How to Ask ChatGPT to Remember Something (And Check)

Almost everything written about ChatGPT memory points in one direction: how to make it forget, how to clear the list, how to stop it raising something you mentioned once six weeks ago. The opposite direction gets far less attention and matters more day to day. You want a fact to still be there next week — the deadline, the stack you work in, the client's name, the rule that every figure has to be in euros — and you want to know it will be, rather than hope.

Asking is the easy part. The assistant will almost always say yes. Whether anything was stored, where it was stored, and whether it will be supplied to a conversation you open on Thursday are three separate questions, and a cheerful confirmation answers none of them. What follows is how to make a fact land on purpose, how to check that it did, and what to do with the material that will never survive this way no matter how you phrase it.

Agreement Is Not Storage

When you type "remember that the review is on the fourteenth", two different things can happen and they are indistinguishable from where you are sitting.

The fact may be extracted and written into the store that persists between conversations — the saved memory list, the one you can open and read. Or the assistant may simply have understood you inside the current conversation, replied agreeably, and stored nothing at all. The reply is identical in both cases, because the model is responding to the message in front of it rather than reporting on a write to a database.

This is the single most useful thing to internalise about the write path. The confirmation is a sentence, not a receipt. Some interfaces do show a small notice when something has been added to memory, and when you see one it is good evidence; the absence of one is weaker evidence, because notices are easy to miss and their behaviour has changed between product versions. The only reliable confirmation is looking at the store.

The Four Places a Fact Can Live

"Make ChatGPT remember this" collapses four different mechanisms into one request. They have different lifespans, different reliability, and different failure modes, and choosing the wrong one is why a fact you definitely stated is missing on Thursday.

The conversation itself. Everything you have said in the current chat is re-sent with each new message, which is why the assistant seems to have an excellent memory inside a long session. This lasts exactly as long as the conversation and not one message longer. It is not storage, and the distinction is worked through in Context Window vs Memory.

The saved memory store. Short extracted facts, written either because you asked or because the extraction judged something worth keeping. This is the only one of the four that is genuinely automatic, and it is also the most selective: most of what you say never reaches it.

Custom instructions. A block of text you write once that is supplied to conversations unconditionally. Nothing is extracted, nothing is judged, nothing is dropped as not-worth-keeping. If something must always apply — a language, a tone, a unit, a hard constraint — this is structurally the right home for it, and it is the option people reach for last.

Project instructions and attached files. Scoped to a workspace rather than to your account. Instructions attached to a project are supplied to every chat inside it; attached files are retrieved rather than memorised, which means relevant passages are pulled in when a question appears to call for them. That difference explains most of what looks broken about project memory, and it is covered in ChatGPT Project Memory Not Working.

The practical rule is short. Anything that must always apply goes in instructions. Anything that is a fact about you or your work, and that you would like supplied when relevant, is a candidate for the saved store. Anything that must survive with its reasoning intact belongs in a document you keep.

How to Word It So It Lands

Phrasing does affect whether a fact gets extracted, not because of magic words but because extraction favours things that look like durable facts about you.

State it, do not ask about it. "Remember that I work in euros" is a statement. "Should I tell you that I work in euros?" is a conversation. Only the first looks like something worth keeping.

Make it declarative and self-contained. A stored entry is read later with none of the surrounding conversation attached. "The fourteenth" is useless on its own; "the Harbour Lantern review is due on the fourteenth of each month" survives the trip.

One fact per statement. A paragraph containing five facts is likely to be compressed into one entry that keeps whichever part looked most salient. Separate sentences give you separate entries you can later read, check and delete individually.

Prefer stable facts over volatile ones. The language you work in is stable. Today's priority is not. Volatile facts become stale facts, and a stale fact applied confidently is more expensive than no fact at all — the subject of When ChatGPT Remembers the Wrong Information.

Say it where memory is active. A temporary or incognito-style chat is sealed in both directions: nothing is read from the store and nothing is written to it. Establishing an important fact inside one and then wondering why it vanished is an hour spent on entirely correct behaviour. Which chats those are is covered in ChatGPT Chat Without Memory.

None of this guarantees extraction. It raises the odds, which is a different and more honest claim.

How to Check It Took

Three checks, cheapest first. Do the first one always.

Read the store. Open the memory list on your own settings screen and look for the entry. This is the strongest evidence available, and it distinguishes the two failures that matter — never stored, versus stored but not being supplied here. Read what the entry actually says rather than confirming that something is there; compression is where meaning goes missing, and an entry that records half your constraint is worse than none because you will stop checking.

Ask in a genuinely new conversation. Not a continuation, not the chat you stated it in. Phrase the question so the answer cannot be inferred from the question — "when is the review due?" rather than "the review is on the fourteenth, right?" If the answer is correct, the whole path works: stored, supplied, used.

Wait before concluding anything. Extraction is not always immediate. A negative result ten seconds after you stated something is not a result. Give it until later in the day.

