AI
← HomePersonal Agent Stack
One memory layer for every agent I run, so I stop re-explaining my life to each one.
The problem
Every AI tool I used started from zero. I kept re-explaining my work, my projects, and my deadlines to each one. The context existed, it just lived in my head and in a dozen apps.
The build
One memory layer that every agent reads from and writes to.
- Shared memory. An Obsidian knowledge base, exposed through a local MCP server, acts as memory for Claude, Claude Code, and Hermes alike.
- Agents. Hermes runs in Docker. Claude Code handles engineering work. Both work from the same notes.
- Schedules. Agents run on cron for recurring work, like checking deadlines and preparing drafts.
- Capture. Messy quick notes from my phone get triaged into tasks and deadlines in Notion. This part is still in progress.
Guardrails
- Agents draft. I send. Nothing goes out under my name without me.
- Anything irreversible needs my approval first.
- Before I approve a command, the agent explains what it does in plain language.
- No secrets go into the shared memory.