How I Turned 10,000 Notes Into an AI Research OS for Agents

Good morning, everyone!

I’m preparing a new talk for AI Engineer World’s Fair in San Francisco with my good friend Paul Iusztin, and I wanted to share an early rehearsal version with all supporters here on the paid subscription before we give it in person.

The talk is about turning a large personal knowledge base into usable context for AI agents.

At some point, simply pasting notes, documents, videos, repositories, and saved links into Codex or Claude Code stops working. Not because the models are bad, but because the raw material quickly exceeds the context window. To make it useful, you need memory and context engineering: organizing sources, indexing what matters, preserving provenance, and loading only the right context at the right time.

Paul and I walk through a system that turns an Obsidian vault of roughly 10,000 notes, documents, videos, and repositories into a file-based research wiki agents can query, update, and reuse.

The main idea is:

One-shot agents use context. A research OS builds memory.

The full implementation is open source here:
https://github.com/iusztinpaul/ai-research-os-workshop

It is also closely related to our Agent Engineering course, which I recommend taking to develop a good AI engineering toolset:
https://academy.towardsai.net/courses/agent-engineering

For paid subscribers, I’m sharing the private rehearsal version…

For everyone else, thanks a lot for your ongoing support! I’ll come back with a new video quite soon ;)