Original video in Chinese.
What to do now (updated October 14, 2026): The body and Key Takeaway below preserve the original video’s practical account. Plugin versions, installation steps, and subscription allowances mentioned there may no longer apply. Start by separating tasks: keep your long-term original work in local files that you can back up and export. For AI retrieval, organization, and writing back, see the Obsidian AI workflow; to connect a client to your notes, see the Obsidian MCP setup guide. If your main task is reading and checking external material, see how Obsidian and NotebookLM divide the work. “Other people’s material” does not mean “no privacy or copyright risks.” Before uploading, check its source, permissions, confidentiality requirements, and the service’s data-handling rules. Do not upload sensitive material by default. Test retrieval and verification with a small set of material you have permission to use before expanding. This update adds no new plugin tests or subscription promises.
Key Takeaway
- My own content: put it in Obsidian. Local Markdown, free, fully under my control, no risk of being held hostage by the cloud; easy to connect to AI (for example, with MCP plugins), and it feels wonderfully free to use.
- Other people’s content: put it in NotebookLM. No privacy worries, strongest performance + Deep Research automatically expands the search (for example, related materials for YouTube podcasts), and it’s included with a Google One subscription.
- The optimal solution, based on experience: distinguish by source so you don’t regret Evernote/Notion later; after two years of trying, I’ve confirmed this combination is the best. It’s light and efficient, and I recommend trying it and fine-tuning it.
In 2026, if you want to build a personal AI knowledge base, remember these two conclusions of mine:
My own stuff goes into Obsidian.
Other people’s stuff goes into NotebookLM.
I believe this is definitely the optimal solution for a personal knowledge base. After I posted this on X, a lot of people said they felt the same way too.
Think about it: everything in your knowledge base comes from just two sources, your own stuff and other people’s stuff.
If it’s your own stuff, then of course it should be under your own control, or at least you should have a copy of it in your own hands.
Many years ago, there was a note-taking app called Evernote. Back then I was a heavy user and stored everything in it. Later it became a huge hassle, and I regretted it deeply.
So today I don’t trust any note-taking app that doesn’t keep a local copy. Take Notion, for example: no matter how polished it is or how popular it gets, I won’t use it.
What’s more, if there’s no local copy and the content isn’t in Markdown format, then I can’t freely use AI tools to work with those notes.
In other words, not only is all my stuff on your servers, but if I want to use AI later, I still have to buy your service. Isn’t that basically walking right up and letting people hold you hostage?
Look at Obsidian—it doesn’t have these problems.
Obsidian itself is completely free. All your notes live on your computer in Markdown format.
Connecting AI to Obsidian is simple and very flexible. I recommend two methods.
First, use the plugin version of Claude Code directly.
You heard that right: Claude Code has a plugin version called Claudian. All you need to do is download these three files from the GitHub repository and put them in a folder. Then open the folder where your Obsidian plugins are stored and drag the Claudian folder in. That completes the installation.
Next comes configuration. If, like me, you’re connecting to a third-party model provider, then you need to fill in the environment variables in settings. For example, I’m using Kimi here.
If you’re on macOS Homebrew, the first launch may throw a code 127 error. I ran into that myself and only solved it after asking Grok. In fact, there are just two things you need to do: fill in the full path and forcibly inject PATH in the environment variables. Then this problem is 100% solved.
Once you’ve got Claudian working, you can use Claude Code directly inside Obsidian. Especially after creating custom commands, you can have AI automatically handle a lot of repetitive work.
For example, I have AI summarize and distill the articles I save each week, compile them into a report, and save it to a designated folder.
Second, use Claude Code or OpenCode to open the note repository.
If you choose this method, Obsidian moves into a supporting role. On the front end, we interact through Claude Code or OpenCode.
I used to use Claude Code all the time, but over the past few days I’ve started switching to OpenCode.
You can think of OpenCode as an open-source alternative to Claude Code. It supports all kinds of models, including locally run ones. You can even log in directly with your subscription. Its TUI also looks better than Claude Code’s.
After you install the oh-my-opencode plugin, you can really feel the power of multi-agent collaboration. And this plugin fully supports Claude Code’s configuration, including custom commands, Skills, MCP, and more, so there’s no need to migrate.
Since the agents in this plugin are designed for programming, I also made some adjustments to their system prompts. I won’t go into that here; I’ll share it in Knowledge Planet.
So whenever I sit down, I use a setup like this:
The lower screen is the iPad, used as an extended display showing OpenCode. I use it to open my note repository.
The upper screen is Obsidian, mainly for editing and viewing documents. When I occasionally need Claude Code, I open the plugin.
My own stuff goes into Obsidian, then I connect it to AI. Try it the way I described, and you’ll know how great it feels.
As for other people’s stuff, then there’s no need to worry about privacy. Since they’re all PDFs I got from who-knows-where, why worry so much? Just take my advice and use NotebookLM, and upload all the documents there.
NotebookLM is extremely powerful, and there’s a good chance it’s more useful than other tools on the market. And once you subscribe to Google One, you get a higher quota, which is basically free value.
I already talked about how to use NotebookLM in detail in a previous video. I’ll just mention one thing here: its Deep Research is really excellent!
For example, I import links to YouTube podcasts. When I discover certain points or topics that really interest me, I have NotebookLM expand the search. It finds a lot of related content and adds it all in together. That way, I save myself the trouble of having to go look for materials separately.
All of the above is my summary from experience. Over the past two years, I’ve tried all kinds of personal AI knowledge base solutions. Some weren’t powerful enough; some were too heavy and too troublesome. In the end, only these two setups are the most suitable for personal use. You can try them first and then make adjustments based on your own needs.
OK, that’s it for this episode. If you want to learn about AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!