ChatGPT greets you by name, knows your coding stack, and remembers your kid's birthday. It feels personal. Then you try a different assistant and walk in as a total stranger.
That is the catch nobody puts on the box. An AI that remembers you only remembers you inside its own walls, and the profile it built usually stays behind when you leave. This post breaks down what "remembers you" actually stores, why moving it is still mostly manual, what gets trapped when you stay, and one irony worth holding onto: the most portable option today is the one a competitor built to pull you away from your current assistant.
What "remembers you" actually stores (and what it doesn't)
It stores a short profile of facts and preferences about you, not a copy of every conversation. OpenAI's own documentation makes the distinction explicit: ChatGPT memory is meant for high-level preferences and details, and should not be relied on to store exact templates or large blocks of verbatim text. Think sticky note about who you are, not a transcript of what you said.
OpenAI splits this into two layers. "Saved memories" are details you explicitly ask it to remember, like your name, your role, or your dietary preferences. "Reference chat history" is information the model pulls from past chats to make future answers more relevant. Saved memories stick until you delete them. Details inferred from past chats can shift over time as the model updates what it finds useful.
Mechanically, the system injects that profile into the model's context every time you send a message. Each request assembles a context window: system instructions, your saved memory, a summary of recent conversations, and your current message. The model does not recall you the way a person does. It rereads a compact dossier about you on every turn, then writes a reply conditioned on it.
Other vendors do the same thing under different labels. Anthropic frames Claude's memory around your processes, client needs, project details, and priorities, leaning toward professional context. Google's Gemini learns from your past chats, gated behind a personal Google Account with activity history switched on. Different words, same move: a maintained profile, refreshed and reused.
Key reframe: "AI that remembers you" is a personal-context profile the model rereads each turn. It is not a recording of your conversations, and it is not the model permanently rewiring its weights to know you.
It is a profile, not a retrained brain
You did not teach it. You filled out a form it rereads. A common misconception is that the AI learns you by retraining itself on your data. It does not. The base model stays the same for everyone. Your personalization lives in a separate store that gets read at inference time and handed to the model as input. That is why you can open settings, see a list of remembered items, and delete one in seconds. You are editing a profile, not performing surgery on a neural network.
Good news for control: you can ask ChatGPT to forget a saved memory at any time, and it stops being used in future chats, though OpenAI notes it may keep a log of deleted saved memories for up to 30 days for safety and debugging. Claude shows a memory summary you can view and edit in one place, plus incognito chats that never write to memory. Gemini lets you manage what it knows by deleting the underlying activity.
Can you move AI memory between vendors?
Mostly not yet, and never in one click. Your AI memory lives in one vendor's silo, and true cross-vendor portability is still emerging rather than settled. No shared standard lets a profile follow you from one assistant to the next automatically. What exists today is a handful of early, mostly manual bridges, and they only carry a slice of what the model actually knows about you.
Here is what most switching guides won't tell you: the clearest exit anyone offers was built by a rival to poach you. Anthropic shipped an import and export feature for Claude memory, letting you bring memory details over from a different AI tool or export Claude memory for backup or migration. Anthropic's own help documentation flags it as experimental and still in active development. The export exists. It is a beta feature, not a portability standard.
The mechanics give it away. Anthropic's import flow has you run a prompt in your current assistant that asks it to list what it remembers about you, then paste that output into Claude. It is copy and paste, performed by you. The accumulated understanding is not packaged in a portable file you own. You rebuild it by hand, one paste at a time.
And the standard "export my data" button does not save you. When you run a vendor's normal data export, the conversation logs come out, but the memory profile, the model's distilled sense of who you are, does not reliably travel with them in a clean, reusable form. You get raw chats and still lose the personalization layer built on top. That gap, raw data without the derived profile, is exactly where lock-in hides.
Ask your assistant "What do you remember about me?" and keep that summary somewhere you control. Treat any vendor memory as borrowed, not owned. A profile you can re-paste is the only one you can actually take with you today.
What gets trapped: AI memory lock-in
Three things get trapped, and none of them are the conversations. What turns you back into a stranger at a new assistant is the derived context: your stated preferences, your project background, and the tone the assistant learned to use with you. Those are the inputs that take longest to rebuild from scratch, which is what makes leaving feel expensive.
- Preferences: how you like answers formatted, your reading level, the languages and frameworks you work in, the things you have told it never to suggest again.
- Project context: ongoing work, client constraints, recurring goals, and the background facts you would otherwise re-explain at the start of every session.
- Tone and relationship: the voice it settled into with you, the running shorthand, the assumptions about your expertise that make replies feel calibrated rather than generic.
The switching cost is the time and friction of re-teaching all of this to an assistant that starts blank. Industry commentators call it digital lock-in: when an assistant knows your communication style, history, and preferences, moving to a competitor means starting from zero. The friction is not a bug in any one product. It is the natural by-product of a memory layer that lives wherever you built it.
