zen-brain, a personal second brain

Context

My documents, emails and tasks were spread across several drives, mailboxes and apps. The idea: centralise everything, reach it from my PC as well as my phone, then plug in an AI agent to summarise emails, track tasks and automate whatever can be automated.

The project contains no documents: only the rules, routines and tools that frame the agent.

The method

Trusting an agent is something you build with a narrow scope and actions you can undo. So every step follows the same order:

• Audit read-only before acting: a full inventory of what exists, changing nothing.
• Simulate the filing in a “before → after” table that I approve before anything moves.
• Copy, never move: copies are verified with SHA-256 checksums, and nothing is deleted.
• Write routines in Markdown: “file my inbox”, “find my document”. The agent follows a playbook with guardrails (one file read at a time, and when in doubt the file stays where it is).
• Limit the tools: a small local server gives access to the Drive, and only the tools listed in the routine are allowed.
• Local OCR (Windows) to read scans.
• Local index to find a document without rereading everything.

Sensitive information follows separate rules, and secrets are never read or copied.

Where it stands

It’s an ongoing experiment: the audit, the plan, the verified copy and the upload are done, and the testing phase has started.

What I take from it

Handing an AI agent your own data forces you to think like you would for a critical system: a narrow scope, written reports, reversible actions, and a human approving every irreversible action.