For me, 2024 was the “year of notes” – if for no other reason that it saw the publication of my book Duly Noted. Its first anniversary is coming up, so this is a good time to reflect on the state of my personal knowledge management system.
TL;DR: it looks much like it did in 2023. Obsidian remains my primary repository. The biggest difference: I’m taking more handwritten notes.
This might be surprising if you’ve read Duly Noted. The book emphasizes hypertext note-taking, which necessitates marked up text. (Particularly, hyperlinks.) It’s difficult to add links or semantic structure to handwritten notes.
So what gives? Have I gone back on what I advocate in the book?
Not at all. Duly Noted also emphasizes that ‘notes’ come in many varieties to suit different cognitive activities and that our systems should evolve to suit our needs. This requires being mindful about how we’re using them in practice.
And when I reflected on how I was taking notes, I realized something important: I pay better attention in meetings and interviews when writing longhand. Yes, I can type faster – but that only leads to verbatim capture, which isn’t what I look for in these scenarios.
Of course, this isn’t surprising. Scientific studies suggest we focus our attention differently when writing by hand. Still, I tried typing up meeting minutes for the sake of cohesiveness. I wanted everything to be captured in Markdown to take advantage of Obsidian’s linking features.
But results invariably fell short. Usually, I’d end with solid notes for the first third or half of a meeting, but petering out toward the end. This was most noticeable when I was an active participant. I can’t effectively type and talk. With handwritten notes, it’s easier for me to scribble even while actively engaging in conversation.
That doesn’t mean I’ve abandoned Obsidian for meeting minutes. Far from it. For me, handwritten doesn’t mean analog. I take handwritten notes almost exclusively on my iPad using Notability. I’ve configured it to back up notes as PDF files to a Dropbox folder. From there, scripts on my Mac import them into Obsidian notes, where I add metadata including the relevant project and the people present.
These ‘metadata wrappers’ give me the bulk of the benefit of using a hypertext system while preserving the ability to focus during the meeting. I can still leverage Obsidian’s powerful linking features to list all notes related to a project or all meetings where a particular person was present. If I need any other links, I just type them up in the ‘wrapper’ note.
Of course, this technique isn’t new. I gave a talk back in April on how to do this. I’ve just leaned into it: I’m now exclusively taking meeting minutes in longhand.
I’ve also experimented with using AI to transcribe these longhand notes. AI is astonishingly good at handwriting OCR, and advances on the API side now make it possible to automate it in the background. Still, I haven’t gone beyond experiments due to privacy concerns. (I’ll likely bake this approach into my workflows once local models get powerful enough.)
My other note-taking practices remain the same. I still capture book notes as Markdown in a dedicated Obsidian vault. I annotate and highlight in Kindle, where I do the bulk of my reading. Kindle notes flow into Obsidian via Readwise, as before.
The biggest change there has been deeper integration of AI. I’m now routinely using GPT to help me better understand my reading. Whenever I finish a book, I reflect on what I learned from it. This reflection now includes a ‘conversation’ with an AI about the book. This practice has greatly increased my ability to process and retain what I’ve read.
Again, not surprising. The last chapter in Duly Noted advocates using AI as an amanuensis; I’m merely leaning into that and exploring its possibilities. But I wrote than chapter in mid-2023, and much has happened since then.
Among the developments: using AI to build knowledge graphs from an unstructured corpus and then using those graphs for RAG interactions with the text. I’ve already experimented with that technique on my blog, podcast, and a client project: all public texts that are already on the internet.
Once local models become powerful enough, I’ll apply that technique to my main Obsidian vault as well. Having the repository consist primarily of plain text will make that much easier. And I can now see a path to transcribing handwritten notes so they can be part of the plain text corpus.
It’s an exciting time for anyone who values learning. Tools for thinking are better than they’ve ever been. I expect they’ll get better still this coming year. I’m looking forward to new developments in 2025 – and sharing what I learn with you.