Notice what is not on the trapped list: the literal answers and exchanges. Those are usually exportable as conversation logs. The valuable, sticky part is the compressed profile sitting on top of them, and that piece has no clean exit. We dig into the ownership questions in who owns your AI memory, and into the mechanics of extraction in can you export your AI memory.
| Aspect | What it stores | Portability today |
|---|---|---|
| ChatGPT (OpenAI) | Saved memories plus referenced chat history; preferences and details, not verbatim transcripts | Manual: paste-out via a prompt; OpenAI does not currently document a memory export API |
| Claude (Anthropic) | Processes, project details, client needs, priorities; memory summary you can view and edit | Import and export feature, experimental and still in active development |
| Gemini (Google) | Learns from past chats; gated behind a personal account with activity history on | Manage by deleting source activity; no standardized cross-vendor export |
The pattern holds across all three: each vendor stores a profile, not your raw words, and each handles exit on its own terms. The most portable path today is the one a competitor built to win you over, which tells you exactly how much leverage portability gives the side that offers it.
The honest trade-off: intimacy versus dependence
The trade-off cuts both ways. Deeper memory makes the assistant more useful, and more useful makes it harder to leave. Those are not two facts in tension. They are one fact seen from two angles. Every increment of personalization that saves you re-explaining yourself is an increment of stored value you would forfeit by switching.
On the intimacy side, the gains are concrete. A model that knows your stack stops suggesting the wrong language. One that knows your audience stops pitching answers at the wrong level. OpenAI describes memory as making ChatGPT more helpful over time precisely because it carries forward what you have shared.
On the dependence side, the cost is structural. The profile sits on the vendor's servers, under the vendor's deletion policy and export options. You can usually view and prune it, which is real control. But you do not hold the canonical copy, and no neutral format makes it trivially yours. The memory feels like a relationship. It is actually a tenancy.
There is no clean winner here. The sober move is to go in with eyes open: enjoy the personalization, audit what is stored, and keep your own copy of the profile so dependence stays a choice rather than a trap. Memory off entirely is also a valid setting, and every major vendor offers it.
Where MemX fits
If the silo problem is what worries you, the fix is owning the source material rather than renting a profile built from it. That is the gap MemX targets. MemX is a consumer AI memory app, a personal second brain where you capture what matters and ask for it back later in plain English.
- Capture photos, PDFs, documents, voice notes, and WhatsApp messages into one place you control.
- Retrieve by asking a plain-English question; MemX answers from your own saved material and cites the source memory it used.
- Search inside images and scans, because MemX reads the text in them, so contents are findable, not just filenames.
- Set reminders by saying them naturally, and MemX writes them to your calendar, recurring ones included.
- Your data is encrypted, never used to train AI, and inaccessible to MemX staff. There is a free tier with no credit card, on Android, iOS via TestFlight, and WhatsApp.
That does not magically standardize memory across every chatbot. Nothing does yet. But it keeps the raw material, the notes, scans, and messages your context is built from, in a store you can query directly, instead of trapped as a profile inside one vendor's walls.
01Does AI store everything I tell it word for word?
No. Memory features store high-level preferences and details, not verbatim transcripts. OpenAI explicitly says ChatGPT memory should not be relied on to keep exact templates or large blocks of text. It keeps a compact profile and rereads it on each turn.
02Can I move my AI memory to a different assistant?
Partially, and mostly by hand. Anthropic offers an experimental import and export feature for Claude, and you can paste your remembered details across via a prompt. There is no universal standard yet, and OpenAI does not currently document a memory export API, so portability is emerging, not solved.
03Which AI assistants have memory?
ChatGPT, Claude, and Gemini all offer it per their own documentation. ChatGPT keeps saved memories and references past chats, Claude stores project and client context, and Gemini learns from past chats when activity history is on. Each maintains a profile, not a transcript.
04If I delete a chat, does the AI forget what it learned?
Not necessarily. With ChatGPT, saved memories from a chat can still be used after you delete the chat itself. To remove them, delete the specific memories in settings. Different vendors handle this differently, so check each one's controls.
05Can I turn AI memory off completely?
Yes. Every major assistant lets you disable it. ChatGPT lets you turn saved memory off in settings, Claude offers incognito chats that never write to memory, and Gemini's personalization depends on activity history being switched on, which you can leave off.
AI that remembers you is a profile the vendor maintains and rereads, not a recording and not a retrained brain. It is genuinely useful and genuinely sticky, and the part that traps you, your preferences, project context, and tone, is the part with the weakest exit. Portability is starting to appear, led by competitors who want your business. Until it matures, keep your own copy and treat any vendor memory as borrowed.
Written by Arpit Tripathi, who works on AI memory at MemX.
