# Custom GPT Knowledge Files: Limits, Formats and Tips > Custom GPT knowledge files explained: how many files you can upload, size limits, which formats work best, and what to do when the knowledge keeps changing. Canonical: https://valta.si/blog/custom-gpt-knowledge-files [Memory](https://valta.si/blog/category/memory) Valta · October 26, 2026 · 5 min read Custom GPT knowledge files are documents you upload in the GPT editor’s **Knowledge** section so the GPT can search them when it answers. OpenAI’s help text allows up to 20 files per GPT, up to 512 MB each, and up to 2 million tokens per text or document file. The GPT doesn’t read every file on every message: it searches them and pulls in the parts that match the question. That makes knowledge files good for reference material that rarely changes, and awkward for anything you update every week. This guide covers the limits, the formats that work best, and the trade-offs. ## How do you add knowledge files to a custom GPT? 1. In ChatGPT, open **GPTs** and click **Create**, or edit a GPT you own. 2. Go to the **Configure** tab. 3. Under **Knowledge**, click **Upload files** and pick your documents. 4. Write in the **Instructions** when the GPT should use them, for example “Answer pricing questions from pricing.md.” 5. Click **Update** (or **Create**) to save. The instruction in step 4 matters more than people expect. Without it, the GPT may answer from general knowledge instead of searching your files. ## What are the limits on custom GPT knowledge files? | Limit | Value | |---|---| | Files per GPT | Up to 20 | | Size per file | Up to 512 MB | | Text per file | Up to 2 million tokens (text and documents, not spreadsheets) | These figures come from OpenAI’s help documentation and can change, so check the [GPTs FAQ](https://help.openai.com/en/articles/8554407-gpts-faq) before building around them. In practice, the file count fills up before the size does: 20 files go fast if you upload one document per topic. ## Which file formats work best? GPTs accept common formats: PDF, Word, plain text, Markdown, CSV, spreadsheets and more. Some read better than others: - **Markdown and plain text** search well. Headings split the content into clear sections the GPT can match to a question. - **PDFs** work when they contain real text. Scanned pages without text come back empty. - **Spreadsheets and CSV** work best with Code Interpreter turned on, so the GPT can calculate instead of guessing. - **Slides** often lose their meaning, because the text on a slide is short and the context was in the talk. If you have the choice, convert to Markdown with clear `##` headings and one topic per section. [Markdown for beginners](https://valta.si/blog/markdown-for-beginners) covers the basics. ## How do you get better answers from knowledge files? - **Merge small files.** Since the limit is the number of files, combine related notes into one file with a heading per topic. - **Name sections the way people ask.** A heading like “## Refund policy” matches “can I get a refund?” better than “## Section 4.2”. - **Remove what’s outdated.** The GPT can’t tell an old price list from a new one. Delete the old file instead of uploading a second version. - **Point to the file in the instructions.** “For brand voice, follow brand-voice.md” works better than hoping it finds the right file. - **Test with real questions.** Ask the questions your users will ask and check the answer against the file. ## Can people download a custom GPT’s knowledge files? They can sometimes get the content. If Code Interpreter is on, a user can ask the GPT to list or share its files, and it may comply. Even with it off, a user can ask the GPT to quote its sources and get large parts of the text back. Treat anything you upload to a shared or public GPT as readable by its users. Keep private data, client details and anything under NDA out of it. [How to share ChatGPT with your team](https://valta.si/blog/how-to-share-chatgpt-with-your-team) covers sharing GPTs inside a workspace. ## What’s the problem with knowledge files that change? Knowledge files are a snapshot. When your pricing, offer or brand voice changes, the GPT keeps answering from the old file until you upload a new one by hand. And the files only work inside that one GPT: your regular ChatGPT chats, Claude and Cursor can’t see them. That’s fine for a style guide you edit twice a year. It’s a chore for working notes that change every week, such as decisions, client details or current projects. ## What’s an alternative for knowledge that keeps changing? Keep the changing part in a notes vault that your AI reads live, and keep the stable reference in the GPT. Valta is a vault of Markdown notes that ChatGPT, Claude and Cursor connect to over MCP. You add it once in ChatGPT, and then in any chat: - Ask it to save something: “save this as our current pricing.” - Ask it to find something: “what’s our refund policy?” It searches the vault and answers from the latest note. - Update a note in one chat, and the next chat, in ChatGPT or Claude, reads the new version. | | Custom GPT knowledge files | Valta vault | |---|---|---| | How it’s updated | You re-upload files by hand | Your AI saves and edits notes as you work | | Where it works | Inside that one GPT | ChatGPT, Claude, Cursor and other MCP apps | | Limit | 20 files per GPT | Free: 100 notes. Pro: unlimited (fair use 10,000) | | Format | Many file types | Markdown notes with tags and links | | Export | Your original files | Markdown zip any time | [ChatGPT custom instructions](https://valta.si/blog/chatgpt-custom-instructions) and [ChatGPT memory full](https://valta.si/blog/chatgpt-memory-full) cover the other two places ChatGPT keeps context, and [the setup guide](https://valta.si/blog/connect-valta-to-claude-chatgpt-and-cursor) connects a vault in a few minutes. ## FAQ ### How many knowledge files can a custom GPT have? Up to 20 files per GPT, according to OpenAI’s help documentation. ### Does a custom GPT read all its knowledge files on every message? It searches the files and pulls in the sections that match the question, so a well-structured file with clear headings gets better answers. ### Can I update a knowledge file without rebuilding the GPT? Delete the old file in the Knowledge section, upload the new one and click Update. The rest of the GPT stays the same. ### Do custom GPT knowledge files work in regular ChatGPT chats? They’re only available inside the GPT you uploaded them to. Projects have their own files, and regular chats can’t see either. ### What’s the best format for custom GPT knowledge files? Markdown or plain text with clear headings, one topic per section. PDFs work when they contain real text, not scans. ## Give every AI the same up-to-date notes [Create your free Valta vault](https://valta.si/signup): one vault, 100 notes, no card. Your AI saves the notes that change, and finds the latest version in ChatGPT, Claude or Cursor. **About this article.** Written by the team that builds Valta, a vault of Markdown notes that Claude, ChatGPT and Cursor can all read and write. Spotted something wrong or out of date? Write to [support@valta.si](mailto:support@valta.si). ## One vault for every AI you use. [Start free](https://valta.si/signup) More in Memory [MEMORY](https://valta.si/blog/category/memory) Oct 26, 2026 · 5 min ## [How to Search ChatGPT History on Web and Mobile](https://valta.si/blog/how-to-search-chatgpt-history) How to search ChatGPT history on web and mobile, why some chats don't show up, and how to keep the answers you'll need again where any AI can find them. [MEMORY](https://valta.si/blog/category/memory) Oct 20, 2026 · 5 min ## [Claude vs ChatGPT: Which to Use for Content Work](https://valta.si/blog/claude-vs-chatgpt) Claude vs ChatGPT for creators and solo marketers: which wins for writing, research, images and voice, and how to use both without losing context between them. [MEMORY](https://valta.si/blog/category/memory) Oct 20, 2026 · 5 min ## [ChatGPT to Claude Migration: Move Memory, Chats and Context](https://valta.si/blog/chatgpt-to-claude-migration) A ChatGPT to Claude migration guide: how to transfer ChatGPT memory to Claude, what happens to your chats, and how to stop losing context next time.