How to make ChatGPT remember across chats

Q: How do I view and delete what ChatGPT has saved in memory?

Open Settings → Personalization → Memory . The current system shows a running memory summary , updated automatically as you chat, with a "last updated" stamp at the top. Correct it by typing into the text box at the bottom, or highlight a passage to fix that specific line. Its three-dot menu holds Delete and turn off memory — OpenAI says this clears the memories shown on that page but does not delete your past chats, and that if you turn memory back on later "ChatGPT may create new memories from chats that remain in your chat history" (web now, rolling out on mobile). The legacy saved memories system instead shows a list of individual entries you delete one at a time or clear all at once. Ask in a normal chat too — "what do you remember about me?" — because OpenAI's guidance says the summary "will not include everything that ChatGPT remembers." And deleting one line isn't erasure: to fully remove something, OpenAI says you have to delete every source where it appears — past chats, archived chats, files, the summary — and disconnect any connected apps that hold it. (Memory FAQ, read 25 August 2026.)

Q: How do I word a "remember this" prompt so it actually sticks?

Lead with the word remember , put one fact in one sentence, and make the sentence stand on its own — "remember that Jordan Kim is Acme's CFO," not "remember what I said about him a minute ago." Softer framings like "don't forget this" or "keep this in mind" read as ordinary conversation and may pass without anything being stored. Then check rather than assume — OpenAI's documented way to find out whether something was kept is to ask in chat, so ask, and say it again more plainly if the answer comes back blank. Facts that will still be true next quarter hold best — your role, your clients, how you want answers formatted. A figure from Tuesday's call belongs somewhere you can reread it, not in a distillation.

One settings screen, one habit, and a way to check it worked. Here's the current setup as of August 2026 — and the one thing that stays out of reach even with everything turned on.

The short version: go to Settings → Personalization → Memory and turn on what that screen offers — Reference saved memories and Reference chat history on accounts that show two switches, or the single memory control that stands in for both on others. Then stop relying on ChatGPT noticing what matters — say "remember that…" for anything you'd be annoyed to re-explain. Then check what actually stuck, because what it keeps and what you assume it kept are rarely the same list. Full steps below, then the boundary none of this crosses.

Step 1 — turn on the memory controls

Continuity across chats comes from two separate mechanisms. Open Settings → Personalization → Memory — the path is the same on the web and in the iPhone app — and turn on what that screen offers you:

Reference saved memories. The discrete facts ChatGPT stores about you: your role, the clients you've mentioned, how you like answers formatted. On the legacy saved-memories system this switch is labeled Enable memory.
Reference chat history. From April 2025 OpenAI gave the model the ability to reference chat context beyond the saved-memories list — widening the pool a fresh chat can pull from to your earlier conversations as a whole, rather than only the facts it formally saved, so you stop re-explaining the engagement every Monday.

Two things to expect on that screen. First, you may not find two switches. OpenAI's Memory FAQ describes a single control — "you can choose to enable or disable memory at any time" — while its Projects article still names Reference saved memories and Reference chat history as separate personal settings that have to be on for project memory to work. Which you see depends on which memory system your account is currently using, and you can move between them yourself: Settings → Memory → Saved memories reverts to the legacy list, and Try improved memory inside it returns you to the current one. Switch on whatever your screen actually offers rather than hunting for a label from someone else's screenshot. Second, memory is per account — the work login on your laptop and the personal login on your phone are strangers to each other, so switch it on wherever you actually work. Confirming takes ten seconds; assuming has cost people a quarter's worth of context.

Step 2 — tell it what to remember, in words that stick

This is the step most people skip, and it's the one that does the work. Left to itself, ChatGPT stores only the details it expects to want later — a call made silently, on criteria you don't set and can't inspect. An explicit request is the reliable route in.

The phrasing that works is unglamorous. Lead with "remember," state the fact plainly, and keep it to something that will still be true next quarter:

"Remember that I run the Acme account and their steering committee meets on the second Tuesday of each month."
"Remember that when I ask for a client summary, I want three bullets and the open questions — no preamble."
"Remember that Jordan Kim is Acme's CFO and the pricing pushback came from him, not the CEO."
"Remember that I'm working across seven active engagements — always ask which client I mean if it's ambiguous."

Two habits make this hold. First, check that it landed instead of assuming. OpenAI's Memory FAQ documents no confirmation message when memory updates — what it tells you to do is the opposite: "If you want to know if ChatGPT has remembered something, just ask in chat." So ask, in that session or the next one. The memory summary also carries a "last updated" stamp at the top, which tells you when it last changed, though not what changed. Second, be selective on purpose. Standing facts survive well: your role, your clients, your formatting preferences, recurring dates. Volatile specifics don't — the exact number a client pushed back on, the wording of a commitment made three chats ago. Memory stores a distillation, not a transcript, and the transcript is where the specifics live.

