Is ChatGPT Memory Down Right Now, or Just Your Account?
Memory stopped working an hour ago. You cannot tell whether half the internet is having the same morning or whether something about your account has quietly broken, and the answer changes what you should do next entirely. If it is an outage, every minute you spend deleting entries and flipping toggles is wasted and potentially destructive. If it is your account, waiting for a fix that is never coming is equally wasted.
There is no indicator that tells you which. The interface behaves the same way in both cases: things do not get saved, older facts do not get used, and nothing announces itself as broken. So you are left doing what everyone does, which is searching to see whether other people are complaining, and reading forum posts from a different month.
This post gives you a triage that separates the two in about five minutes, using only checks you can run yourself. It is deliberately evergreen: it does not tell you whether anything is down today, because by the time you read this that answer would be wrong. It tells you how to find out.
Why This Is Genuinely Hard to Tell
Worth naming up front, because it is not you being unobservant.
A platform-wide incident and an account-level problem produce the same symptoms because the failure happens on the same code path. A write that fails server-side during an incident and a write that never fires because your list is full both result in nothing new being saved, with no error either way. A retrieval that times out during a degradation and a retrieval that returns nothing because the store is empty both result in an assistant that does not seem to know you.
On top of that, memory operates without receipts. There is no log of writes, no timestamp on entries, no notification when something fails, and no way to replay what happened. You get a list you can read and delete from, and that is the entire diagnostic surface. Every cause collapses into the same observation.
So the triage below does not try to look inside. It works by elimination from the outside in: platform, then account, then session, then store.
Step 1: Check the Platform Before You Touch Anything
Sixty seconds, and it determines whether the rest of this is even worth doing.
The official status page. OpenAI publishes service status at status.openai.com. Read it first. It is the only source in this list with any authority, and a posted incident ends your investigation immediately.
Its limitations, which matter. Status pages lag. A partial degradation affecting one subsystem for a subset of accounts frequently never gets posted at all, and when it does it is posted after people noticed. A green status page is weak evidence that nothing is wrong; a red one is strong evidence that something is.
Third-party outage trackers and social reports. Useful as a second signal because they are driven by users rather than by an incident process, so they spike earlier. They are also noisy, and people report "ChatGPT is down" when they mean their tab froze. Look for a sharp rise in the last hour rather than a steady trickle, and look for reports describing memory specifically rather than the product in general.
Check: has anything changed on the status page, and is there a burst of reports in the last hour naming the same symptom you have?
If yes: stop. Change nothing. During an incident, every setting you toggle and every memory you delete is a change you will not be able to distinguish from the recovery afterwards, and deletions do not come back. Note what you saw, close the tab, and re-test in a few hours.
If no: continue, but hold the possibility loosely. An unposted partial degradation is common enough that you should not treat a green page as proof.
Step 2: Rule Out the Session and the Client
Cheap, provable, and it eliminates a surprising share of reports before you get anywhere near your account.
Run these in order and stop as soon as one changes the picture:
- Hard reload. Not a normal refresh.
- A different browser, signed into the same account.
- A different device on a different network — a phone on mobile data is the ideal control, because it shares nothing with your desktop but the account.
- Disable any extension that modifies the ChatGPT interface, then reload.
Check: does the behaviour follow you across all four?
If it does not, the platform is fine and your account is fine. You had a client, cache or extension problem, and you have already fixed it.
If it does, you have established something genuinely useful: whatever is wrong is attached to the account, not to the machine. That narrows the remaining causes considerably and rules out an entire category of advice you will otherwise waste time on.
Step 3: Confirm You Are Testing Properly
More "outages" die at this step than at any other, and all of them are avoidable.
Test in a genuinely new chat. An open conversation still contains everything that was in it before anything changed, so it will keep behaving the old way regardless of what is or is not working. Judging memory from a chat you already had open proves nothing.
Check that the chat is not temporary. A temporary conversation contributes nothing to memory and draws on nothing, by design. It looks like an ordinary chat.
Check the account you are signed into. Memory is account-scoped. A second identity, a work login, a different browser profile — each has its own store, and an empty one is indistinguishable from a broken one. Read the email address in the account menu rather than assuming.
Check that memory is switched on at all. Settings do occasionally change under you after an update, and the controls have moved between product versions. Read your own screen rather than matching it against a label from an article.
Check: a brand-new, non-temporary chat, in the right account, with memory on. Does it still fail?
If any of those four was wrong, you never had a memory problem. Fix it and re-test once.
Step 4: Separate Which Memory Failed
"Memory is broken" covers several different failures with different causes, and mixing them up sends you down the wrong path.
It forgot something from earlier in this same conversation. That is the context window, not the memory feature. Long conversations run out of room and the earliest exchanges stop being visible. No outage is involved and no setting will change it. This one is covered properly in ChatGPT Context Loss Mid-Conversation.
