Vant builds Vant: an agent OS that remembers its own work
How our agent operating system keeps its memory in git, serves it over MCP, and has been used to build itself through 2,649 commits and 71 merged pull requests.
The problem with amnesia
Every AI session starts from zero. The model is capable, but it does not know what the last session learned, which approach failed, or what the project decided last week. We wanted the opposite: an agent that wakes up with the whole history and writes down what it learned before it sleeps.
Memory as files, not a database
Vant stores its brain as plain markdown in the repository: identity, goals, lessons, errors, and preferences. Git history becomes the memory lineage, so every belief an agent holds has a commit behind it. Stop using Vant and the memory is still there, readable with cat. Three systems cover the need: the markdown brain, a key-value store with TTL, and semantic brain search with git-backed citations.
The agent loop
On wake, the agent loads the brain. It works. On sleep, it writes what it learned. Agents connect over MCP, so any client on any stack uses the same memory: the server listens on localhost and auto-wires the module surface into tools, covering brain read and write, the store, search, and migration status.
It builds itself
The proof is the repository. Vant has 2,649 commits and 71 merged pull requests, and agents did most of that work on Vant itself: they read the brain, changed the code, ran the test suite, and wrote the result back. Succession and trust levels control how much each generation inherits, and the horcrux embeds the whole encrypted brain in an image so it can be restored anywhere.