We gave our agents a shared company memory in plain markdown. What we learned after 4 months
We're a small AI studio in Germany. Since June our whole company runs on one Obsidian vault that we and our Claude agents both read and write. It's at 600+ notes now, plain markdown in git with a small local search index on top.
A few things that turned out to matter more than we expected.
Agents can write but they don't get to decide. Every change they make shows up as a git diff and one of us approves it before it counts. Without that rule we wouldn't trust anything in there.
Numbers need a source. When our morning briefing agent reports a number it has to say where it came from, and if it can't find a source it writes "not verifiable" instead of guessing.
Raw material never gets edited. Call transcripts, PDFs and exports go into one folder and stay as they are. The agents write short notes from them somewhere else and link back, so you can always check what the original said.
It gets messy faster than you'd think. Our last check found 49 broken links and 59 notes missing basic metadata. A script finds it, an agent suggests fixes, we approve.
Decisions get their own note with what we decided, why, and what would make us change our mind. We have 21 of those now and they save a lot of "wait, why did we do this" conversations.
How are you handling shared memory for agents across a team? Do you let them write directly or do you gate it too?
We also made a free example version of this setup with a fictional company in it (MIT). Happy to share it if anyone wants to look.