My second brain, and why I still need my first

Ever since Andrej Karpathy published his idea of an LLM-maintained knowledge base built in Obsidian, these systems have become one of the defining elements of the tech, LLM and AI space. Everyone and their grandmother seems to be building themselves a 'second brain'.

I am, of course, included in that 'everyone'. In my defence, my second brain in Obsidian was under construction long before LLMs and agentic AI took off, since the platform has been my note-taking application of choice for over four years now. What started as a desire to have data sovereignty over my own writings and ideas has turned out to be one of the best service choices I have made over my academic career.

Two things people forget

On the topic of second brains, I have noticed two key requirements that people seem to forget or disregard.

1. There needs to be a genuine use case for your second brain. Many systems seem to exist just to aggregate resources and websites automatically and collect them as markdown files.

You have not built a second brain. You have built google.com.

If you end up in a situation where you no longer know the scope and breadth of the information stored in your second brain, is that brain really yours? Therefore, only build a knowledge base for something you genuinely do, whether that be an academic degree, a position at work, or a specific long-term project you are working on.

2. When building a second brain, do not neglect your first. The second brain should be a tool to help you develop your first brain, the one in your head. If the second brain no longer supports your first brain, your first brain has become the second one.

So, my own studying and note-taking system, which I promised to go through in my earlier post: what led me to it, how it works, and how you can apply similar techniques to get better academic results, study more efficiently, and end up with more knowledge in your first brain.

Why take notes at all

Note-taking is a key mediator of learning. For decades, schools preached the VARK learning styles. In truth, you are not a visual learner. You are just smart and lazy, which has let you sit in class and still ace the exam. For many students this method stops working quite suddenly once the topics get more complicated.

Learning is just the brain forming connections, and the way you get the brain to form connections is by straining it and forcing it to do work. Note-taking is a surprisingly good way of achieving this, given that it is done right: not mindlessly copying lecture slides, but actually processing the information and forming new ideas and connections.

Statement 1: notes need to last

I struggle to find motivation to take notes if I feel that I am not creating anything persistent. This happens if I take notes just for the sake of paying more attention or processing the information more actively. Persistent notes also let me recap lectures and topics, prepare better for exams, and quickly look up tidbits I have forgotten, e.g. during exercise sessions.

So, take notes and save them, right? But there are so many ways to do this. Paper notes are not an option, as organising them would be a nightmare. I despise anything Microsoft Office related and prefer to avoid those apps. Proprietary note-taking apps spook me, since the service could just shut down and leave my notes in a mess.

Obsidian felt like the logical choice: it is very customisable, has plenty of ways to organise notes, and the notes are plain markdown files, so I can take them elsewhere if I one day find the service unappealing. But how to organise the notes? For a long time, before Obsidian, I organised my digital lecture notes by course.

Statement 2: folders do not work

Folder-based note structures have one key limitation: what do you do when a note or topic falls under two categories? For example, when I have a note relevant to two courses, which course folder do I place it in? Or both?

The solution I found was to just forget the courses. I take lecture notes by topic and sort the topics into courses with tags. This way a single note can be accessed from the point of view of both courses.

Statement 3: maintenance is tedious

Maintaining the knowledge base is tedious. How can I get the most out of it for learning with the least work necessary? And how can the system adapt to different courses with different difficulties, some of which need more recap than others?

The answer to this is my current system.

The system

My current system has three parts:

1. A file base with a collection of wikilinked markdown notes following a set of formatting conventions, course folders with submissions and attachments, and a bibliography folder containing a literature note and the attached PDF for every item in my Zotero library.
2. A Zotero collection of scientific literature.
3. A Hermes agent orchestrating the maintenance of all this.

Everything lives on my home server, from where the notes are synced to my laptop, tablet and so on. Together, these parts let me get school work done and study for exams efficiently. I chose Hermes for the orchestration to give myself a way out when Anthropic and fellas inevitably hike their prices, as it lets me switch to a local model with ease. Here are some examples of what I can do with it.

Ordinary coursework based on lectures

My lecture notes flow from raw notes into topic notes in the vault through a single custom ingestion skill run as a daily cron job. The agent processes the information in the raw note, including citations and figures, into topic notes, and every topic note gets its course as a field in the frontmatter. This lets a topic note belong to several courses, while each course can still collect its notes into a database for a quick view of what the course contains.

In general, I do one of three things depending on the lecture style:

1. If the lecture slides are information-heavy, I have the agent prepare a raw note from the slides, which I then fill in during the lecture with whatever extra the professor says. I simply cannot write fast enough to copy everything from the slides.
2. If the slides are thin on text and the professor conveys most of the information by speaking, I write the raw note myself during the lecture.
3. If the lecture content is complicated and I have the time, I write the notes by hand and later turn them into a digital raw note. This adds a second round of recap, which really helps with tough courses, but it obviously takes more time.

Recapping for an exam

For exam recap, I usually read through all the contents of a course in chronological order, after which I test my knowledge by having the agent quiz me on the topics in my notes. Sometimes I generate mock exams to work on, and sometimes I compose collections of notes into single pages. This all depends on the nature of the course.

Writing assignments, essays and a thesis

I use custom research skills to search PubMed and OpenAlex for research papers relevant to a query. I can also give the agent .ris exports of Scopus searches. The result is a sort of research assistant, which can sort through thousands of papers to filter out irrelevant publications and put together cited answers to research questions. This lets me quickly see which sources offer which pieces of information and write based on those. It has come in extremely handy while writing my thesis, saving dozens of hours that would otherwise have been spent just getting to the relevant publications.

Does it pass its own test?

Coming back to the two requirements from the start of this post, I think my system passes both, though the second one takes some effort to keep up. The use case is clear: everything in the vault is there because of a course, a submission or my thesis, and I know roughly what is in it because I have written most of it myself. The literature notes are the exception, but even those are tied to papers I have chosen to keep in my Zotero library.

The first brain is the harder part. When an agent does the maintenance, it is very easy to let it do the thinking as well, and a topic note the agent has written from my raw note is not something I have learned just because it exists. This is why I still write the raw notes myself, by hand when the course is tough, and why exam recap ends with the agent quizzing me rather than summarising things for me. When writing, the agent finds and sorts the sources, but the text itself is mine.

If you want to try something similar, you do not need a home server or an AI agent to start with. Pick one course, take your notes by topic instead of by lecture, link them together, and save them somewhere you actually own. The automation can come later, once you know what you need it for. I will most likely keep changing parts of my own system as courses and tools change, but notes I write myself, in a format I own, have been the core of it for four years now.