It is not saving new facts. The write path. Candidates: memory off, temporary chat, list full, a server-side failure, or the system simply not judging the fact worth saving.
It is not using facts that are visibly in the list. The read path. If you can see the entry on the settings screen and the assistant behaves as though it does not exist, storage is fine and retrieval is not. This pattern is a meaningful outage signal, because it is one of the few that is hard to explain by user error.
It is using facts that are wrong or stale. Not an outage at all. Old entries sit alongside new ones and nothing expires.
Check: which of the four is yours? Write it down in one sentence before continuing. The full sorting exercise, with fixes per category, is in ChatGPT Memory Not Working? Complete Fix Guide.
Step 5: The Boring Account-Level Cause
If you have reached this step, the single most likely explanation is capacity, and it is not an outage.
The saved memory list is bounded. When it fills, new entries stop landing, silently, with no error and no notification. Everything already stored keeps working, which is exactly why this presents as "memory partly broke on a specific day" — the day the list filled up.
Check: open the list, scroll to the end, and look at the newest entry. If it is weeks old in an account you use daily, and the list is visibly long, you have your answer and it has nothing to do with anyone's servers.
Fix: prune, deliberately, oldest and most finished first. Deletion has no undo, so copy the list into a text file before you start. Which entries to remove and which to protect is worked through in ChatGPT Memory Is Full: What to Prune and What to Keep.
Step 6: What an Actual Outage Looks Like
For completeness, the pattern that genuinely indicates a platform problem rather than an account one:
- It started abruptly, at a time you can name, with no change on your side.
- It reproduces across devices, browsers and networks.
- Entries are visible in the list but are not being used — the read path, not the write path.
- Other parts of the product are also slow, erroring, or dropping requests.
- Reports naming the same symptom spiked within the same hour.
Three or more of those together is a reasonable call. One on its own is not.
What to do about it: nothing, and that is the hard part. There is no user-side fix for a server-side failure. The correct response is to change nothing, avoid deleting anything, and re-test in a few hours in a new chat. Anything you "fix" during an incident becomes a confound afterwards. Specific error strings you may see during one, and what each actually indicates, are catalogued in ChatGPT Memory Error Messages Explained.
The Triage, Condensed
- Status page and outage reports. Posted incident, or a spike naming your symptom? Stop and wait.
- Reload, second browser, second device, extensions off. Does it follow you?
- New non-temporary chat, right account, memory on. Still failing?
- Name which memory failed — context window, write path, read path, or stale content.
- Check list length and the date of the newest entry. Long and stale means full.
- Score it against the outage pattern. Three or more signals, wait it out. Fewer, it is your account.
Every step is free and reversible, and none of them is "clear all memories". That step destroys the evidence you would need to diagnose anything and is almost never the answer. The wider set of account-level problems and their fixes is in How to Fix ChatGPT Memory Problems.
Why You Keep Having to Do This
The deeper issue is not any particular incident. It is that your working context lives in a store you cannot inspect, cannot version, cannot export and cannot repair. When it misbehaves, your only available move is to guess between causes you have no instrument to distinguish, using a page maintained by the same party whose service is in question.
That is a reasonable design for a convenience feature. It is a poor foundation for anything you actually depend on. A morning spent wondering whether memory is down is a morning where the answer was unavailable to you by construction, and no amount of skill at triage changes that.
Context That Does Not Depend on Anyone's Uptime
The structural alternative is to keep the context you rely on in documents you hold, and supply them deliberately.
That is what MindLock does. You save the conversations that mattered with Ctrl or Cmd plus S, or with the one-click button from the sideloaded Chrome extension, upload them, and distillation produces memory documents you can open and read in full. Starting a new chat, you generate a context block from the relevant memories and paste it as the first message — in ChatGPT, or in whichever assistant is working today. On the free tier that processing happens on your own machine and the documents sit in local storage in your browser, which means an outage somewhere else does not reach them.
Two caveats stated plainly. Nothing is automatic: saving is a keystroke you press, and memory documents update when you re-run distillation rather than continuously. And local distillation needs a WebGPU-capable GPU; without one you would use the cloud option on the paid tier. How conversations get in is walked through in Importing Conversations, and the split between the tiers is on the pricing page.
The Short Version
You cannot tell an outage from an account problem by looking, because both take the same path to the same silence. So eliminate from the outside in.
Read the status page and look for a spike of reports naming your symptom. Reload, then try a second browser and a second device to rule out the client. Test in a genuinely new, non-temporary chat in the right account with memory on. Name which memory actually failed. Then check whether your list is simply full, which is the most common account-level cause and is not an outage at all.
If three or more outage signals line up, change nothing and come back in a few hours. If they do not, it is your account, and it is probably capacity. Either way, do not clear your memories to make the symptom go away — that is the one step you cannot take back, and it answers nothing.