Troubleshooting playbook · 9 min read
"ChatGPT stopped remembering me" is one of the most common searches on the internet right now, and the top-ten Google results are all repeating three of the same tips. Below are the eight failure modes we actually see — in the order we check them when a Konshus member says memory just broke — plus the fix for each, when to escalate, and the one fix nobody wants to write about.
The short version
Walk down this list in order: (1) Memory setting, (2) Temporary Chat, (3) memory-cap full, (4) wrong account, (5) stale client cache, (6) retrieval didn't fire, (7) browser extension stripped the request, (8) account-level pause. The first six catch 90% of cases. The last two are where people give up and switch tools.
This is the most common one and the one everyone misses because they check it last. Open Settings → Personalization → Memoryand confirm both switches are on: "Reference saved memories" and "Reference chat history". They are two independent features, and either can flip off after a client update, a device change, or (rarely) a policy sync from another device on the same account. If either one is off, ChatGPT will still act like a normal chatbot — it just won't pull from anything long-lived.
Fix: turn both on. Then send a probe: "What do you remember about me?" If the reply lists things, memory is working; if it says "I don't have anything saved yet," you were probably starting fresh in a chat that never wrote memories in the first place — move to step 3.
The lightning-bolt icon next to the model picker toggles Temporary Chat, which is essentially an incognito mode: nothing is written to memory, nothing is added to history, nothing is used for training. It's easy to accidentally start one — the icon is tiny, and some deep-link entry points default to it. If your last three chats aren't showing up in the sidebar, this is almost certainly why.
Fix: exit the chat, start a new one from the sidebar, and confirm the lightning-bolt icon is not lit before you type. Also worth checking on mobile — the toggle state doesn't always sync between web and app.
OpenAI doesn't publish the cap, but heavy users hit it around 200–400 entries depending on entry length. Once full, ChatGPTsilently overwrites older memories with newer ones — no notification, no confirmation, no audit trail of what was dropped. If it feels like ChatGPT used to know something specific about you six months ago and doesn't now, this is the cause 80% of the time. See our ChatGPT memory limits field note for the numbers behind this.
Fix: open the memory panel and prune ruthlessly. Every entry you delete is one less slot fighting for attention. Also worth doing periodically even when things are working — the model retrieves better from a smaller, cleaner set.
Memory is per-account, not per-device. If you've ever signed into a work SSO tenant, a Team workspace, or a second personal account, each has its own memory store. Two of the last three "help, my memory disappeared" tickets we've seen were people who'd logged into a workspace last week and forgot they never logged back out.
Fix: go to Settings and check the email at the top. If it's not the account you thought, sign out fully, sign back in, and confirm before assuming data is lost.
The web client caches the last-known memory state client-side. After a policy change, an account switch, or a slow sync, the cached state can lag reality — you see "memory off" in the UI even though it's on server-side, or vice versa. This is rare but harmless once identified.
Fix: hard-refresh (Cmd/Ctrl-Shift-R) on web, or force-quit and reopen on mobile. If it still looks wrong after a reload, it's not a cache issue and you should move on.
Even when memory is on and populated, the model has to decidein each turn whether the current prompt is relevant enough to pull memories in. That's a probabilistic decision. A question like "give me some ideas" often skips retrieval entirely; a question like "given what you know about me, give me some ideas" almost never does. If memory feels intermittent — sometimes it knows you, sometimes it doesn't — this is why.
Fix: phrase prompts explicitly. "Based on what you remember about me…" or "Using my saved preferences…" nearly always forces retrieval. If you want deeper mechanics, see AI memory vs context window explained.
Aggressive ad-blockers, privacy extensions, and enterprise MDM profiles occasionally strip headers or block the memory-panel endpoint. If memory works in an incognito window (with extensions disabled) but not in your normal window, an extension is the cause.
Fix: allow-list chat.openai.com in the offending extension, or switch to a browser profile without the extension active. On managed devices, ask IT to whitelist the memory endpoint.
Rare but real: accounts under review (usually for content-policy reasons, but sometimes just for unusual usage patterns) can have memory writes paused silently. Reads still work, so existing memories show up, but nothing new is being saved. If steps 1–7 check out and new memories still aren't appearing after a week, this is the case.
Fix: contact OpenAI support with account email and a description of what's happening. If you're mid-flag, they'll tell you what to do; if not, they'll usually restore writes within a few days.
You just read eight fixes. Every one of them is patching a symptom of a system you don't control. OpenAI can — and does — change memory behavior in every model release. GPT-5 shipped with subtle retrieval differences from GPT-4o. The next model will ship with more. Whichever fix works for you today may need to be re-applied after the next update, and there is no notification when that happens.
The durable fix isn't a setting. It's owning the memory layer itself — keeping a copy of everything ChatGPT knows about you in a place you control, and injecting it back into every new chat, every new model, and every new provider you try. That's what Konshus is. When ChatGPT breaks, your memory doesn't break with it — you paste your Whisper into the chat and keep going. When you switch to Claude, the same context comes with you.
Konshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.
Meet KonshusKonshus is one way to solve this — a persistent memory vault and portable persona that follows you across ChatGPT, Claude, Gemini, and whatever ships next.
Meet Konshus