What ChatGPT, Claude, Gemini and DeepSeek Remember

All four major assistants now keep memory between conversations, and all four use it for the same job: building a working profile of you. ChatGPT synthesizes one in the background, Claude keeps editable entries you can read line by line, Gemini plugs into your Google account, and DeepSeek retains your history on servers in China. What none of them do is remember other people as people.

That distinction is easy to miss, because “memory” sounds like it covers both. It implies an assistant that recalls whatever came up. In practice it recalls whatever came up about you, because that is the only kind of fact that pays for its own storage on every future turn.

What each one actually holds

Assistant	What its memory holds	Organized around	Can you see and edit it?
ChatGPT	Your preferences, projects and constraints, synthesized in the background from all your chats	You	Yes, on a memory summary page
Claude	Your role, working style, technical preferences and project details, as individual entries	You, split per project	Yes, entry by entry, with pause and reset
Gemini	Past chats, plus whatever Gmail, Photos, YouTube and Search reveal	You, and your Google account activity	Yes, with a toggle per connected app
DeepSeek	Account-level chat history, retained as long as the account exists	Your chat history	History yes, with a stated right to opt out of training
A relationship memory app	Dated notes about the people you actually know	Each person	Yes, and exportable or deletable per note

The categories give it away

Claude publishes what its memory is for: your role and professional context, your communication preferences and working style, your technical preferences, and your project details. Four categories, four kinds of fact about the person typing.

OpenAI is just as direct. Announcing Dreaming in June 2026, it described memory as what helps ChatGPT learn “your preferences, projects, and constraints” and set three goals for it: carry your context forward, follow your preferences, and stay current as time passes. The worked examples in that announcement are a vegetarian diet, a trip to Singapore, a birthday party next Saturday, and an underwater camera rig. Every one is a fact about the user.

This is not an oversight, it is economics. Memory earns its keep by lowering the cost of the next prompt, and a fact about you applies to every conversation you will ever have. A fact about your friend’s mother applies to one, and possibly never again. Ranked by that logic, the people in your life lose every time.

A thread is not an index

Even where the raw material is there, the shape is wrong. Chat history is stored the way it happened, as a list of sessions with the newest first. You can search it, but search only helps when you already remember there is something to find.

The question that matters months later is rarely “what did I discuss in March”. It is “what do I know about Priya”, and no assistant organizes around that. This is the same structural problem a notes field has, and it is covered in more depth in AI that remembers conversations.

Gemini is the exception, and how it gets there matters

Gemini is the one assistant that reaches for other people on purpose. Personal Intelligence connects Gmail, Google Photos, YouTube and Search, and Google says it can offer insight into you, your relationships, and the people in your world. Its memory import will even carry across a detail like a sibling’s name from another assistant.

That is a real capability. It is also worth understanding where the material comes from, because Personal Intelligence reads your digital exhaust. It knows the people who generate traffic inside your Google account, which is a specific and quite skewed sample of the people in your life.

The colleague who copies you on forty threads a quarter is vivid. The friend you had dinner with on Sunday, who mentioned their mother’s scan came back clear and that they are quietly interviewing again, produced no email at all. The relationships that matter most are usually the ones conducted in person, and an in-person conversation leaves nothing behind for an inbox to read.

One practical limit worth knowing: these connected experiences are for personal Google accounts, not Workspace business, enterprise or education accounts. If your working life runs on a Workspace account, the feature does not reach it.

The transfer test

The fastest-rising question about AI memory in the US right now is how to move ChatGPT’s memory into Claude. Anthropic added an import for exactly that in May 2026, covering ChatGPT, Gemini and Copilot, and the mechanics are simple: run a prompt in the old assistant, then paste the output into the new one.

What arrives is a summary of preferences. Your tone, your role, your ongoing projects, the instructions you kept having to repeat. What does not arrive is the substance of eighteen months of conversation, because it was never held as substance. It was held as a description of you.

That makes the transfer the cleanest test of what this kind of memory really is. A store that moves in a single paste is a paragraph, not a database.

The split that works

Use the assistant’s memory for the job it was built for. It is genuinely good at that. It stops you re-explaining your role, your constraints and your preferred format at the start of every session, and drafting quality goes up once it knows how you write.

Then keep people somewhere organized around people, where the unit is a person rather than a session, the note is dated and attached to them, and it outlives whichever assistant you happen to be using this year.

Here is the kind of thing that has no home in a chat profile:

Coffee with Priya, 3 June. Left the agency, consulting solo now and underpricing herself. Mother’s scan came back clear. Wants an intro to someone who has run a two-sided marketplace. Said to check in after her Lisbon trip in August.

Paste that into any of the four and you will get a warm, usable draft back today. Ask any of them about Priya in November and you are starting from nothing, because none of them filed it under Priya. An app built for relationship memory does file it under Priya, and answers from what you wrote rather than filling the gap with something plausible.

The two are complements, not rivals. Keep the drafting where it is good, and keep the people somewhere built to hold them. For the difference between recall built on your own notes and context imported from elsewhere, see AI relationship memory vs contact enrichment.

Key takeaway: ChatGPT, Claude, Gemini and DeepSeek all remember you, not your people. Their memory is a profile built to make the next prompt cheaper, it is organized by session rather than by person, and the one that does reach for relationships gets them from your inbox, which is blind to the conversations that happen face to face.

FAQ

Does Claude remember previous conversations?

Yes. Claude keeps memory as individual entries covering your role, working style, technical preferences and project details, and it updates them as you chat. You can read, edit and delete those entries in settings, pause memory to stop it growing, or reset it to clear everything. Each project keeps its own memory space, while chats outside a project share a common pool.

What does ChatGPT’s memory actually store?

Two things. Saved memories are details you explicitly asked it to keep, and Dreaming is a background process that synthesizes memory from your chats without you asking. Both aim at the same target: your preferences, projects and constraints. You can review the result on the memory summary page and edit it there.

Can I move my memory from ChatGPT to Claude or Gemini?

You can move a summary of it. Claude added an import in May 2026 for ChatGPT, Gemini and Copilot, and Gemini offers its own import. What transfers is your preferences and context, not the detail of past conversations, so treat it as carrying over a profile rather than an archive.

Is DeepSeek safe for personal notes?

DeepSeek’s privacy policy states that it collects, processes and stores personal data in the People’s Republic of China, and that it uses content to train and improve its models, alongside a stated right to opt out of that training. Its policy also asks users not to provide sensitive personal data about themselves or anyone else. That is worth reading closely before pasting anything personal, and it applies doubly to notes about other people, which is covered in telling AI about other people.

Which AI assistant has the best memory for relationships?

For remembering people specifically, none of the general assistants are built for it. Gemini comes closest by connecting Gmail and Photos, but it only sees relationships that generate activity in your Google account. A dedicated relationship memory app indexes by person and works from what you captured yourself, including the conversations that never touched email.

Should I just paste my notes into a chat each time?

It works for a single draft and it does not scale. You are limited by how much you can paste, you have to remember which notes are relevant before you can include them, and nothing accumulates. That is the gap a person-indexed store fills, as covered in can ChatGPT be your personal CRM.

Intriq keeps dated notes filed under each person, answers from what you actually wrote, and stays private by default. Start with the personal CRM hub or the AI relationship assistant hub.