Your agent proposes. You approve. Approved knowledge compiles into a cited wiki that lives in your repo — and every new session starts knowing it.
vouch is the memory your agent has to earn. Agents propose claims with cited, content-hashed sources; a human approves them through a review gate; approved claims compile into wiki pages; and each session recalls what your team actually decided. Knowledge is plain YAML and markdown in .vouch/ inside your repo — diffs, blame, and an append-only audit log included.
pipx install vouch-kb

Start here

Quickstart

Init a knowledge base, connect your agent, and run the propose → approve → recall loop in minutes.

Your first knowledge base

A verified, end-to-end tutorial — every command was run against the current build before it went in.

The review gate

The load-bearing invariant: why every write waits for a human, and what that buys you.

Connect a coding agent

One command per host — Claude Code, Cursor, Codex, Windsurf, Zed, and any MCP-aware agent.

Why vouch

Review-gated

Every write becomes a pending proposal. The approver must differ from the proposer, and there is no side door around proposals.approve().

Git-native

Knowledge is YAML + markdown in your repo. Branch it, PR it, blame it — one decision is one commit.

Cited, hashed sources

A claim without evidence is a validation error. Sources are content-hashed at registration and re-verifiable later.

Local-first

No cloud, no account, no telemetry. The search index is a derived, disposable SQLite file.

With and without

Without vouch, every session starts from amnesia:
you    > deploy the fix
agent  > Sure — what's the deploy process for this repo?
you    > (explains make release, tagged commits, canary… again)
With vouch, the session starts knowing what you approved:
[SessionStart] vouch recall
<vouch-approved-knowledge>
[claim/deploy-tagged-commit] production deploys run `make release`…
[claim/canary-first] deploys hit canary before fleet
</vouch-approved-knowledge>

agent  > Tagging v1.4.2 and running `make release` — canary first,
         per claim/canary-first.
The difference is not just recall — it’s that everything recalled was reviewed by a human and cites a source you can still verify.

The loop

  1. Propose — the agent files a claim with cited, content-hashed evidence.
  2. Review — it waits in the queue: pending, visible, diffable.
  3. Approve — a human stamps it; the decision is one audited commit.
  4. Compile — approved claims compile into cited wiki pages.
  5. Recall — every new session starts knowing what you approved.
Read the review gate for the invariant, the object model for what lives on disk, and compile for how pages — the product — get built.