ChatGPT Memory Does Not Remember What You Read
People increasingly assume their AI already knows what they have been reading. It does not, and the difference matters if you are deciding where to keep the things you save.
What ChatGPT memory actually stores
OpenAI's own Memory FAQ describes two layers:
- Saved memories: details you ask ChatGPT to remember, or that it decides are useful, like "prefers TypeScript" or "is planning a trip to Lisbon". They live separately from your chats.
- Reference chat history: ChatGPT can draw on relevant information from past conversations. OpenAI notes this changes as ChatGPT updates what it considers useful, so it is not a fixed record.
It can also use files you add and content from apps you connect. Claude has comparable features: memory of past conversations and projects with files you upload.
Notice what is missing: the web pages you read in your browser. If you never pasted an article into a chat, your assistant has no copy of it. The one partial exception is OpenAI's Atlas browser, which can keep opt-in browser memories of pages you visit inside Atlas. That covers one browser, with details the AI chose to keep, not the page itself.
Memory remembers you. A library remembers sources.
These are two different jobs.
| AI memory | A saved library | |
|---|---|---|
| What it keeps | Facts about you, distilled from chats | The pages, videos and PDFs you chose to keep |
| Who decides | The model, mostly | You, at save time |
| Exact text | No, a summary at best | Yes, if the tool stores it |
| Can it cite a source? | No source to cite | Yes, the saved page |
| Survives the page dying | Not applicable | Only if the text was kept |
Memory is good at "remember I'm vegetarian". It is bad at "what did that benchmark article say about Postgres 18 write latency?" The second question needs the article, word for word, and a link back to it.
Why "just ask the AI" fails on your own reading
When you ask an assistant about something you read, one of three things happens:
- It searches the web and finds something similar but not your source.
- It answers from training data that may be out of date.
- It confidently paraphrases something it never saw.
None of those is your source. If you are writing, deciding or citing, you need the page you actually read.
What to do instead
Keep a library of what you read, with the text, and ask that library.
- Save as you go. One link at a time, or the whole window when a research session ends. In Markwise, the Chrome extension saves the window as a Focus Session, and the important pages keep their text. See Close your tabs. Your AI remembers them.
- Search by what you remember. Search by meaning finds "the article about shrinking Docker images" when you type "container optimization".
- Ask with citations. Markwise Copilot (Pro) answers from your saves and lists the ones it used, so you can open the source.
You can also ask from inside ChatGPT or Claude. Add Markwise as a connector (an MCP server at https://api.markwise.app/mcp), sign in, and your assistant can search your saved pages, read their text and cite them. In ChatGPT, custom connectors need a Business, Enterprise or Edu plan; Claude allows one on its free plan. Recall and Readwise ship connectors too; our Recall comparison and Readwise Reader comparison cover the trade-offs.
The short version
Your AI's memory is about you. Your reading is about the world. Keep the reading somewhere that stores the words and can show its sources, and let the memory feature do what it is good at.