I built AGENTS.db which keeps the spirit of AGENTS.md while scaling it into a layered, append‑only, vectorized flatfile database for LLM agents.
Instead of one mutable markdown file, context lives in layers:
- Base - immutable, human‑verified source of truth
- User - durable human additions
- Delta - proposed / reviewable changes
- Local - ephemeral session notes
Higher layers override lower ones (`local > user > delta > base`), with full provenance and fast local semantic search.
No server. No SaaS. Works offline. Source‑control friendly. Exposes an MCP server so agents can read/write context safely instead of rewriting docs.
This is an early reference implementation targeting a public spec, and I’m trying to pressure‑test whether this is a better long‑term primitive than “just keep adding to AGENTS.md”.