ThumbGate

npm package: mcp-memory-gateway — install with npx mcp-memory-gateway init

— Solo dev? Free tier has everything you need. Team or multi-repo? Pro syncs prevention rules across machines and team members. $49 one-time.

Thumbs down a mistake. It never happens again.

The safety net for vibe coding. Give your AI agent a thumbs-down and it auto-generates a prevention rule. Give a thumbs-up and it reinforces good behavior. Pre-action gates physically block the agent before it repeats a known mistake — a reliability layer for one sharp agent, without another planner or swarm.

Honest disclaimer: this is not RLHF weight training. ThumbGate is context engineering plus enforcement. Feedback becomes searchable memory, prevention rules, and gates that block known-bad actions before they execute.

Works with Claude Code, Cursor, Codex, Gemini, Amp, OpenCode, and any MCP-compatible agent.

Live Demo Dashboard | Landing Page | Verification Evidence

Most memory tools only help an agent remember. ThumbGate also enforces.

The problem without it:

BEFORE: Agent force-pushes to main. You correct it. Next session, it force-pushes again.

With ThumbGate (mcp-memory-gateway):

AFTER: Gate blocks the force-push before it executes. Agent can't repeat the mistake.

• recall injects the right context at session start.
• search_lessons shows promoted lessons plus the corrective action, lifecycle state, linked rules, linked gates, and the next harness fix the system should make.
• search_rlhf searches raw RLHF state across feedback logs, ContextFS memory, and prevention rules.
• Pre-action gates physically block tool calls that match known failure patterns.
• Session handoff and primer keep continuity across sessions without adding an extra orchestrator.

Free and self-hosted users can invoke search_lessons directly through MCP, and via the CLI with npx mcp-memory-gateway lessons.

See it in action

$ npx mcp-memory-gateway serve
[gate] ⛔ Blocked: git push --force (rule: no-force-push, confidence: 0.94)
[gate] ✅ Passed: git push origin feature-branch

Quick Start

# One command install — auto-detects your agent
npx mcp-memory-gateway init

# Or add the MCP server directly
claude mcp add rlhf -- npx -y mcp-memory-gateway serve
codex mcp add rlhf -- npx -y mcp-memory-gateway serve
amp mcp add rlhf -- npx -y mcp-memory-gateway serve
gemini mcp add rlhf "npx -y mcp-memory-gateway serve"

# Wire PreToolUse enforcement hooks
npx mcp-memory-gateway init --agent claude-code
npx mcp-memory-gateway init --agent codex
npx mcp-memory-gateway init --agent gemini

# Health check and inspect lessons
npx mcp-memory-gateway doctor
npx mcp-memory-gateway lessons
npx mcp-memory-gateway dashboard

How It Works

1. You give feedback    →  👎 "Force-pushed and lost commits"
2. ThumbGate validates  →  Rejects vague signals, promotes actionable ones
3. Rules auto-generate  →  "Block git push --force to protected branches"
4. Gates enforce        →  PreToolUse hook fires → BLOCKED before execution
5. Agent improves       →  Same mistake never happens again

Pipeline: Capture → Validate → Remember → Distill → Prevent → Gate → Export

Pre-Action Gates

Gates are the enforcement layer. They do not ask the agent to cooperate — they physically block the action.

Agent tries git push --force
  → PreToolUse hook fires
  → gates-engine checks rules
  → BLOCKED: no force pushes to protected branches

Built-in gates:

• push-without-thread-check — block push if PR threads unresolved
• force-push — block git push --force to protected branches
• protected-branch-push — block direct pushes to main/master
• package-lock-reset — block destructive lock file changes
• env-file-edit — block edits to .env files with secrets

Define custom gates in config/gates/custom.json.

What Actually Works

Actually works	Does not work
recall injects past context into the next session	Thumbs up/down changing model weights
session_handoff and session_primer preserve continuity	Agents magically remembering what happened last session
search_lessons exposes corrective actions, lifecycle state, linked rules, linked gates, and next harness fixes	Feedback stats automatically improving behavior by themselves
Pre-action gates block known-bad tool calls before execution	Agents self-correcting without context injection or gates
Auto-promotion turns repeated failures into warn/block rules	Calling this "RLHF" in the strict training sense
Rejection ledger shows why vague feedback was rejected	Vague signals silently helping the system

