Ask ChatGPT to help with your CV on Monday and it will cheerfully forget your name by Thursday. That was the reality for most of 2023. Memory fixed a lot of it, and the change is bigger than it sounds. It’s the difference between talking to a clever stranger and talking to a colleague who already knows which project you’re behind on.
Here’s how it actually works, where it earns its keep, and the parts that will quietly irritate you if you never open the settings.
Two systems, not one
Most confusion around ChatGPT memory comes from people treating it as a single feature. It isn’t. There are two mechanisms running in parallel, plus a third thing that behaves similarly but isn’t memory at all.
- Saved memories are short notes ChatGPT writes to itself about you: your job, your time zone, the fact that you’re learning Spanish. You can read every one of them, edit them, and delete them individually.
- Reference chat history lets the model search back through previous conversations for relevant context. It’s closer to retrieval than to recollection, and it’s the reason ChatGPT might suddenly remember a holiday you mentioned last spring in a completely unrelated thread.
- Custom instructions are rules you write yourself and which apply to every conversation. They aren’t memory, but they do the same job for stable preferences.
Saved memories get pulled in when they seem relevant. Custom instructions are always on. That difference matters more than it first appears, and it changes how you should phrase things.
How to see exactly what it knows
On the web, open Settings, then Personalization, then Memory. On mobile, tap your profile icon and follow the same path. You’ll get a toggle for reference chat history and a list of saved memories with timestamps next to each one.
If you’ve never looked at that list, do it now. It’s usually a mix of things you’d have chosen to tell it and a few you wouldn’t. Deleting an entry takes one click, and there’s a “delete all” option if you want a clean slate. If you’re new to the interface entirely, this practical starting guide for ChatGPT beginners walks through the settings screens in order.
You can also manage memory in plain language. Say “forget that I live in Berlin” and it should remove the entry, but check the list afterwards. Models occasionally acknowledge a deletion without performing it.
Reference chat history is the part that surprises people
Until 2025, memory was basically a list of notes. Reference chat history turned it into something closer to a search engine over your own conversations, and it rolled out gradually: Plus and Pro accounts first, free accounts later with tighter limits.
That’s why ChatGPT can now surface a detail you mentioned eight months ago and never saved. Useful when it recalls the exact wording of a brief you’ve since lost. Less useful when it brings up something personal in a thread you were planning to screenshot and share with your team.
Two practical consequences. First, it’s probabilistic, not a database lookup. It may miss context you know is there, or apply it at the wrong moment. Second, temporary chats bypass it entirely, in both directions. Nothing is read and nothing is written. When I’m drafting anything sensitive, that’s the mode I use.
Where memory genuinely earns its keep
The feature pays off when you stop repeating yourself. A few things it handles well:
- House style. Tell it once that you write in British English, dislike the word “leverage”, and keep paragraphs under four lines. It holds across sessions without being re-briefed.
- Long-running projects. Names, dates, budgets, the client who wants everything in the passive voice. “Project Kestrel launches 14 March, budget £40k, no mention of the agency to the client” saves ten minutes a week.
- Code conventions. Naming patterns, framework versions, the fact that you’re on Python 3.11 and not 3.9.
- Personal constraints. Dietary needs, mobility requirements, your working hours, your time zone for scheduling.
- Recurring admin. Your job title, your company size, the fact that you’re job hunting and don’t want it mentioned in the cover letters it drafts.
Stack that on top of automated workflows and it becomes more than convenience. It’s the difference between an assistant that knows the context and one that needs a briefing every time, which is exactly the gap discussed in this piece on what ChatGPT agents actually do and where they fall over.
Where it goes wrong
Three failure modes show up repeatedly.
Stale facts. Tell it you’re vegetarian for one dinner party and you’ll get vegetarian recipes for two years. It won’t notice you’ve changed your mind unless you say so explicitly.
Over-application. A memory about tone that suited a job application gets applied to a technical spec, and suddenly your API documentation reads like a cover letter.
Confident false memories. Say something hypothetical (“imagine I run a bakery”) and it may file that as fact. The list won’t always show it, but the behavior will.
The fix isn’t to switch memory off. It’s to audit the list occasionally and correct errors the moment you spot them: “That’s out of date. Update your memory to say I moved to Porto in February.”
How to write memories that survive contact with reality
Short, declarative, dated and single-topic. That’s the whole trick. “Remember: I’m based in Lisbon, WET, so schedule everything in that time zone” works. A rambling paragraph about your working patterns gets half-read and misapplied.
One memory per idea, so you can delete the stale ones without losing the rest. Add the date you set it, which makes spring-cleaning far easier. And for anything that will be true for years, use custom instructions instead of memory: instructions apply every single time, memories only surface when the model judges them relevant.
Privacy, training, and what leaves the room
Saved memories aren’t used to train OpenAI’s models, and the company states that plainly. Your general chat history setting still governs what happens to conversations themselves, and it’s worth checking separately if that matters to you. On Team, Enterprise and Edu plans, admins decide whether memory is available at all.
Worth knowing: memory is per account. It doesn’t follow you into a shared workspace, and it doesn’t cross into a different login. If you’re weighing tiers, the free plan now includes memory with tighter limits, while the paid tiers get the full version. There’s a clear breakdown of the differences in what you actually get on ChatGPT Free, and a blunter assessment of whether the top tier is worth it in what ChatGPT Pro buys you for $200 a month.
It isn’t unique to ChatGPT anymore
Gemini, Claude and Copilot all have some version of persistent context now, and they handle it differently enough that switching tools means rebuilding your profile from scratch. If you’re comparing, this year-long ChatGPT versus Gemini comparison covers how each one remembers and where the results diverge in daily use.
The five-minute monthly cleanup
Open the memory list once a month and read it properly. Delete anything tied to a project that’s finished, a city you’ve left, or a job you no longer have. Merge duplicates. Scan for anything you’d rather not see quoted back at you in a shared window.
Then ask it directly: “What do you remember about me?” The answer is a decent diagnostic. If it lists things you never told it, reference chat history has been working harder than you assumed. If it lists almost nothing, your saved memories are too vague to be retrieved. Either way, you’ll know within a minute whether the profile it’s built matches the person you actually are, which is the only test that really matters.