vouch is a git-native knowledge base that puts a review gate between agent proposals and human approval. It's the difference between treating agent-generated knowledge like code (reviewed, cited, auditable) and treating it like a chatbot dump (one-way, unverifiable, lost next session).
The Problem
LLM agents are great at learning context—but terrible at knowing when they're wrong. They'll confidently assert false facts, hallucinate citations, and propose contradictory claims. Meanwhile:
- mem0 / Letta store knowledge in a service (vendor lock-in, no audit trail)
- LLM-Wiki tools skip evidence entirely (claims with no sources)
- Obsidian plugins have no review gate (any agent write sticks)
You end up choosing between agent amnesia and unverified noise.
The Solution: Review-Gated Knowledge
Vouch treats agent-written knowledge the way your repo treats agent-written code:
-
Agents propose claims to
.vouch/proposed/(draft state) - You review via PR — is this claim true? Is the source cited? Is it contradictory?
-
You approve → it moves to
.vouch/claims/(permanent, committed) - Next session, agents fetch from approved claims, not proposals
Everything is durable, reviewable, auditable:
- Plain YAML/Markdown — diffs cleanly in PRs
- Content-addressed sources — cite code, transcripts, URLs; duplicates de-duplicate
- Append-only audit log — every mutation is signed with actor + timestamp
- Per-agent attribution — you know which agent claimed what
Why This Matters
For teams running Claude Code, Cursor, and CI bots in the same repo:
- Agents share context instead of re-explaining on each session
- One source of truth — not scattered across Slack notes, Discord, and nobody's laptop
- Legal/compliance readiness — "who decided this, citing what, when?" is answerable
- Decision records without ceremony — crystallize a session's insight into a claim with one command
Top comments (2)
github.com/vouchdev/vouch/
Some comments may only be visible to logged-in visitors. Sign in to view all comments.