Your Obsidian Vault Hires an AI Intern

What it does

This is essentially a structured Obsidian vault template and orchestration layer. You capture ideas in an inbox, track projects, and take daily notes in plain markdown. An external CLI agent called Hermes reads those files, runs skills like inbox sorting or stale-project detection, and writes results back into your vault—including a running lessons.md file meant to make the system “self-improving” over time.

The interesting bit

The memory layer is stubbornly low-tech: no vector database, no embeddings, just markdown files on disk. The agent appends its operational lessons to a permanent file inside the vault, creating a feedback loop where the “improvement” is durable, inspectable, and survives even if the AI model behind it changes.

Key highlights

• Uses standard Obsidian folders (inbox/, projects/, daily/, etc.) as both memory and task queue
• Ships with five built-in skills: inbox processor, daily brief, project health, weekly review, and a “learn” skill that updates lessons.md
• Routes AI work through OpenRouter, so you can swap models without vendor lock-in (free tier works)
• Optional X/Twitter content module for drafting and archiving posts
• All state lives in markdown files you already own

Caveats

• Requires three separate moving parts to function: Obsidian, the external hermes-agent CLI, and an OpenRouter API key
• The “self-improving” claim rests on a simple feedback loop written to a single markdown file; the README doesn’t clarify how or if the agent’s behavior materially changes over time
• It is more of a vault skeleton and glue code than a standalone application

Verdict

Worth a look for Obsidian power users who want their notes to trigger action instead of gathering dust. Skip it if you need a fully packaged app or a model that actually retrains itself.