Step 3 — verify what actually stuck

Turning memory on tells you nothing about what it captured. Check, in two places:

Ask it directly. Type "what do you remember about me?" into a normal chat. This is OpenAI's own documented method — "If you want to know if ChatGPT has remembered something, just ask in chat" — precisely because the settings page doesn't show everything.
Open the memory screen. Back at Settings → Personalization → Memory, the current system shows a running memory summary you correct by typing into it or highlighting a line; the legacy system shows a saved memories list of individual entries. Read it once properly — most people find something stale steering their answers.

Worth knowing before you trust either view. OpenAI's help center says the memory summary "will not include everything that ChatGPT remembers," and that memory is "a continually updated synthesis of context from your past chats, which may be broader than what can be shown as individual items in a summary." There's also a per-answer view — tapping the book icon under a response lists the sources that personalized it, including memories, past chats and files — but OpenAI notes those "may not show every factor or source that shaped a response." Inside a Project it's thinner still: OpenAI says project memory shows no list of memories at all. So you can see a lot of what ChatGPT knows about you. You can't open a complete inventory of it — and what the sources view shows is which chat, file or memory shaped an answer, not the client email or contract behind the fact itself, unless you put that document into ChatGPT yourself.

Step 4 — keep it accurate as the work moves

The first three steps get the mechanism running. What degrades it over months isn't a setting — it's that the facts you saved quietly stop being true. Three maintenance habits, in order of what they pay back:

Correct facts out loud when they change. The current system is built to handle the passage of time on its own — OpenAI's worked example is a memory revising itself from "you're going to Singapore in July" to "you went to Singapore in July 2026" once the trip ends. What it can't infer is the change it never saw: a cadence that shifted, a CFO who left, an engagement that closed. Say those the same way you said them the first time — "remember that Acme's steering committee now meets on the first Thursday; the second-Tuesday cadence is out of date." A stale memory left alone doesn't go quiet; it keeps steering answers with the same confidence as a current one.
Give each engagement its own Project — and choose its memory setting. OpenAI's Projects article (read 25 August 2026) says a project runs on either default memory or project-only memory, chosen when you create it and changeable afterwards under Project settings → Memory, with changes taking a few hours to apply. Project-only stops one client's chats shaping answers about another — previously saved memories aren't referenced inside it, and its chats can't be referenced from outside. Default keeps your standing preferences available in there instead. Decide that per engagement rather than accepting whichever one you were given.
Read the memory screen before it matters, not on a schedule. Skimming it the morning of a client call is worth more than a recurring reminder to tidy up. You're checking one thing: whether anything in there is about to put a wrong number or a closed engagement into an answer you're going to act on.

What your plan changes (and what it doesn't)

Memory isn't documented as a Plus-only feature: OpenAI's Memory FAQ describes one set of controls at Settings → Personalization → Memory and doesn't gate them by plan. What OpenAI has tied to plan is rollout and headroom. Announcing the current memory system on 4 June 2026, it said the update was "available to Plus and Pro users in the US today, and will roll out to additional countries and Free and Go users over the coming weeks," and that the efficiency work behind it also let OpenAI "increase memory capacity for Plus and Pro users." Other plan limits sit beside memory without being part of it: files per project run 5 on Free, 25 on Go and Plus, and 40 on Pro — a Projects number, not a memory one.

The two systems also differ in what you can inspect and how much upkeep they ask of you. The legacy saved-memories list is a set of discrete entries you delete individually or clear all at once; OpenAI's stated reason for moving on from it is that it "often became stale and relied on users to manually manage updates," with entries that could contradict each other. The current system is described instead as a continually updated synthesis of context from your past chats, curated automatically, and the Memory FAQ states no user-facing entry limit for it. If you're still on the legacy list and managing it by hand, here's what to do when ChatGPT memory is full.

If you've worked through all four steps and it's still losing the thread, the diagnostic order is in why ChatGPT forgets previous conversations. If the opposite is happening — old projects surfacing in new questions — the controls are in how to stop ChatGPT bringing up old conversations.

What Steps 1 to 4 still can't reach

Assume you've now done all of it properly. The controls on, one account, an explicit "remember that…" behind every standing fact, a Project per engagement, a memory screen you keep current. Open a new chat and try these four:

"Pull up what Acme's CFO actually wrote about the pricing model." That was an email. It was never in a chat.
"What changed between the redline they sent Tuesday and the version we signed?" Two attachments, three replies deep in a thread you never pasted anywhere.
"Remind me what I decided walking back from the site visit." That's a voice memo sitting on your phone.
"Where did that come from?" The sources view will name the chat, file or memory it drew on — but behind a fact you only ever typed into a chat, there's no original message to open, just its own distillation of a conversation you had with it.

Every step in this guide improves how well ChatGPT remembers ChatGPT. That's the honest limit, and nothing on the Personalization screen moves it. You can hand over your inbox or drive account-wide through its connected apps, but that's a full-account handover rather than the four threads that actually matter — and it still won't point at the specific message an answer came from.

For client-facing work, most of what's worth remembering was never typed into a chat window in the first place.

