I can search 185+ notes, every article I have ever captured, every YouTube video I have bookmarked, every past conversation I have had with Claude — all from one place. Semantic search, not keyword matching. I type a question in natural language and get back the relevant thing I wrote, read, or discussed six months ago.
This is what I built and how it works.
The Problem
I keep an Obsidian vault with notes on everything: investment theses, technical architecture decisions, product strategy, research, personal preferences. The vault grew to 185+ files across every project I work on. At some point, a folder of markdown files stops scaling. You forget what you know. You re-research things. You lose context between sessions.
The search built into note-taking apps is keyword-based. If you do not remember the exact words you used, you will not find what you are looking for. I needed something that understood meaning, not just text.
Vault Sync and Embeddings
A sync process reads every markdown file from the vault, extracts the content and metadata, and embeds it using a multimodal embedding model. Text and images live in the same vector space, so semantic search works across content types. The embedded notes are stored alongside their vectors in the database, and retrieval uses built-in vector search — no separate vector database, no external service to manage.
When I search "what was my thesis on gold exposure," it finds the note even if I never used the word "thesis" in it. That is the difference between keyword search and semantic search, and it is the reason this system exists.
Source Capture
This is the part I use most. I built a capture pipeline that takes any URL — article, YouTube video, tweet, Reddit post, Instagram reel, TikTok — and processes it into the knowledge base. The pipeline extracts the content, summarises it with AI, classifies it into the right vault section, generates tags, embeds everything, and stores it.
One URL in, fully indexed knowledge out.
I also built a mobile app so I can capture sources on the go. See an interesting article on my phone, share it to the app, and it is in the knowledge base by the time I sit down at my desk. The capture habit is the whole game — the system is only as good as what you feed it.
The MCP Server
This is what ties the knowledge base into my actual workflow. An MCP server exposes tools to Claude: vault search, source capture, note creation, session history, and more. When I am in a Claude conversation and need to reference a decision I made three months ago, Claude searches the vault semantically rather than me digging through folders.
The mental model: Claude has access to everything I have ever written down. It does not have perfect memory, but it has searchable memory — which is better than what most people have of their own notes.
The Council
The most interesting piece. The council is a multi-model RAG system where multiple LLMs answer the same question against my knowledge base, and then a reviewer model synthesises the best answer.
Seven models across four providers, each answering independently against the same retrieved context. The reviewer picks the strongest reasoning from each response and combines it into a final answer. It is like having a research team that always disagrees productively — you get the benefits of diverse model perspectives without having to prompt each one individually.
Why This Exists
The honest answer: I was tired of context loss. Every new Claude conversation starts from zero. Every project has decisions buried in old chat logs. The vault is the persistent layer, and the MCP server makes it accessible without copy-pasting context manually.
The less obvious benefit: it changed how I make decisions. When you can ask "what did I decide about X last time and why?" and get an actual answer grounded in your own notes, you stop relitigating the same questions. The knowledge compounds. Every note you write, every article you capture, every conversation you have — it all becomes searchable context for the next decision.
That is what a knowledge OS actually is. Not a note-taking app with AI bolted on. A system where everything you know is always available, always searchable, and always connected to the tools you use to think.