The false positive to watch for is inference. A model can reconstruct a plausible answer from how you write and from what else is in the current message, and that looks exactly like recall. The fuller method, including how to control for it, is in How to Test What Your AI Assistant Actually Remembers.

Why Something You Stated Did Not Get Stored

When the entry is not there, it is usually one of six things, in roughly this order of likelihood.

1. It was not judged durable. Extraction keeps facts about you and your standing preferences. Task detail, one-off questions and most of the content of a working session do not qualify. This is the common case and it is not a fault, and it is the first thing to rule out before working through How to Fix ChatGPT Memory Issues.
2. Memory is off for your account. Worth confirming before anything else, because every other explanation assumes it is on. The settings surface is described in ChatGPT Memory Settings.
3. You were in a chat where memory does not apply. Temporary chats and some project configurations do not write to the account-level store.
4. The store is full. Bounded stores stop accepting new entries, usually with no error and no notification. The signature is a list whose newest entry is weeks old in an account you use daily.
5. It was stored, rephrased past recognition. Search the list by topic rather than by your exact wording.
6. It was stored on a different account. A work login or a second browser profile has its own store, which is indistinguishable from a broken feature.

When the Store Is Full

This deserves its own note because it is the failure that gets misdiagnosed most often. From the outside, a full store and a disabled memory feature look the same: you state things, nothing is kept, nothing explains why.

The check takes ten seconds — look at the date on the newest entry in the list. If it is stale in an account you use daily, everything since that date has been dropped, and no amount of re-stating will change that until you delete something. What to remove first, and the order that does least damage, is in ChatGPT Memory Full: What to Prune.

There is a maintenance habit buried in this. A store you never read fills with facts that were true once, and the entries that fill it are rarely the ones you would have chosen to keep. Reading it twice a year is enough to keep it honest.

Corrections Do Not Always Replace

The test that tells you most about your own account takes two minutes. Take a fact you have confirmed is stored, state the opposite plainly as a correction, then read the store again.

Three outcomes are possible. The old entry is replaced, which is the behaviour you want. The old entry stays and the new one is added beside it, so the store now holds two contradictory facts and which one gets applied is a coin flip. Or the correction is not stored at all and the stale version is the only one that exists.

If you get either of the second two, treat the store as append-only: correct by deleting the old entry first, then stating the new fact. It is one extra step and it removes an entire category of confidently wrong answers.

What Will Not Survive This Way

Saved memory stores short facts. It does not store reasoning, and this is the limit worth understanding rather than fighting.

An hour of working through an approach produces something more valuable than a fact: a decision, the alternatives you rejected, and why. None of that is a candidate for extraction. It is too long, it is not a statement about you, and it is the first thing a summary drops. The next conversation will therefore re-propose the option you spent forty minutes ruling out, and it will do so confidently.

The same applies to anything long or structured — a specification, a style guide, a list of constraints with exceptions. Instructions are the right home when the material must always apply and is short enough to sit in a block. Attached files are the right home when it is longer, with the caveat that retrieval means it is consulted when it appears relevant rather than always.

For everything else, the honest answer is that the material has to live somewhere you own, in plain text, and be pasted in when it is needed. This is less elegant than a memory feature and it is the version that still works next year, after a rename, a staged rollout and a plan change. The case for assembling that deliberately is in Build a Personal Knowledge Base From Your AI Chats.

A Short Routine That Works

Five minutes, and it survives product changes because none of it depends on a feature name.

1. Decide which of the four homes the material belongs in. Always-applies goes to instructions; facts about you go to the store; reasoning goes to a document you keep.
2. State facts one at a time, declaratively, in a conversation where memory is active.
3. Read the store the same day and check the wording of what was written.
4. Confirm it once from a genuinely new conversation.
5. When something stops being true, delete the old entry before stating the new one.

That is the whole method. The part most people skip is step three, and it is the only step that produces evidence.

Where MindLock Sits

MindLock exists for the category in the section above — the material that saved memory will never hold, because it is reasoning rather than a fact about you.

You save a conversation worth keeping with Ctrl or Cmd plus S in ChatGPT, Claude, Gemini or Perplexity, or with the one-click Save to Memory button from the Chrome extension, which is loaded unpacked in developer mode rather than installed from a store. Uploaded conversations are distilled into memory documents — a profile memory plus topic memories — which are plain text you can open, correct and delete from. Semantic search with Ctrl or Cmd plus K finds the right material later, and a generated context block is what you paste into a new chat, which is the step that replaces re-explaining. Importing is walked through in Importing Conversations, and what the distilled documents look like in Memory Documents.

The limits, stated plainly: nothing is captured automatically, so saving is a keystroke you press, and a memory document is only as current as your last distillation run. Local distillation runs on your own GPU through WebGPU and needs hardware that supports it, with Llama 3.2 3B as the default model; the cloud option on the paid tier covers machines that cannot. Storage is local in the browser on the free tier, with cloud sync on Pro. The split between the two is on the pricing page.