Your Vault Does Not Fit in the AI Notebook — Four Vendors' Own Limits, and What Each One Does With the Notes Past Them

If you use any of the major AI notebooks — Google's Gemini Notebook, ChatGPT Projects, Claude Projects, or Microsoft Copilot Notebooks — you have a file cap. On a free tier, it can be as low as five files per project. On the top paid plan, it can reach 600 sources. A vault of a few hundred notes will exceed at least one of those ceilings, and when it does, each tool responds differently. Two refuse. One silently excludes. One switches from reading everything to searching for what it thinks is relevant.

The four behaviours matter more than the four numbers. A refusal is visible. A silent switch is not.

Every figure in this post was read on the vendor's own help page on 2026-10-28. The numbers will change. The behaviours are less likely to.

The cap numbers, from the vendors themselves

Every AI notebook publishes the limits on its own help pages, but they are scattered across different articles, and the third-party summaries that appear in search results often disagree with each other. The table below reads from primary documentation only — four vendors, their stated caps, their stated file sizes, and whether markdown appears in their supported type list (all figures read 2026-10-28).

Three things stand out immediately. ChatGPT's free plan is the tightest at five files per project — tight enough that most people hit it on first use. Gemini Notebook's free cap of 50 sources is an order of magnitude looser, but 50 notes is still a small fraction of most active vaults. Claude states no file count at all, which looks permissive until you read what happens at the context window boundary.

The markdown question deserves its own note: none of the four vendors explicitly names .md as a supported file type on the help pages read on 2026-10-28. Microsoft's page lists specific types and .md is not among them.⁷ OpenAI's article names DOCX, PDF, and TXT.⁴ Google and Anthropic do not enumerate a type list on the pages read. If your vault is raw .md files, verify your tool's actual import behaviour before assuming they will be accepted.

Four tools, four different endings

When a tool hits its ceiling, four behaviours are possible: refuse the import, display an error on upload, silently exclude some content from grounding, or switch from reading all files to retrieving only the relevant passages. Each of the four tools does something in that range — and not the same thing.

Gemini Notebook refuses to import sources that exceed the plan cap. For sources you have already added but subsequently lost access to — because a Google Drive file was deleted or a sharing permission was revoked — the tool accepts them as occupying a source slot but stops consulting them. Google's own help page puts it precisely: "Inactive sources will count towards source limits but will not be referenced throughout your notebook."⁸ This statement describes sources for which you lost Drive access or deleted; it does not describe over-cap behaviour. Sources that exceed the plan cap fail to import. Sources you lose access to after adding are counted but not consulted.

ChatGPT Projects returns a visible error when you reach the storage cap.⁴ Uploads are also rate-limited: free users are capped at three file uploads per day, and all users share a rate limit of 80 files every three hours.⁴ You can only upload 10 files at a time.⁹ What happens when the total content of your project files exceeds ChatGPT's context window — not the file-count cap but the token ceiling — is not documented on the three OpenAI help pages read on 2026-10-28. Whether those files are read in full or searched is not stated in the current documentation.

Claude Projects behaves differently depending on your plan, and that difference is significant. On free plans, projects hit a hard context window limit. Anthropic's own documentation states: "Previously, projects had a knowledge capacity limit based on the context window, meaning it wasn't possible to add more content once the limit was reached."⁶ On paid plans — Pro, Max, Team, and Enterprise — Claude automatically enables a retrieval mode when your project knowledge approaches that limit. Anthropic describes it this way: "Now, when your project knowledge approaches the context window limit, Claude will automatically enable RAG mode to expand your project's capacity by up to 10x."⁶ The mechanism: "Instead of loading all project content into memory at once, Claude intelligently searches and retrieves only the most relevant information needed to answer your questions."¹⁰ The retrieval switch is invisible. There is no notification that the tool has stopped reading all your files and started searching them instead.

Microsoft Copilot Notebooks operates on a two-tier system. Copilot Chat users can add up to 50 references, and all 50 are used for grounding.⁷ Microsoft Copilot users can add more than 300 references, but only the first 300 are used. Microsoft's documentation states: "While Microsoft Copilot users can add more than 300 references to their notebook, only up to the first 300 will be used for grounding."⁷ References beyond 300 are accepted without error and silently excluded from every answer the tool produces.

What the retrieval switch actually means

The paid-Claude behaviour — silently switching from reading every file to retrieving relevant passages — deserves honest framing before treating it as a straightforwardly good thing.

Retrieval-augmented generation is a trade-off, not an upgrade. Research published at EMNLP 2024 found that "when resourced sufficiently, LC [long-context] consistently outperforms RAG across all six tasks in our study."¹¹ A separate study accepted at ICLR 2024 found that "a 4K context window with retrieval can be nearly as performant as a 16K context window that has been finetuned" — retrieval is useful when context is constrained, not a strict improvement over full-context reading.¹²

In plain terms: if the AI reads every file, it has all the context. If it retrieves relevant passages, it has what its retrieval index judged relevant for this particular query. A note that does not surface in the index for a given question will not influence the answer, even if it contains the fact you needed. The retrieval index is not your vault.

