DEV Community

Igor Ganapolsky
Igor Ganapolsky

Posted on

I built a proentropic memory layer for AI coding agents — every mistake makes the system stronger

#mcp #ai #opensource #devtools

Every AI coding agent has the same problem: it makes a mistake, you correct it, and next session it makes the exact same mistake again.

I built ThumbGate to fix this. It's an open-source MCP server that turns thumbs-up/down feedback into pre-action gates — hard enforcement that physically blocks the tool call before execution.

How it works

  1. Your agent makes a mistake → you give 👎 with context
  2. ThumbGate auto-generates a prevention rule
  3. Next time the agent tries the same thing → PreToolUse hook fires → BLOCKED

The key insight: every mistake makes the system stronger. More errors = more rules = more reliable agent. It's proentropic — built to get stronger from chaos.

What's included (free)

  • recall — injects past failures at session start
  • Pre-action gates — hard blocks, not prompt suggestions
  • Thompson Sampling — adapts which gates fire
  • Domain skill packs (Stripe, Railway, DB migrations)
  • Hallucination detection — decomposes claims into verifiable sub-claims
  • PII scanning — blocks sensitive data before export
  • Works with Claude Code, Cursor, Codex, Gemini, Amp, OpenCode

Install 

npx mcp-memory-gateway init --agent claude-code
Enter fullscreen mode Exit fullscreen mode

GitHub: https://github.com/IgorGanapolsky/ThumbGate
npm: https://www.npmjs.com/package/mcp-memory-gateway

Pro ($19/mo, 7-day free trial) adds a personal dashboard and DPO export for fine-tuning.

Top comments (0)