Building an AI Second Brain in Obsidian: Our Setup
Updated · May 13, 2026
Most second brain systems die the same death: beautifully organized, almost never opened. You spend an hour tagging a note and then forget it exists. We rebuilt our Obsidian workflow from scratch eight weeks ago with one goal — make the AI do the connecting work, so we’d actually want to open the vault. Here’s what we shipped and what we’d do differently.
The rules of the experiment
One blank vault. A $30/month ceiling. Every tool had to work from inside Obsidian — no jumping to a browser tab, no copy-pasting between apps. If the workflow pulled us out of the editor, the tool failed.
We ruled out Notion AI before we started. Not because it’s bad — Notion’s AI summaries are genuinely useful — but because we wanted Markdown files on our own machine. Obsidian stores everything as plain .md files. That single decision shaped every other choice in the stack.
What we needed the system to do: capture information without friction, retrieve notes without remembering where they were filed, surface connections across the vault automatically, and generate synthesis on demand. Simple list. Surprisingly hard to actually build.
Building the capture layer first
We almost built an elaborate folder hierarchy and tagging taxonomy. Classic Obsidian trap. We killed it on day three when we realized we were spending more time organizing than writing.
Instead, the capture layer runs on two things. Readwise handles everything we read: Kindle highlights sync automatically, and the Readwise Reader browser extension clips web articles in one click. The Obsidian plugin drops everything into a “/Readwise” folder as clean Markdown — we’ve never touched that folder manually. At $7.99/month (or $99/year), it’s the easiest decision in the stack.
For original notes — meetings, research, half-formed ideas — we use daily notes with a Templater template that prompts for one or two context tags. Everything lands in a flat “/Notes” folder with a date prefix. No subfolders. Finding the note is the AI’s job, not the filing system’s.
Which AI plugins are worth installing?
Two plugins. Everything else is overhead. We tested seven total. The other five either duplicated functionality, broke our writing flow, or required so much configuration that the setup became the project. Here’s what each of the two survivors actually does and why we kept them.
Smart Connections is the core of the system. Install it, connect your OpenAI API key, and it builds a semantic index of your entire vault. You can then ask “what have I written about decision-making under pressure?” and get relevant notes back — not by keyword match, but by meaning. During our testing it surfaced a connection between a book highlight and a meeting note written six months apart that we had completely forgotten. That kind of link is the whole point of a second brain.
The plugin has a free tier that runs embeddings locally using a small model via Ollama. The quality gap against OpenAI’s text-embedding-3-small model is noticeable but the privacy tradeoff is real. For a vault of around 800 notes, we pay about $3–5/month in API costs. Well within budget.
Obsidian Copilot handles the chat interface. Select a note or set of notes, open the Copilot panel, and ask questions in plain English: “Summarize what I know about this client” or “What did I decide the last time I worked through this problem?” It routes to whichever LLM backend you configure — we use Claude Sonnet via the Anthropic API for most queries. In our experience it stays tighter and hallucinates less on domain-specific notes than GPT-4o does.
We also tried Text Generator for drafting notes from bullet points. Cut it after a week. We already have ChatGPT and Claude for that work — having it inside Obsidian added friction without adding capability.
What does the day-to-day workflow actually look like?
Our routine is 15 minutes every Friday. We ask Smart Connections what was captured this week, then use Copilot to find connections between the new notes and the older vault. About 60% of AI-suggested connections turn out to be genuinely useful — enough to surface things we’d have missed, not so many that triage becomes a job.
Dataview (free plugin) generates automatic queries across the vault: all notes tagged “#decision” from the last 30 days, all open questions without a resolution marker, all book highlights that haven’t been processed into permanent notes yet. These lists surface the knowledge we’ve been avoiding. Pair a Dataview query with a Copilot question and you get something close to a thinking partner that knows everything you’ve written.
One more piece worth mentioning: a Zapier automation that pipes starred Slack messages into a Markdown file that drops into Obsidian each morning. Ten minutes to set up on Zapier’s free tier. Saved hours of manual copy-paste over the past two months.
What we’d change next time
The biggest mistake was migrating an existing knowledge base. We imported years of old notes from Notion and Evernote. The noisy vault confused Smart Connections — it kept surfacing irrelevant old content alongside fresh notes. If we rebuilt tomorrow, we’d start fresh and import selectively, only notes we’ve actively referenced in the past year.
We also underestimated Obsidian Sync at $8/month. We tried iCloud sync first. It caused vault corruption twice in three weeks. Either pay for the official Obsidian Sync or set up git-based sync. Don’t improvise this part — losing your knowledge base to a sync conflict is exactly the kind of failure that makes people abandon the whole system.
Local LLMs via Ollama are worth testing if privacy is a genuine concern — client notes, sensitive research, anything you’d rather not send to a cloud API. On an M3 MacBook, the quality for general knowledge work still falls noticeably short of GPT-4o-mini. The gap is narrowing but it’s not closed.
The final stack
- Obsidian — vault and editor: free (personal use)
- Obsidian Sync — cross-device sync: $8/month
- Smart Connections plugin — semantic search across vault: free plugin + ~$4/month OpenAI API
- Obsidian Copilot plugin — in-vault AI chat: free plugin + ~$8/month Anthropic API
- Readwise — book and web highlight capture: $7.99/month
- Zapier — Slack-to-Obsidian automation: free tier
- Total: ~$28/month
Frequently asked questions
Do I need a paid Obsidian plan to use AI plugins?
No. The Obsidian app is free for personal use. Smart Connections and Obsidian Copilot are free community plugins — you pay for API access to the LLM you choose, not for Obsidian itself. Obsidian Sync is optional but strongly recommended over third-party cloud sync services.
How many notes do I need before Smart Connections becomes useful?
Semantic search starts earning its keep around 150–200 notes. Below that threshold you can find things manually without much trouble. Above 200, the AI’s ability to surface forgotten connections becomes genuinely valuable — building that density takes two to three months of consistent capture.
Is there a privacy risk to indexing my vault with OpenAI?
Yes, with caveats. Note content is sent to OpenAI’s embedding API when Smart Connections builds its index. For most personal knowledge work this is acceptable. For sensitive client or legal notes, switch Smart Connections to local embeddings via Ollama — quality drops, but nothing leaves your machine.
Eight weeks in, the vault has become something we actively want to open rather than feel obligated to maintain. The AI layer doesn’t replace good note-taking habits — it just makes the payoff for those habits dramatically more visible, faster. Start with Smart Connections and one consistent capture routine. Everything else can wait until you have 200 notes worth indexing.
This article contains affiliate links. If you subscribe through one, we may earn a commission at no extra cost to you. It never changes what we recommend — we only link to tools we actually use. Full disclosure.
