MCP Memory Gateway

Deterministic pre-action governance for AI agents that blocks repeated mistakes before they execute. Hooks into PreToolUse to intercept risky commands like force pushes or destructive file operations, matching them against prevention rules built from your own thumbs-down feedback. Works with Claude Code, Cursor, Codex, Gemini CLI, Amp, Cline, and OpenCode. The core mechanic is simple: agent tries a bad action, you reject it once, ThumbGate blocks it permanently in future sessions before any tokens are spent. Also generates a versioned BRAIN.md artifact that consolidates lessons, guardrails, and gates into agent-readable context. Useful when you're paying per-token for frontier models and tired of debugging the same hallucinated function calls or policy violations across sessions.