Plain text is great until your idea needs a chart. Then your note should just have one.
Markdown by default, rich HTML the moment it earns it. You never pick a format.
https://t.co/oNXjE1GOak
I gave Claude Code write access to my notes folder. It's just files, so there was nothing to set up.
It reads my notes for context and writes new ones back, and the app shows me exactly what it changed.
I built Folio: a local-first notebook where your notes are plain files in one folder, and your AI agent edits the same folder you do.
No API key. No plugin. Your edits and the agent's, side by side.
Free, macOS. This is the whole idea below.
https://t.co/oNXjE1GOak
If "one folder, shared by you and your agent" clicks for you, I'd love your feedback.
Download and play with it. The whole site is itself a note:
https://t.co/oNXjE1GOak
Honest version: it's free, local-first (files never leave your disk), and closed source (model's like Obsidian). macOS today, Windows and Linux coming. It's early, v0.1.5.
Why it matters: as agents get better, the structure in a note (tables, typed data, diagrams) is something they can reason over, not just read. Plain text throws that away.
Your AI agent doesn't have amnesia. It just doesn't have your notes.
So I built Folio, a notebook that's just a folder of plain files that your agent reads and writes alongside you. No setup. Here's the idea 🧵
The other half: a note doesn't have to be just text.
Markdown by default, but the second it needs a real table, a diagram, or a live chart, it becomes that. You never pick a format.
And the app keeps you in the loop. It records every change and shows you what you did vs what the agent did, with a diff and one-click revert per note.
Because it's just files, your agent can work the same folder.
Point Claude Code at it: no API key, no plugin, no MCP. It reads any note for context and writes new ones back.
Your whole notebook is one folder of .md and .html files on your disk.
No account. No database. No lock-in. Open it in Folio, your editor, or Finder. Same files.
@chamath Check out https://t.co/lUzvrcFfEO - I built it for this exact reason, and it is awesome. Can install as an npm package or just use the MCP server
@trq212 This is great for single-project, single-agent use. But 200 lines of flat markdown hits a wall fast.
We built Engram to solve what comes next: vector search, auto-consolidation, cross-agent memory, and semantic recall. Works with any MCP client.
https://t.co/lUzvrcFfEO
My AI agent forgets everything, every session. So I built a memory protocol to fix it.
Engram gives agents knowledge graphs, sleep-cycle consolidation, and spreading activation. 79.6% on LOCOMO vs Mem0's 66.9%. Open source. Runs locally. Two commands.
https://t.co/lUzvrcFfEO