How to Migrate Your NotebookLM (Now Gemini Notebook) Notebooks to ChatGPT Without Losing Context (2026)

Gemini Notebook is a grounded question-answering box. Google is unusually direct about the boundary: "Gemini Notebook answers questions based on the information provided in your uploaded sources. If the answer isn't in the source material, it won't provide a response."

That refusal is the product's best feature, and it is the first thing you lose. ChatGPT's documented personalization surfaces work the other way around — the memory summary is, per OpenAI, "a continually updated synthesis of context from your past chats," and saved memories are "always considered in future responses" unless deleted. Synthesis is the opposite of refusal-to-extrapolate. Neither behaviour is wrong; they are answers to different questions. But if your notebook's value came from knowing that every claim traced to a source you uploaded, that guarantee does not come along.

Now the mechanics, which are where the losses actually happen.

Your uploaded files transfer, if you still have them. Files you uploaded directly became copies inside the notebook. Google's limits tell you the shape: "Each source can contain up to 500,000 words or up to 200MB for uploaded files. You can include up to 50 sources (for Free users)."

Your Drive-linked sources may not be yours to move. These are the auto-synced kind: "Sources imported from Google Drive are auto-updated and will sync every few minutes. Changes to your original document will automatically update when you open your Notebook." Convenient, and conditional. Google spells out the condition: "You can only import files if you have view access or more. If you lose access to a file in Google Drive or if the file is deleted, the source will be inaccessible, and you will no longer be able to view or interact with the source in your notebook. This applies to both owned and shared notebooks."

Read that last sentence twice. A notebook you own can lose a source because somebody else changed a sharing setting. If a colleague's spec was one of your fifty sources, your notebook's answers quietly stop being grounded in it.

Your YouTube sources are the most fragile. Google supports "only public YouTube videos with captions, either user-uploaded or auto-generated," and notes that "videos uploaded less than 72 hours prior may not be available to import." A video that goes private takes its transcript with it.

Your questions and answers do not transfer as a mechanism. The notebook's job was retrieval: "When your notebook contains many sources, Gemini Notebook retrieves the most relevant information based on your question, then builds a response from it." The retrieved answers were generated, not stored as canon.

Your conclusions transfer only if you wrote them down. This is the one that hurts. Everything you learned by reading fifty sources — which vendor's claim held up, why you discarded the second approach, what the numbers actually meant — existed as chat output. It was never a source. The distinction is the same one covered in why RAG isn't memory: retrieval finds what is written; it does not remember what you decided.

Step 1: Separate the copies from the pointers, before anything changes

Open each notebook and go source by source. For every one, answer a single question: is this a file I uploaded, or a Drive document I linked?

Uploaded files are copies. You control them, and if you have lost the original, the notebook is now the only place it exists — get it out first.

Drive links are not copies. Re-download those documents from Drive while you still have access, because the notebook's version is a synced reflection, not an archive. If you are working on a shared notebook, do this for sources somebody else contributed too; the access that makes them readable today is somebody else's setting.

While you are there, check for a quieter problem. Google notes that "inactive sources will count towards source limits but will not be referenced throughout your notebook." An inactive source is occupying one of your fifty slots and contributing nothing to answers — which means any conclusion you thought it supported was not actually grounded in it.

Step 2: Write down the conclusions, then rebuild the reference layer in ChatGPT

Before you touch ChatGPT, spend twenty minutes per notebook writing out what you concluded. Not summaries of the sources — the decisions. "We went with the second vendor because the first one's SLA excluded the failure mode we actually see." "The 2024 figures are not comparable to 2025 because the definition changed." These are the sentences you will otherwise re-derive from scratch, and they are the only part of a notebook that was never stored anywhere. The sorting pass this calls for is the same one described in turning project docs into AI memory, applied to research rather than to a codebase.

Then move the material into ChatGPT in two lanes, because OpenAI documents two different surfaces.

Standing facts and preferences go into Custom Instructions, which OpenAI describes as being for explicit information and instructions you want applied. Keep this small; it is in play on every response.

Reference documents go in as uploaded files, ideally inside a Project so the scope is contained. OpenAI documents how that scope behaves: with project-only memory, "chats can reference other conversations in the same project, but they cannot reference conversations outside the project, and chats outside the project cannot reference conversations in it." That gives you a per-topic boundary roughly analogous to a notebook, without the source ceiling.

