The concept of the "second brain"/"knowledge base" is one of my absolute favorite nerd snipes of all time. As a writer and blogger, I love everything about the idea of attempting to hoard every single piece of knowledge I accumulate into a structured, tagged, and metadatafied text mound repository. The memory talks today at the AI Engineer World's Fair, along with some of the online track talks, led me to the realization that the human knowledge base and the AI concept of an AutoWiki are a match made in heaven.
Let's start from the beginning so we don't lose anybody: a knowledge base, or a "second brain" if you do it a certain way, is a collection of data in various states from raw, incoming knowledge to processed, synthesized, integrated learnings. Oftentimes these data bits are stored as plain Markdown files (here your AI antennae should be perking up). If you're extra cool, you use an app like Obsidian or Neovim to operate on these files. One popular method is to set up an incoming inbox where you can jot down new information or idea fragments in seconds so that you don't lose anything by thinking, "Oh cool, I'll make a note of that later." Then, when you get some time to process things, you go through your inbox, fill in blanks, do more research, and eventually, you end up with a more processed, more personalized piece of knowledge that is all linked up to lots of other pieces of processed knowledge in your knowledge base. You may also generate even more ideas or thoughts for the inbox during this process. Then, when you go to create or learn something, you already have all of these related, processed knowledge chunks to draw from. If this seems like a cult mindset that interests you, look up the word Zettelkasten.
You know what else processes and synthesizes Markdown files? That's right. AI. It's, like, the main thing it does. Do you know one of the hardest problems to solve in a manually-managed knowledge base? (The more librarian-minded amongst us might also argue that this is one of the most fun problems.) Finding and combining stuff. You have all these linked bits, but then you're forced to a) click through link after link until you're thoroughly lost, b) write up your own summary notes that you then link to the raw notes, growing your web, c) see if you can get lucky with tag or string searches, or d) cry.
One of the online talks, Turn 10,994 Notes into Memory, by Paul Iusztin and Luis-François Bouchard proposes one approach: give your agent access to your notes to answer questions for you. And to answer the question of that being way too much context to consume for each question, they propose giving your agent a little scratch area to set up its own memory wiki. It's possible you don't want it in your main knowledge base updating and editing files unattended. But that doesn't mean it can't have its own knowledge/memory base. And the extra neat thing they propose is a semantic index file so that the agent can know whether to look at its own docs, or dive back into your full knowledge base for more information. You would have to do some work and schedule some staleness/wiki refresh runs before deep AI work sessions, but there is a synergy to these two tools that I hadn't considered before and I will be experimenting with my own knowledge base.