The latest articles on DEV Community by plind (@plindjunior). https://dev.to/plindjunior https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4010696%2Ffa58483f-d608-4e82-9237-4577d32680de.jpg https://dev.to/plindjunior en plind Wed, 01 Jul 2026 09:47:54 +0000 https://dev.to/plindjunior/stop-agents-from-breaking-your-knowledge-base-meet-vouch-a-review-gated-kb-for-llm-workflows-37bn https://dev.to/plindjunior/stop-agents-from-breaking-your-knowledge-base-meet-vouch-a-review-gated-kb-for-llm-workflows-37bn <p><strong>vouch</strong> is a git-native knowledge base that puts a review gate between agent <em>proposals</em> and human <em>approval</em>. 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).</p> <h3> The Problem </h3> <p>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:</p> <ul> <li> <strong>mem0 / Letta</strong> store knowledge in a service (vendor lock-in, no audit trail)</li> <li> <strong>LLM-Wiki tools</strong> skip evidence entirely (claims with no sources)</li> <li> <strong>Obsidian plugins</strong> have no review gate (any agent write sticks)</li> </ul> <p>You end up choosing between agent amnesia and unverified noise.</p> <h3> The Solution: Review-Gated Knowledge </h3> <p>Vouch treats agent-written knowledge the way your repo treats agent-written code:</p> <ol> <li> <strong>Agents propose</strong> claims to <code>.vouch/proposed/</code> (draft state)</li> <li> <strong>You review</strong> via PR — is this claim true? Is the source cited? Is it contradictory?</li> <li> <strong>You approve</strong> → it moves to <code>.vouch/claims/</code> (permanent, committed)</li> <li> <strong>Next session</strong>, agents fetch from approved claims, not proposals</li> </ol> <p>Everything is durable, reviewable, auditable:</p> <ul> <li> <strong>Plain YAML/Markdown</strong> — diffs cleanly in PRs</li> <li> <strong>Content-addressed sources</strong> — cite code, transcripts, URLs; duplicates de-duplicate</li> <li> <strong>Append-only audit log</strong> — every mutation is signed with actor + timestamp</li> <li> <strong>Per-agent attribution</strong> — you know which agent claimed what</li> </ul> <h3> Why This Matters </h3> <p>For teams running Claude Code, Cursor, and CI bots in the same repo:</p> <ul> <li> <strong>Agents share context</strong> instead of re-explaining on each session</li> <li> <strong>One source of truth</strong> — not scattered across Slack notes, Discord, and nobody's laptop</li> <li> <strong>Legal/compliance readiness</strong> — "who decided this, citing what, when?" is answerable</li> <li> <strong>Decision records without ceremony</strong> — crystallize a session's insight into a claim with one command</li> </ul> ai programming agents