Obsidian as memory for agents

Most knowledge tools lock their contents behind an interface. Obsidian does the opposite: a vault is an ordinary folder on your disk, every note a Markdown file. That makes it the obvious memory for agents – they can read files, and nothing more is needed.

A structure that works for both

You do not need an elaborate system. Three things are enough:

Knowledge/
  START.md        ← map: which folder holds what
  RULES.md        ← how reading and writing work here
  Projects/
    Store/OVERVIEW.md
    Tax/OVERVIEW.md
  Sources/        ← originals, minutes, archive

START.md is the map. RULES.md addresses everyone who writes – agents included. Every project has an overview that points to the details instead of repeating them. What the individual notes should look like was the subject of the previous station.

How the agent gets to the vault

File access. A coding agent or a desktop assistant you grant the folder to reads and writes the files directly. In its rules file – next station – you write a single sentence: “Before substantive work, read RULES.md in the vault, then only the overview of the project concerned.”

The Obsidian CLI. Since version 1.12 Obsidian ships an official command line, enabled in the settings. It has commands for files, search, tasks and Bases and takes the target vault as a parameter. An agent with terminal access can use it to search and to create notes from templates without guessing the folder structure. Obsidian CLI

Bases. A base is a table view over notes, saved as a .base file in YAML. The data itself stays in the Markdown files and their properties. For an agent that means: it maintains properties such as status or updated in the notes, and your overview updates from them. Bases

Five rules for two authors

As soon as an agent writes, you have a co-author who is fast and forgets nothing – mistakes included. So RULES.md should say:

1. Read first, then change. Fetch the current content, add to it precisely, check back afterwards.
2. Every new statement with date and source.
3. Delete nothing silently. What is superseded is marked as superseded or moved to the archive.
4. No credentials. Only a pointer to where they are kept.
5. No second memory. The agent creates no memo files of its own elsewhere. If the vault cannot be reached, it says so instead of improvising.

Plus one piece of insurance: put the vault under version control or make sure there is a backup with history. Then every change an agent makes is visible and reversible.

Stored locally does not mean processed locally

The vault sits on your machine. The model that reads it usually does not. Every note an agent opens goes as text to the provider of its model. So separate what an agent may see from what stays private – most simply with two vaults or one folder you never share.

If you want to keep the processing in-house too, you need a local model. That is the learning path Local AI.