Docs · Repo memory
Corvi learns your repo's conventions.
The first plan for a repo starts cold. The next one shouldn't — Corvi remembers the durable facts it learns and brings them to the next run. Nothing is silent, and you can forget anything it kept.
What it learns
After a run, Corvi extracts the durable facts about a repository that help the next plan land cleaner — the conventions, the patterns, the things-that-bit-you. These are facts about how your code is organised and how work gets done in it, not a copy of the code.
Where it's kept
Learned facts live with the repository, in a .corvi.md file under a machine-owned section. It's plain text in your tree — read it, commit it so a teammate inherits it, or leave it untracked. Corvi only ever rewrites its own section; anything you write above it is left alone.
Never silent, always opt-in
Repo memory is opt-in, and Corvi never writes a learned fact behind your back. After a run it shows what it proposes to remember and you approve it. What gets kept is bounded and deduplicated, so the memory stays a short, useful list rather than a growing pile.
Review and forget
Open what Corvi has learned at any time and read it back as a structured list. Drop a fact you don't want carried forward and it's removed exactly — look-alike facts aren't co-deleted, and the rest is preserved. Memory is a default you can always audit, never a black box.
It feeds the next plan
On the next run, the kept conventions inform the draft so the plan fits how your repo actually works. As with everything Corvi proposes, the plan is still a draft you approve — see the planning model.