My Zettlr Second Brain

Everyone is excited about Karpathy’s second brain technique so I thought I’d share mine.

Those of you closer to 20 than 40 probably don’t know that rusty feeling where recall suddenly slows and some fact or detail you know you know can’t be pulled up mid conversation.

Reading the commonwealth saga must have stuck with me more than I realized. In the commonwealth characters struggle with their brain’s storage capacity when lifetimes extend beyond 200 years for the wealthiest. They all have memory cell implants that continuously record their brains contents and not being able to (or wanting to) hold all those memories in one brain comes up a lot in the plot.

I’ve been personally trying to offload some long term storage using Zettlr since long before personal AI assistants were available. Zettlr is free and open source, has a vim mode, a dark mode, and includes a lot of other features making Zettlekasten easier. Zettlekasten as an organization technique for paper notes long predates personal computers. It should not come as much of a surprise that AI can be super helpful organizing a bunch of markdown notes that are already following a strict and well documented linking convention and format.

Despite being a long-term Zettlr user and advocate I never spent the time necessary linking, managing metadata, all of which are necessary to create a useful index for all this data. I used it as a quick local Notion-like tool where I threw any thoughts with a couple of inconsistent tags and then rarely revisited what I wrote.

My first attempt to improve the situation with AI was to allow Claude Code to do a linking pass creating relationships based on all the content that was already there. This started to slowly improve my linking and tagging, allowing me to more effectively use Zettlr’s interface for searching by tag and understanding content relationships using graphs.

Older models with their 200k context window did alright but since my data was not well linked we filled the window too quickly searching for things. Content was not easily discoverable and filenames like 20250916204826.md don’t exactly provide the tool with a lot of data to run with.

The big unlock for me was getting 1million token context windows and to start using claude code as my primary interface for Zettlr.

I added a CLAUDE.md to my main Zettlr folder that helped claude know where the juicy nuggets and linchpin docs were and started to co-write all my notes with claude’s help. Turns out that fed with the right guidance claude is much more inclined to be uptight about linking than I ever was.

I now have a constantly open claude session or three in my Zettlr folder where I’m getting career advice, asking questions about docs, refining docs, and creating new docs. The more claude works in here the better linked everything gets.

One accidental feature I unlocked was in my attempt to back all this important data up to google drive. Turns out that all those gemini tools in the sidebar of your inbox and docs are perfectly happy to pull in relevant looking much-edited files from anywhere in your drive if given permission and all of the sudden my personal TODO lists and goals started getting mentioned in e-mail and doc summaries.

The end result is far from perfect. I still would like to make an MCP to replace my half-baked Zettlr skill and teach the model exactly how I would like it to interact with this tool and follow the thoroughly documented rules. This is far from a true memory system but surprisingly effective when seeded with a few years worth of human-generated notes.

I may end up indexing, tokenizing, vectoring or otherwise storing this data in a more LLM friendly format, or it may live forever as a bunch of linked markdown docs. For now the model’s abilities to consume a lot of data are scaling faster than my desire to formalize this system.