The canary-in-the-vault test is one way to measure this gap: plant a plausible fact in a note near the end of your bundle and ask for it by name. A silent miss — the AI not returning the fact — tells you that this note was not consulted for this query. If the tool switched to retrieval without telling you, that test becomes more important, not less.⁶

How to decide what goes in the bundle

The practical response to any cap is to decide which notes go in, build the bundle intentionally, and test that the tool read what you gave it.

The standard workaround is to combine topic-related notes into fewer, larger files before uploading — fewer source slots consumed, more context per slot. A single file holding a full project's notes uses one Gemini source slot; the same content split across fifty individual notes consumes fifty slots on plans that allow it, and immediately exceeds the cap on plans that don't. Bundling by topic keeps related context together, which also helps retrieval-based tools find the right passages.

For the selection problem — which topics go in at all — treat the AI notebook as bounded context rather than a full-vault search. Fill that context with the notes most likely to bear on the questions you plan to ask. The notes you leave out will not influence the answers, silently or otherwise.

If you want to build a retrieval pipeline where you control exactly which notes are indexed, can inspect the index, and can audit which passages were retrieved for any answer, the approach is described in A Personal RAG You Can Actually Audit. A hosted notebook is a shortcut. A folder of Markdown files and an open retrieval stack is the auditable version.

A folder of Markdown files stored on your own device has no source cap. You choose the bundle. You choose the retrieval logic. Nothing is silently excluded from grounding because a plan tier was exceeded.

Frequently asked questions

How many sources can I add to Gemini Notebook (NotebookLM)?

On a free Google account, the limit is 50 sources per notebook.¹ The cap rises with your Google AI Plan: Standard gives 50 sources, Plus 100, Pro 300, and Ultra plans allow 500 to 600 sources depending on the storage tier.² Each source can hold up to 500,000 words or 200 MB.¹ All figures read on Google's own help pages, 2026-10-28.

How many files can I upload to a ChatGPT project?

Free accounts can add up to 5 files per project. Go and Plus plans allow 25. Edu, Pro, Business, and Enterprise plans allow 40.⁴ Separate GPTs allow up to 20 files each.⁵ Free users are also limited to 3 file uploads per day, and the per-account rate limit is 80 files every three hours.⁴

What is the Claude project knowledge limit?

Claude states no file count cap. The limit is the context window. On free plans, that is a hard stop — the documentation says adding more content was not possible once the limit was reached.⁶ On paid plans (Pro, Max, Team, Enterprise), Claude automatically enables retrieval mode, expanding capacity "by up to 10x."⁶ Retrieval mode is paid-only; free projects still hit the context-window wall.

How many files can I add to Microsoft Copilot Notebooks?

Copilot Chat users can add up to 50 references, and all 50 are used for grounding. Microsoft Copilot users can add more than 300 references, but only the first 300 are used for grounding.⁷ References beyond 300 are accepted but silently excluded from answers. Markdown (.md) is not listed among the supported file types in Microsoft's documentation (read 2026-10-28).⁷

What happens when my AI notebook is full or reaches the limit?

It depends on the tool. Gemini Notebook fails to import sources past the plan cap. ChatGPT Projects displays a storage error when you hit the file cap.⁴ Claude on paid plans silently switches from reading all files to retrieving relevant passages.⁶ Microsoft Copilot Notebooks accepts references past 300 but silently excludes them from grounding.⁷ Only the error and the import failure are visible. The retrieval switch and the silent exclusion are not.

Can I upload Markdown files to these AI notebooks?

Markdown (.md) is not named in Microsoft Copilot Notebooks' listed supported file types.⁷ OpenAI lists DOCX, PDF, and TXT — no .md.⁴ Google and Anthropic do not enumerate a type list on the pages read (2026-10-28). The absence of .md from a stated list is not an explicit prohibition, but it is also not confirmation of support.

How do I test whether the AI notebook actually read all my notes?

Plant a specific, plausible fact in a note near the end of your upload set and ask the tool for it by name. A silent miss — the tool not returning the fact — does not prove the retrieval pipeline is broken for all queries, but it tells you this note was not consulted for this one. The full test design and its documented failure modes are in Bury a Canary in Your Notes to Test Your RAG.

References

The cap is not a flaw in the product. It is the product's honest edge — the point where the tool stops being a full-vault assistant and becomes a bounded context with a retrieval heuristic. Know where the edge is, build your bundle before you hit it, and test whether the notes that matter were actually read.

MNMNOTE keeps your notes stored locally on your own device — no source cap, no silent exclusions, open Markdown you can bundle however you need. mnmnote.com