I have tried a lot of PKM systems.
For a long time, my favorite was Roam Research. The graph, the daily notes, the feeling that every thought could connect to another thought — it clicked with me. Roam changed how I thought about notes.
But for agent memory, Obsidian wins.
Not because the app is flashy. I barely open it manually anymore. Pi does most of the work now: reading notes, writing memory, filing research, updating project context, and pulling the right pieces back into the session.
The reason Obsidian works is simpler: it is a local Markdown vault.
Plain files. Real folders. No opaque database I have to trust. No export anxiety. No “where did the platform put my thoughts?” problem. If an agent writes something, I can inspect it. If I want to move it, grep it, sync it, back it up, or edit it with another tool, I can.
That matters more now than it did when I first started using Obsidian.
Memory is becoming a big topic in agent systems. Everyone wants agents that remember preferences, projects, workflows, decisions, and personal context. That is useful, but only if the memory layer is inspectable. Otherwise you are building future behavior on top of hidden guesses.
This is where I feel a little vindicated.
We were thinking about Obsidian as an agent brain before “memory” became the obvious thing to talk about. At the time, it felt like a practical choice. Now it feels like the right architecture.
My setup is not “open Obsidian and write notes.”
It is: keep Obsidian as the brain, let Pi operate on it, and make every durable memory auditable as a Markdown file.
Roam is still special to me. It shaped my taste for connected thinking.
But right now, for agent work, a local .md vault is hard to replace.
Maybe impossible.