Anyone using AI agents with their Zettelkasten? Thoughts on graph access?

gimalay

I've been experimenting with AI coding agents (Claude Code, Cursor, etc.) and started wondering: could my Zettelkasten serve as context for these tools?

The idea came from frustration. When I ask an agent for help, it doesn't know my projects, my conventions, my past decisions. I end up copy-pasting context manually. But I already have a system for organizing that knowledge — my Zettelkasten.

So I started thinking: what if the same links I create for my own navigation could help an agent understand what's related to what? Instead of dumping raw files, the agent could follow my structure.

I'm using a tool that I built. It's called IWE and it lets agents query my notes:

iwe find "cognitive biases"

iwe retrieve -k psychology/confirmation-bias --depth 2

The find command searches notes. The retrieve command follows links to expand context — if "confirmation bias" links to "motivated reasoning", the agent gets both in a structured response.

What I find interesting is that the connections are mine. When the agent follows a link, it's following a decision I made, not an automated association. The same structure that helps me think becomes the structure the agent navigates.

I'm still figuring out if this is genuinely useful or just a clever idea that doesn't pan out in practice. Some questions I'm sitting with:

• Are others here using AI tools alongside their Zettelkasten?
• Does giving an agent access to your graph feel useful, or does it feel like mixing concerns?
• Is there value in the agent "understanding" your personal structure, or is raw search good enough?

Would love to hear how others think about this intersection — or if you've deliberately kept AI tools separate from your knowledge work. Both perspectives are interesting to me.

GitHub if you are interested: https://github.com/iwe-org/iwe

ctietze

You're on the right track as far as I can tell: Progressive summarization makes books available to our human brains through excerpts and summaries in stages. Each stage "compacts context" in a way. LLMs work well in the opposite direction, assuming infinite, unbound, pre-digested knowledge but no clue about what's relevant -- then work from summaries to details as needed.

To make an agent use a/the test-driven development methodology, you don't need to copy-paste books about it into the context. You only tell it to use red/green tdd, and that'll put it right on track.

Eventually, you'll want to end up with look at my zk and the LLM will then know where to look, how to orient itself, and get to the details.

Structure notes are for both the human thinker and the agent to find their way