Why AI forgets you, and how to keep a memory that travels

Last verified 23 July 2026. AI memory features change quickly; dated claims were checked on this date.

You told it everything last week. Today it asks who you are. This post explains why assistants forget, what each one actually keeps, what you can and cannot move between them, and what a memory that belongs to you would look like.

The short answer

AI assistants forget because most of what you tell them lives in a temporary working space, not in storage. ChatGPT, Claude and Gemini each now offer a memory feature, but each remembers different things, by different rules, in different regions. Moving memory between them is partly possible, with real limits. A memory you hold yourself, outside any one assistant, is the only version that survives switching.

Why assistants forget: three different things called “memory”

When people say an AI “remembers” or “forgets”, they usually mean one of three mechanisms, and the confusion between them causes most of the frustration.

The first is the context window. This is the assistant’s working memory: the text of your current conversation, held only while you are in it. It is why an assistant can follow a long discussion perfectly and then know nothing next morning. Nothing in the context window persists on its own.

The second is saved memories. These are specific facts the assistant stores deliberately: you are vegetarian, you live in Manchester, you prefer short answers. You can usually see and delete these in settings.

The third is chat-history reference. The assistant reads back across your past conversations and builds an impression of you. You did not dictate this profile and you cannot always see all of it.

An assistant can have any combination of the three switched on. When it “forgets”, the thing you told it usually lived in the first mechanism and never reached the other two.

What ChatGPT, Claude and Gemini actually remember

As of 23 July 2026, verified against each vendor’s own documentation:

Feature availability changes often and varies by plan and region. Checked 23 July 2026.

Can you move your memory between assistants?

Partly, and the honest answer is more useful than the headlines.

Anthropic added an import tool to Claude (in Settings, under Capabilities) that reads an export from ChatGPT, Gemini or Grok and builds Claude memories from it. It takes up to around a day to process, and Anthropic labels it experimental. Independent hands-on testing found the result is not a copy but a re-interpretation: Claude reads what the export says about you and writes its own version.

What actually transfers is a snapshot of preferences and facts, once, in one direction. Your conversations do not move. The inferred profile does not move. Nothing stays in sync afterwards: the day after importing, your two assistants begin drifting apart again.

Two further limits matter for readers here. ChatGPT’s data export (Settings, then Data Controls, then Export) gives you your conversation transcripts, but saved memories have historically not been included in that export and there is no memory API. And Gemini’s import routes were not available in the UK or EEA at the time of checking.

So the accurate summary: you can carry a one-time, re-interpreted snapshot from one assistant to another, in some regions, on some plans. You cannot yet carry a living memory.

The portable alternative: memory that lives outside the assistant

This is not a hypothetical concern. Rewind AI became Limitless, Meta acquired Limitless in December 2025, and the service ended in the UK and the EU on the day the acquisition was announced. The Rewind desktop app stopped capturing two weeks later, on 19 December 2025.

There is a second approach: keep the memory yourself, in a store you control, and let assistants read it rather than own it.

The mechanism that makes this practical is the Model Context Protocol, an open standard that lets AI assistants use outside tools. A memory store that speaks MCP can serve the same facts to Claude today, ChatGPT tomorrow, and whatever you use next year. Switching assistants stops costing you your history, because the history was never inside the assistant.

AI Twin is built on this model: a typed personal memory you fill deliberately, held in one place with an audit log, exposed to the assistants you choose through MCP. It is one example of the approach rather than the only one, but the approach itself is the point: the model is swappable, the memory is yours.

Questions people ask

Can I export my ChatGPT memory?

You can export your conversation transcripts from Settings, under Data Controls. Saved memories have historically not been included in that export, and there is no memory API. The practical route to reusing them elsewhere is an import tool on the receiving side, such as Claude’s, which re-interprets rather than copies.

Does moving memory to Claude move my conversations?

No. The import reads an export and writes Claude’s own version of your preferences and facts. Your conversation history stays where it was, and the two assistants do not stay in sync afterwards.

Why does ChatGPT forget what I told it last week?

Most of what you type lives in the context window, which is working memory for the current conversation only. Unless the detail was written into a saved memory or picked up by chat-history reference, it was never stored.

Does Claude remember previous conversations?

Yes. Claude’s memory from chat history rolled out through 2025 and reached all users, including free accounts, on 2 March 2026, switched on by default, with settings to view, edit or disable it and an incognito mode for conversations you want kept out.

Is AI memory available in the UK?

Unevenly. ChatGPT’s chat-history reference was excluded from the UK and EEA at its April 2025 rollout, and Gemini’s memory-import routes were not available in the UK or EEA when we checked. Claude’s memory is available in the UK. Always check the current state; this page shows its verification date above.