Deleted a ChatGPT Memory and It Still Shows Up: Why
You opened the settings, found the entry, deleted it, and the next conversation brought it up anyway. This is one of the more unsettling things memory does, because deletion is supposed to be the one operation with no ambiguity in it, and here it appears to have been ignored.
In almost every case it was not ignored. The entry really is gone, and the fact arrived from somewhere else. There are five other places it can live, and the reason this is confusing is that all five produce an identical symptom: the assistant knows something you thought you had removed. Working out which one you are looking at takes a couple of minutes and determines whether there is anything to fix at all.
The Six Places a Fact Can Come From
Only one of them is the saved memory list you just edited.
The saved memory store. Discrete facts extracted from your conversations and kept in a list you can open and read. This is the one that deletion acts on, and when you delete from it the entry is actually gone.
Custom instructions. A block of text you wrote once, supplied to conversations unconditionally. It is not memory, it is not in the memory list, and clearing memory does not touch it. People forget what is in theirs, because the whole point is that you set it and stop thinking about it. If a fact about you keeps reappearing with suspicious consistency, read your custom instructions before anything else — consistency is the signature, because instructions are supplied every time while stored facts are retrieved only when judged relevant. The full map of the settings involved is in ChatGPT Memory Settings: Complete Control.
Project or workspace instructions. The same mechanism scoped to a container. A fact set at that level will follow you through every conversation inside it and be invisible everywhere else, which produces the maddening version of this problem: deleted, apparently still there, but only sometimes.
The current conversation's own transcript. Covered below, because it is the most frequent cause by a distance.
History referencing, where a product offers it and the account has it switched on. A mechanism distinct from the saved list, with its own control, and clearing the list does not clear what it can draw from. Whether your account has such a thing, and what it is called this quarter, is a question for your own settings screen rather than for any article.
Nowhere — the model inferred it. The biggest false positive of the lot, and the subject of the next section.
Inference Looks Exactly Like Recall
This is the one that sends people back to the settings screen three times.
If you deleted the stored fact that you write software, and the next conversation addresses you as a developer, nothing failed. You asked about a stack trace. The model is not recalling anything; it is reading the message in front of it and responding to what is obviously there. The same goes for your register, your vocabulary, the kind of problem you bring, and the language you write in — all of it is legible in the current message and none of it needs a store.
The test is to remove the cue. Ask the question in a way that gives nothing away, in a conversation containing nothing else, and see whether the fact still appears. If it does not, there was never anything stored, and you can stop looking. The method generalises, and the trap generalises with it — a correct answer is not by itself evidence of memory, which is the argument of How to Test What Your AI Assistant Actually Remembers.
Worth being honest about how much this accounts for. A good share of reports that deletion does not work are this, and the way to tell is not to reason about it but to run the clean test once.
The Conversation You Are Testing In Still Has It On Screen
The second most common cause, and the most quickly fixed.
If you deleted the entry and then went back to the same conversation to check, the check is meaningless. Everything already in that thread is still part of what gets sent on every turn — the fact is sitting there in message nine, in your own words, being supplied to the model regardless of what the memory store now contains. Deleting a stored entry does not and cannot edit the transcript of a conversation that is already open.
Test in a brand-new chat. Not a new branch of the old one, not a regenerated reply: a new conversation with nothing in it. If the fact is absent there, the deletion worked and the old thread is simply still carrying its own history.
The same confusion runs in the other direction, which is worth knowing because it is a separate trap. Deleting a conversation from your history does not delete anything that was extracted from it into the store, because the two are independent. A fact lifted out of a chat in March survives the deletion of that chat in October, and people who clear their history expecting a clean slate are frequently surprised. The reverse holds too: clearing memory leaves every conversation in your history exactly where it was. Both of those are set out in How to Make ChatGPT Forget Something.
The Entry You Could Not Find
Sometimes the deletion did not happen, because the thing you were looking for was not where you looked.
Stored entries are written in the extractor's wording rather than yours. A conversation in which you mentioned moving to a new city can become an entry phrased around relocation, or around a place name, or around a job — and if you scanned the list for the phrase you used, you missed it. Lists also get long, and scrolling a long list looking for a fact you half-remember is an unreliable operation.
Read the whole list rather than searching it, at least once. It is a tedious few minutes and it is the only way to be sure. It also tends to be informative in its own right: most people find at least one entry that is wrong, stale, or was never meant to be durable in the first place, which is a related problem covered in When ChatGPT Remembers the Wrong Information.
The other thing to look for while you are there is duplication. Facts get stored more than once in slightly different wordings, and deleting the one you found leaves its twin in place. If something survives a deletion you are confident about, a duplicate is the likeliest mundane explanation.
Corrections Do Not Replace
A particular case of the above, and the one with the worst consequences.
