Why built-in AI memory doesn’t cover what you watch
Built-in memory in ChatGPT and Claude is a list of things the assistant has learned about you: your job, your projects, how you like answers written. It lives inside one vendor’s product and it is short by design. A 90-minute podcast is tens of thousands of words. Nobody’s assistant memory holds that, and pasting a transcript into a chat only helps for that one conversation.
What you need is a second kind of memory: a library of the sources themselves, searchable by meaning, that any assistant can query when a question needs it. For media that means one thing above all: the transcript.
Step by step
- Pick what should be remembered. Choose the channels, playlists and podcasts you actually learn from. A library of what matters answers better than an archive of everything you clicked.
- Save videos and episodes. In BrainTube, save a video from the Chrome extension or paste a YouTube or podcast link. Each item is transcribed and indexed.
- Follow channels and feeds. Follow a YouTube channel or a podcast RSS feed and new uploads and episodes are added on their own (3 sources on Free, 20 on Starter, 100 on Pro).
- Connect your AI. In Claude.ai add https://mcp.brain-tube.com/mcp as a custom connector and sign in with OAuth. In Claude Code, add the same URL with claude mcp add. Cursor and other MCP clients use the same server.
- Ask in plain words. Ask about an idea, not a title. Search matches meaning and keywords across every transcript, and answers in BrainTube's own chat link to the timestamp the quote came from.
What happens when you save a video or an episode
A link on its own is opaque: a title and a thumbnail. BrainTube treats a saved video or episode as a document. It keeps the transcript, writes a summary, extracts the people, companies and ideas mentioned and links them into a knowledge graph, and indexes everything so a search matches what was said, not only the title. Search is hybrid: meaning and keywords are matched together, so an exact phrase and a loosely remembered idea both find the clip.
The same pipeline runs on articles, PDFs and ebooks, so one question can pull from a podcast, a talk and a paper at once. Once you trust that the episode is captured, you can listen without taking notes and look it up when the question comes.
Getting a backlog in
- YouTube playlists and Liked videos: connect your YouTube account and they import and keep syncing.
- A playlist in one go: bulk import takes 10 videos at a time on Free, 50 on Starter and 150 on Pro.
- Channels and podcasts you follow: new uploads and episodes arrive on their own.
- Watch Later: YouTube doesn’t let apps read it, so move those videos into a playlist first.
Your options, honestly
- Paste the transcript into a chat: free and immediate, gone when the chat ends.
- A research notebook such as NotebookLM: good for a fixed set of sources for one project; answers stay inside the notebook. See NotebookLM alternatives for YouTube.
- A read-later or knowledge app with media capture (Recall, Readwise Reader, BrainTube): keeps transcripts of what you save. They differ in whether other AI assistants can search the library and save into it.
What BrainTube can and can’t do here
It keeps the transcript of every YouTube video and podcast episode you save and makes the library searchable from its own chat, where answers cite the source and link a video quote to its timestamp, and from Claude, Cursor and other MCP clients. Your AI can also save what it finds into BrainTube during the same session.
It only knows what you save. It does not read your YouTube watch history or Watch Later list, and it runs as a web app, a Chrome extension, and iPhone and Android apps.
Related: search every video and podcast you save, search everything you’ve watched on YouTube, how to remember what you watch and read and give Claude long-term memory with MCP.