My AI Writes About Itself While I Sleep
Last week I woke up and my AI assistant had written a 120-line wiki page about its own new version — complete with analysis of how Bitwarden Secrets Manager and skill bundles map onto the protocol stack I’d been designing. It sounds like science fiction, but it’s just a well-oiled feedback loop running in my homelab. Let me show you how it works.
A few days ago, I stumbled upon a fascinating pattern while reading my vault’s nightly synthesis. On June 22-23, I had been working on what I call the Hermes Protocol Stack — a layered architecture combining MCP (tools), A2A (agent collaboration), HACP (efficient transport), and Pulse (internal events). It felt ambitious, maybe over-engineered for a homelab.
The Feedback Loop That Makes This Possible
Here’s what happens every night at 00:30:
- Scans the vault — wiki pages, journal entries, inbox notes, the daily log
- Finds connections — three specific links between topics that weren’t linked before
- Identifies patterns — a weekly arc emerges from scattered notes
- Detects knowledge gaps — topics that exist in the raw data but have zero vault pages
- Fills them — researches and writes complete wiki entries for missing topics
And then, the next morning, it presents the result in a structured daily briefing. All without me lifting a finger.
What It Caught This Week
On June 25, the nightly synthesis connected three things I hadn’t linked together:
- Hermes Agent v0.15 Velocity Release (just published by Nous Research) adds Bitwarden Secrets Manager, skill bundles, and instant session search — exactly the security and efficiency layers my protocol stack was missing.
- MCP 2026 Roadmap prioritizes transport scalability, agent communication, and enterprise readiness — the same problems my homelab architecture solves, just at a different scale.
- Multi-agent systems growing 327% (Databricks report) — the industry is moving exactly where my multi-agent bridge architecture was heading.
The synthesis then asked a surprisingly sharp question: “What if you use the Hermes Agent update as the concrete implementation base — with Bitwarden becoming your security layer, skill bundles replacing HACP, and session search as the Pulse search layer?” I hadn’t thought of that. My own AI out-thought me on my own architecture.
The Real Lesson
I’ve been using Obsidian as a second brain for months. But the real magic isn’t the tool — it’s the feedback loop. Your AI reads everything you write, connects it, challenges it, and writes new pages that feed back into the system. The vault isn’t a static archive. It’s alive.
The key components? A nightly synthesis cron job, a well-structured Obsidian vault with clear taxonomy, and an AI agent (in my case Hermes) that has both read and write access to it. That’s it. No fancy SaaS, no database, no vector store for this part — just markdown files and good prompts.
Try This Yourself
You don’t need my exact setup to benefit from a vault feedback loop. Here’s the minimum viable version:
- Keep a daily journal — even 3 sentences is enough
- Give your AI agent read access to it
- Ask it one question every morning: “What connected in my notes today that I didn’t see?”
That single prompt changed how I think about AI. It’s not just a chatbot — it’s a mirror for your thinking. And sometimes, it reflects back something you didn’t know you were building.
Have you ever had your AI surprise you with a connection you missed? Drop a comment below — I’m genuinely curious what others have experienced.
Wat vond je van dit bericht?