An assistant you can send the rest to

Equerry is a personal assistant that takes the four asks above. Nothing reaches it unless you send it, and what you send is held together rather than in whichever app it arrived in. Taking each ask in turn:

The email you never pasted in. Put your own Equerry address on the client thread — CC it going forward, or forward what's already there. The thread then keeps arriving as it develops, so you're not deciding each morning which messages are worth feeding it, and nothing else in your inbox travels with them.
The attachment three replies deep. Send the PDF or the screenshot through the share sheet and it's kept with the emails about the same client, rather than filed somewhere you'd have to remember separately.
The decision you made walking to the car. Dictate it as a voice memo or type it as a note. Either one lands with the rest of that engagement.
"Where did that come from?" Answers arrive with a source chip. Tap it and the exact email, note, or file opens — you're reading the original, not weighing up whether to trust a summary of it.
A list, not a summary. Everything Equerry has remembered is set out in one place with the source beside each item, and any of it comes out on a tap.
And ahead of the meeting. Ask what you're walking into and the answer is built from everything that has landed since you last looked — including the reply that arrived while you were in the previous meeting.

None of this replaces ChatGPT for what it's genuinely good at — drafting, thinking out loud, working a problem. Turn its memory on, follow the four steps above, and it'll stop opening every chat asking who your client is. Just don't ask it for the parts of your work that never passed through it. For those, see how Equerry handles the working context your day actually runs on, or what ChatGPT does and doesn't keep.

Checked on 25 August 2026 against three OpenAI sources: the Memory FAQ and Projects in ChatGPT help-center articles, and OpenAI's 4 June 2026 post Dreaming: Better memory for a more helpful ChatGPT. Quoted wording is OpenAI's own. Nothing here is drawn from a source outside those three; where OpenAI's pages are silent — on defaults, or on any entry limit for the current memory system — this guide says so rather than filling the gap. Control labels, plan limits and rollout states change often, so treat the labels as a guide and check your own Settings screen.

Frequently asked questions

Is ChatGPT memory available for free users or only Plus?

OpenAI's Memory FAQ doesn't describe memory as a paid-only feature — it documents one set of controls at Settings → Personalization → Memory and doesn't gate them by plan (read 25 August 2026). What OpenAI has tied to plan is rollout and headroom: its 4 June 2026 announcement of the current memory system said it was "available to Plus and Pro users in the US today, and will roll out to additional countries and Free and Go users over the coming weeks," and that the same efficiency work let OpenAI "increase memory capacity for Plus and Pro users." Other limits sit next to memory rather than inside it: files per project run 5 on Free, 25 on Go and Plus, and 40 on Pro — a Projects number, not a memory one. OpenAI moves these lines often, so treat your own Settings screen as the authority for your account rather than someone else's screenshot.

How do I view and delete what ChatGPT has saved in memory?

Open Settings → Personalization → Memory. The current system shows a running memory summary, updated automatically as you chat, with a "last updated" stamp at the top. Correct it by typing into the text box at the bottom, or highlight a passage to fix that specific line. Its three-dot menu holds Delete and turn off memory — OpenAI says this clears the memories shown on that page but does not delete your past chats, and that if you turn memory back on later "ChatGPT may create new memories from chats that remain in your chat history" (web now, rolling out on mobile). The legacy saved memories system instead shows a list of individual entries you delete one at a time or clear all at once. Ask in a normal chat too — "what do you remember about me?" — because OpenAI's guidance says the summary "will not include everything that ChatGPT remembers." And deleting one line isn't erasure: to fully remove something, OpenAI says you have to delete every source where it appears — past chats, archived chats, files, the summary — and disconnect any connected apps that hold it. (Memory FAQ, read 25 August 2026.)

How do I word a "remember this" prompt so it actually sticks?

Lead with the word remember, put one fact in one sentence, and make the sentence stand on its own — "remember that Jordan Kim is Acme's CFO," not "remember what I said about him a minute ago." Softer framings like "don't forget this" or "keep this in mind" read as ordinary conversation and may pass without anything being stored. Then check rather than assume — OpenAI's documented way to find out whether something was kept is to ask in chat, so ask, and say it again more plainly if the answer comes back blank. Facts that will still be true next quarter hold best — your role, your clients, how you want answers formatted. A figure from Tuesday's call belongs somewhere you can reread it, not in a distillation.

What is the difference between ChatGPT Memory and Custom Instructions?

Custom instructions are the standing brief you write yourself — who you are and how you want answers — and they apply to new chats until you change them. Memory is what ChatGPT accumulates: facts it picks up as you work, plus anything you explicitly told it to remember. Rule of thumb: custom instructions govern how it answers, memory governs what it knows about you. They sit on the same Personalization screen and complement each other. One caveat on reach: OpenAI says memory draws on "your chats, files, and connected apps" — so it isn't confined to the chat window, but it only sees what you uploaded or connected to ChatGPT yourself.

Last updated 2026-08-25.

All product names, logos, and brands are property of their respective owners. Equerry is not affiliated with, endorsed by, or sponsored by any other product mentioned; comparisons are provided for informational purposes.