Phase 1 · ChatGPT · Level 2 · Practitioner
Memory and custom instructions
By the end, you'll be able to…
- Turn Memory on or off, and view, edit and delete what ChatGPT has remembered about you
- Write custom instructions that shape every reply without repeating yourself
- Judge what's safe to let ChatGPT remember, and what should never go in
Why it matters
By default you re-explain your job, your preferences and your context in every new chat. Memory and custom instructions let you set that once and have it apply everywhere, a real time-saver. But anything ChatGPT stores about you is also something you should be able to see and control, so this lesson gives you both the shortcut and the off switch.
Two features that stop you repeating yourself
ChatGPT normally forgets everything the moment you start a new chat; each conversation is a blank slate. That's often what you want, but it also means telling it "I work in operations at a mid-size UK firm, keep replies brief and in UK English" over and over. Two features fix that in different ways.
Custom instructions are a short brief you write once that gets applied to every new chat: who you are, and how you'd like ChatGPT to respond. You control the words exactly.
Memory is more automatic: as you chat, ChatGPT builds up a picture of what's worth remembering (that you manage a team of six, that you prefer bullet points, that you're planning a product launch) and quietly draws on it later. You didn't write these; it inferred them. Recently memory was rebuilt to be time-aware: instead of just piling up a list of saved notes forever, it periodically re-synthesises what it knows about you, so newer facts carry more weight and stale ones fade rather than steering replies indefinitely. That's powerful and occasionally unnerving, which is why the controls matter.
Custom instructions: the brief you write once
Find these under your profile picture → Settings → Personalisation → Custom instructions (the exact path shifts, but it's in Settings). You'll see a couple of boxes: roughly, "What should ChatGPT know about you?" and "How should ChatGPT respond?" Fill them in plain language.
Keep it to durable, broadly-true facts and preferences, not one-off details, which belong in the individual chat. Good custom instructions are specific about tone, format and audience, because those apply to almost everything you ask.
About me: I'm an operations manager at Fernway, a mid-size UK company. I mostly write internal updates, emails to colleagues, and short summaries of documents. My audience is busy non-technical staff. How to respond: Use UK English and plain language. Default to short answers; lead with the key point, then detail only if useful. Prefer bullet points and headed sections over long paragraphs. When you're unsure or making an assumption, say so rather than guessing. Don't be effusive; skip the flattery.
Why this works: It front-loads the context you'd otherwise retype every time (role, audience, house style) so every reply lands closer to usable on the first try.
The effect is immediate: every new chat now starts already knowing your role, your readers and your style, so replies land closer to usable without you steering each time. You can override the brief in any single chat just by asking ("give me the long version this time"), so it's a default, not a straitjacket.
How to respond: When facts, figures or names matter, tell me how confident you are and what I should verify. If you're drawing on something that could be out of date, flag it. Never invent a source, statistic or quote; if you don't have it, say so.
Why this works: Telling it up front to show uncertainty and cite sources bakes your checking habits into every reply, so you're less likely to be handed a confident guess.
Memory: what it stores and how to control it
Memory lives under Settings → Personalisation → Memory. There are three things to understand: the switch, the summary, and the list.
The switch turns Memory on or off. With it on, ChatGPT saves things as you go and reuses them. With it off, it remembers nothing between chats. Every conversation is fresh. Both are legitimate; plenty of careful users leave it off precisely so nothing accumulates. Alongside the plain off switch you'll now usually find a "Delete and turn off memory" option, which clears everything it has stored and switches memory off in one step, the clean-slate button for when you want both at once.
The summary is newer and follows from the time-aware rebuild. Rather than only a raw list of notes, ChatGPT keeps a short synthesised summary of what it currently understands about you, and you can read and edit that summary directly: correct a fact it has got wrong, or trim a detail you'd rather it didn't lean on, and it works from your edited version.
The list, often labelled "Manage memories" or "Saved memories", is the more granular view people don't know exists. Open it and you see, in plain English, the specific things ChatGPT has chosen to remember: "Manages a team of six", "Prefers concise replies", "Working on a product launch in Q3". You can delete any single item with a bin icon, or clear them all. Between the summary and the list you've got your audit trail, and it's worth reading occasionally: you'll sometimes find it has remembered something you'd rather it hadn't, or an out-of-date fact ("planning to leave", "new to the team") that's now steering replies oddly.
Two more controls worth knowing. ChatGPT usually tells you when it saves a memory: a small "Memory updated" note appears; click it and you can undo that specific save on the spot. And there's a temporary chat mode (often a toggle near the model name) for a conversation that isn't saved to history and doesn't read or write memory at all, the right choice for a one-off you don't want influencing anything later.
What have you saved in memory about me? List each item plainly, and tell me if any of it looks out of date or too specific to be worth keeping.
Why this works: ChatGPT can recite its own saved memories in plain English, giving you a quick readable check without digging through Settings, though the Settings list is still the place you delete them.
The privacy angle
Memory is convenient, and that's exactly why it needs a rule of thumb. Anything ChatGPT remembers persists across chats and shapes future answers, so treat the memory store like a note you're leaving on a shared desk: fine for "prefers bullet points", not fine for a client's name, a colleague's personal circumstances, or anything confidential. Memory doesn't make a personal account safe for work data; it arguably makes it less safe, because sensitive details you'd have wanted gone after one chat can quietly stick around. Under UK GDPR, personal data about other people especially shouldn't be sitting in a personal AI account's memory. Keep the sensitive stuff out in the first place, review the memory list now and then, and use temporary chats for anything you don't want retained.
Try it now
Common mistakes
- Cramming one-off details into custom instructions. They apply to everything. "I'm writing a proposal for Acme this week" will still be steering replies next month. Keep instructions durable; put the specifics in the chat.
- Forgetting the memory list exists. Memory saves things silently, so people never look, and never notice an out-of-date or oversharing item nudging their answers. Read the list occasionally.
- Letting sensitive details get remembered. A name or account number mentioned in passing can be saved and reused. Use temporary chats for sensitive one-offs, and keep confidential data out of a personal account entirely.
- Assuming a remembered fact is a correct one (over-trust). ChatGPT can remember something it misunderstood, then apply that wrong fact confidently across future chats, with the error now invisible because you never re-stated it. A memory is a note it took, not a verified truth; check anything load-bearing rather than assuming "it knows".
Keeping current
The exact menus, and how much Memory infers versus asks, change regularly. The time-aware rebuild described above landed in June 2026 (OpenAI has referred to the periodic re-synthesis as "Dreaming"), and the editable summary and "Delete and turn off memory" controls arrived with it, so their exact labels and placement may still shift. For current behaviour and controls, see OpenAI's Memory FAQ and the custom instructions guide. Accurate as of 14 July 2026.