The recurring issue I saw: when agents write directly into memory/tasks, quality drifts quickly. It's hard to answer what changed, who approved it, and how to roll it back safely.

Memrail treats writes like pull requests: - dry-run first - diff preview - human approve/reject - commit - audit trail + undo

Current surface: - `/changes`: review inbox (commit/reject/undo) - `/tasks`: execution workspace - `/knowledge`: governed knowledge CRUD

Stack: - FastAPI + SQLAlchemy - Next.js - SQLite default, PostgreSQL optional

Repo: https://github.com/zhuamber370/memrail

I would value feedback on: 1) Where this governance gate should sit in an agent stack 2) Which diff/audit details are non-negotiable for real ops 3) What would block you from trying this