What you will notice immediately is that nothing enforces grounding any more. Gemini Notebook would decline to answer beyond its sources; ChatGPT will happily synthesize. That is a real behavioural change and the reason the conclusions you wrote in the previous paragraph matter so much — they are now the thing keeping the answers anchored. The file-handling side of this has its own well-documented rough edges, covered in when ChatGPT forgets uploaded files.

A notebook is two things stacked together: a pile of reference material, and the understanding you built on top of it. Only one of those is replaceable. You can re-upload a PDF. You cannot re-derive, cheaply, the reason you rejected an approach eight months ago.

So split them deliberately. Let the reference pile live wherever it is convenient — a Project, a drive, a notebook. Put the understanding somewhere addressable that does not depend on one vendor's source limit, one Drive permission, or one product rename.

That is what MemoryLake is for: the decisions, constraints and conclusions your research produced, held in one place every assistant can read, so the next question starts from what you already know instead of from fifty documents.

Step 1: Create an API key

Sign in, open your workspace settings, and generate an API key. This is the credential your assistants use to read the same layer, so create it once and keep it reachable from each tool you research in.

Step 2: Upload your first memories

Start with the conclusions you just wrote out, one per entry, each with the reason attached. Add the constraints that are easy to forget and expensive to rediscover: which datasets are comparable, which source is authoritative when two disagree, which question you already answered and closed.

Step 3: Connect your AI & agents

Connect ChatGPT and whatever else you work in. The reference files can stay wherever they live. The reasoning arrives everywhere, and a new chat opens already knowing what the last six weeks of reading established.

The first difference is that a permissions change stops costing you knowledge. Google is explicit that losing access to a Drive file makes the source inaccessible in your notebook, owned or shared. When the conclusion is recorded separately from the document that produced it, the conclusion survives the document.

The second is the source ceiling. Fifty sources on the free tier is a real constraint, and the usual workaround — delete an old source to make room — is exactly the move that erases the basis for something you already concluded. Keeping the conclusions outside means pruning sources becomes a storage decision instead of a memory loss.

The third is the rename itself. NotebookLM became Gemini Notebook, and product identities will keep moving. Anything you hold in a vendor's container inherits that vendor's roadmap. The same pattern shows up with Gemini forgetting project context at the assistant level rather than the notebook level.

The fourth is the one you will feel weekly: you stop re-uploading. The reflex of dragging the same three PDFs into a new chat because the assistant cannot recall what they said is the problem described in how to stop re-uploading PDFs, and it is solved by holding the extracted conclusions rather than the source files.

Export the copies first, the pointers second. Uploaded files may exist nowhere else. Drive-linked sources still exist in Drive, for as long as your access does.

Treat inactive sources as a warning. They count against your limit and are not referenced in answers. Any belief you hold "because it was in the notebook" needs re-checking if its source was inactive.

Write the conclusion next to the citation. "Vendor B, per their 2025 security page, excludes this failure mode" is durable. "See source 14" stops meaning anything the moment the source list changes.

Re-check anything that came from a video. Public-with-captions is a condition that can lapse without notice, and a transcript you never saved is a transcript you no longer have.

Expect the destination's memory list to be partial. OpenAI notes that the memory summary "will not include everything that ChatGPT remembers based on your chats" and suggests asking in chat if you want to know whether something was retained. Do not treat the visible list as an inventory.

Keep one authoritative version of each fact. A notebook naturally accumulates three documents that disagree. Deciding which one wins, once, in writing, is worth more than any amount of re-reading.

Gemini Notebook is very good at a specific thing: answering only from what you gave it, and telling you when it cannot. That discipline is why people built serious research notebooks in it, and it is also why the migration feels lossy — ChatGPT's documented surfaces synthesize rather than refuse, and the guarantee does not port.

What ports is whatever you wrote down. A source is a copy or a pointer, and Google says so in the first line of its own definition. Your conclusions are neither, which is exactly why they need a home of their own before you start moving boxes.

If the memory side rather than the notebook side is what you are moving, our guide on migrating Gemini memory to ChatGPT covers that path instead.