Three ways to build a personal knowledge base for AI
1. Local RAG on your own computer
A local app indexes your files and answers questions with a model on your machine. AnythingLLM is the simplest version: a free, MIT-licensed desktop app for macOS, Windows and Linux that runs entirely on your computer, transcribes uploaded audio and video with a local Whisper model, and lets its agents call MCP servers. If you already keep notes in Obsidian, add an AI plugin instead: Smart Connections finds related notes with a local embedding model and no API key; Smart Second Brain adds hybrid search, a topic graph and agents that can read and write notes; QVAC Local AI chats with your vault fully offline, with cited answers. Best for privacy. Trade-off: you run it, and only that app can use the knowledge.
2. A hosted notebook
Upload sources into a notebook and chat with them. NotebookLM (Google now documents it as Gemini Notebook) is free with up to 100 notebooks and 50 sources per notebook, takes PDFs, Docs, web pages, audio files and public YouTube videos with captions, and caps each source at 500,000 words. Best for researching one fixed set of documents. Trade-off: the knowledge lives inside one app and one notebook at a time.
3. An MCP knowledge base every AI client can read
Keep the library in one service that runs an MCP server, and connect Claude, ChatGPT, Cursor and other MCP clients to it. Your knowledge then follows you from one assistant to the next. Notion (pages you write), Readwise (highlights and reading) and Mem (AI notes) work this way. BrainTube does it for what you watch and read: save a YouTube video, podcast, article or PDF and it is transcribed and indexed; any connected MCP client can search it, and your AI can save what it finds into BrainTube. Best when you use more than one AI. Trade-off: the library is in the cloud.
Building memory into your own AI app rather than for yourself? That is a different job: see MCP memory servers compared. Comparing capture apps such as Recall? See BrainTube vs Recall.
How to build one in five steps
- Decide where it lives. Your computer (AnythingLLM, Obsidian, Logseq), one app (NotebookLM), or a server every AI client can reach (BrainTube, Notion, Readwise, Mem).
- Get your sources in as text. An AI can only answer from text: notes, full articles, transcripts. Links and titles are not enough. For videos and podcasts, pick a tool that transcribes (AnythingLLM for uploaded files, NotebookLM for captioned YouTube, BrainTube for YouTube, podcasts and audio).
- Index it for meaning, not just keywords. Local apps and plugins build embeddings on your machine; hosted tools do it for you.
- Connect your AI. Local: chat inside the app. Hosted notebook: chat inside the notebook. MCP: add the server URL to Claude, ChatGPT, Cursor or another MCP client once (for BrainTube: connect guide, give your AI memory over MCP).
- Check you can leave. Obsidian and Logseq are already files on disk. Mem exports Markdown. Readwise (full plan) exports highlights to Notion and Obsidian. BrainTube: export your library as Markdown (Obsidian or Notion format) or Anki from Settings, or the whole corpus through MCP.
What to check before you commit
- Does it keep the full text? Transcripts and full text answer questions; bookmarks don’t.
- Which AIs can read it? Only the app itself (local RAG, notebooks), or any MCP client? If MCP, is it read-only or can your AI write back?
- Where does the data live? On your disk, or in a vendor’s cloud.
- What does the free plan include? Notion AI is a limited trial below Business; Readwise has a 30-day trial; NotebookLM, AnythingLLM Desktop, Obsidian and BrainTube have free plans.
- Can you export it? See step 5.
Related: what an MCP memory server is, one memory across ChatGPT, Claude and Cursor, NotebookLM alternatives for YouTube, BrainTube vs Obsidian, BrainTube vs Notion, best second brain apps in 2026, best AI memory tools for ChatGPT and Claude and how to remember what you watch and read.