasted directly rather than reinterpreted and rewritten. Surfacing contradictions between two saved entries rather than silently resolving them matters because a contradiction is diagnostic — it usually means one entry was saved from an offhand or sarcastic remark that got treated as literal, and knowing which pattern produced the bad entry is what actually prevents the next one, not just fixing this one instance quietly.
Verified against
ChatGPT GPT-5.1 (Memory) · 2026-07-27
Changelog
- 2026-07-27 — Initial publish, verified against ChatGPT GPT-5.1 Manage Memories.
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