Telling the assistant that something has changed does not reliably overwrite the old entry. The correction can be stored as a new fact sitting alongside the original, which leaves two contradictory statements in the store and no mechanism deciding between them. Which one gets applied in a given conversation is then effectively a coin flip, and the one you will notice is the wrong one, applied confidently, in a context where it matters.
This is why corrections should be made as deletions. Open the list, delete the old entry, then state the new fact. Two operations rather than one, in that order. Stating the new fact first and trusting it to displace the old is the move that produces the coin flip.
Deleting Everything
If you want a clean slate rather than a surgical removal, the path is different and so are the trade-offs.
Clearing the whole store is a single operation and it is not reversible — there is no undo and no export of what was in there, so if any of it was useful, read the list first and copy out what you want to keep before you clear. People who clear in frustration routinely wish they had done that.
It is also incomplete on its own. Clearing the store does not touch custom instructions, project instructions, your conversation history, or anything the product derives from history under a separate control. A genuine reset is several operations in several places, and the inventory of them is in Make ChatGPT Forget Everything About You.
And it does not prevent recurrence. An empty store starts refilling from your next conversation, by the same extraction that filled it the first time. If the problem is that it keeps storing things you did not want stored, the control you want is the one that governs whether it stores at all, not the one that empties it.
Confirming a Deletion Actually Took
Thirty seconds, and worth doing whenever it matters.
Delete the entry. Reload the settings screen and confirm the list no longer contains it — reloading rather than trusting the view you are already looking at, because an interface can show you a local state that never reached the server. Then open a brand-new conversation and ask a question whose natural answer would use the fact, phrased so that nothing in the question itself gives it away.
If the fact does not appear, you are done. If it does appear, you now have a specific question rather than a vague one: work down the list at the top of this post. Is it in your custom instructions? Are you inside a project with its own instructions? Could it have been inferred from what you just wrote? Is there a duplicate entry you missed? One of those is almost always it.
If you are signed into more than one account, add that to the list. It sounds unlikely and it happens often enough to be worth ruling out early, because every one of these symptoms is also what a second account looks like.
The Order to Check In
Condensed, because the diagnosis is the whole job and it has a fastest path.
- Reload the settings screen and confirm the entry is actually gone from the list. If it is still listed, nothing else in this post applies and the deletion simply did not go through.
- Test in a brand-new conversation, never in the thread where you first noticed the problem. This single step resolves the largest share of cases on its own.
- Phrase the test question so it gives nothing away. If the fact still appears when the question contains no cue, it is coming from somewhere; if it only appears when you have effectively told it, that was inference.
- Read your custom instructions. Unconditional supply, invisible in the memory list, and the most common hiding place for a fact that reappears with perfect consistency.
- Check whether you are inside a project or workspace with instructions of its own. A fact that appears in some conversations and not others points here before anywhere else.
- Read the whole stored list rather than searching it, looking for a duplicate or a differently-worded twin of the entry you removed.
- Confirm which account you are signed into, which costs five seconds and explains an embarrassing number of these.
Working in that order matters, because the first three steps are free and account for most outcomes, while the later ones involve reading things you would rather not read.
What Deletion Cannot Reach
Two limits, neither of which is a fault.
Deletion acts on what a product stores about you going forward. It is not a claim about every copy of anything, and treating it as one leads to unpleasant surprises; if what you need is for something never to have been sent, the control for that is the one that governs what you put in a chat in the first place, not the one that empties a list afterwards. Conversations held with memory switched off, or in a chat type that does not write to the store, are the mechanism designed for that, and they are a decision taken before rather than after.
The second limit is the ordinary one. Nothing you delete affects a conversation that is currently open, because its text is already in the thread. Starting fresh is not merely a better test; it is the only way to get a session that does not contain the thing.
Where MindLock Sits
A deletion you cannot verify is a deletion you do not really have, which is most of what this post is about.
MindLock's position is that a memory store should be one you can open and read in full. Conversations you choose to save are distilled into memory documents — a profile memory plus topic memories — which are plain text. You can read every line, edit a line that has gone stale, and delete anything you no longer want, with no extraction logic between you and the entries and no question about whether the deletion took. On the free tier the store lives in your own browser rather than on a server, so the answer to where the data is does not require anyone's assurance. The documents are described in Memory Documents.
The honest boundary is that MindLock is not a control panel for ChatGPT. It cannot delete anything from your saved memory store, because no third-party software can reach it — that deletion remains yours to make on your own settings screen. What it offers is the other half: a record of your own that does not store things you did not choose, and whose contents are visible rather than inferred. Saving is manual and pasting is manual, for the same reason; the automated part is the distilling, searching and condensing in between. What runs where, and what the paid tier adds, is on the pricing page.