Core MCP Tools

Essential profile

Tool	Purpose
capture_feedback	Accept up/down signal + context, validate, promote to memory
recall	Recall relevant past failures and rules for the current task
search_lessons	Search promoted lessons with corrective action, lifecycle state, rules, gates
search_rlhf	Search raw RLHF state across feedback logs, ContextFS, and rules
prevention_rules	Generate prevention rules from repeated mistakes
enforcement_matrix	Inspect promotion rate, active gates, and rejection ledger
feedback_stats	Approval rate and failure-domain summary
estimate_uncertainty	Bayesian uncertainty estimate for risky tags

Lean install for recall + gates + lesson search only:

RLHF_MCP_PROFILE=essential claude mcp add rlhf -- npx -y mcp-memory-gateway serve

Free and self-hosted users can invoke search_lessons directly through MCP to inspect corrective action per lesson. For broader retrieval across feedback logs, ContextFS memory, and prevention rules, use search_rlhf through MCP or the authenticated GET /v1/search API.

Dispatch profile

Phone-safe read-only surface for remote ops:

RLHF_MCP_PROFILE=dispatch claude mcp add rlhf -- npx -y mcp-memory-gateway serve
npx mcp-memory-gateway dispatch

Guide: docs/guides/dispatch-ops.md

ThumbGate vs Alternatives

Feature	ThumbGate	SpecLock	Mem0	.cursorrules
Blocks mistakes before execution	Yes — PreToolUse gates	Yes — Patch Firewall	No	No
Learns from your feedback	Yes — thumbs up/down	No — manual spec writing	Yes — auto-capture	No
Works across sessions	Yes — SQLite + JSONL	Yes — encrypted store	Yes — cloud	No — per-project
Auto-generates rules	Yes — from repeated failures	No — manual or Gemini compile	No	No
Agent support	Claude Code, Codex, Gemini, Amp, Cursor, OpenCode	Claude Code, Cursor, Windsurf, Cline, Bolt.new	Claude, Cursor	Cursor only
Install	npx mcp-memory-gateway init	npx speclock setup	Cloud signup	Edit file
Cost	Free (Pro $49 for teams)	Free	Free tier + paid	Free
npm weekly downloads	724	98	N/A	N/A

When to use ThumbGate: You want your agent to learn from mistakes automatically and enforce what it learned. One thumbs-down creates a gate.

When to use SpecLock: You have a written spec/PRD and want to lock specific sections from AI modification. Manual constraint authoring.

When to use Mem0: You want cloud-hosted memory shared across apps. No enforcement.

Tech Stack

Core runtime

• Node.js >=18.18.0
• Module system: CommonJS CLI/server runtime
• Primary entry points: CLI, MCP stdio server, authenticated HTTP API, OpenAPI adapters

Interfaces

• MCP stdio: adapters/mcp/server-stdio.js
• HTTP API: src/api/server.js
• OpenAPI surfaces: openapi/openapi.yaml, adapters/chatgpt/openapi.yaml
• CLI: npx mcp-memory-gateway ...

Storage and retrieval

• Local memory: JSONL logs in .claude/memory/feedback or .rlhf/*
• Lesson DB (v0.8.0): SQLite + FTS5 full-text search via better-sqlite3 — dual-written alongside JSONL. Indexed by signal, domain, tags, importance. Replaces linear Jaccard token-overlap with sub-millisecond ranked search.
• Corrective actions (v0.8.0): On negative feedback, capture_feedback returns correctiveActions[] — top 3 remediation steps inferred from similar past failures by tag/domain overlap.
• Context assembly: ContextFS packs and provenance logs
• Default retrieval path: SQLite FTS5 (primary) with JSONL Jaccard fallback
• Semantic/vector lane: LanceDB + Apache Arrow + local embeddings via Hugging Face Transformers

Intelligence layer

• MemAlign-inspired dual recall: Principle-based memory (distilled rules) + episodic context (raw feedback with timestamps). Recall surfaces both lanes ranked by relevance.
• Thompson Sampling: Bayesian multi-armed bandit over feedback tags — adapts gate sensitivity per failure domain based on observed positive/negative signal ratios.
• Corrective action inference: On negative feedback, the lesson DB infers top-3 remediation steps from similar past failures by tag/domain overlap.
• Bayesian belief update: Each memory carries a posterior belief that updates on new evidence — high-entropy contradictions auto-prune.

Enforcement and automation

• PreToolUse enforcement: scripts/gates-engine.js
• Hook wiring: init --agent claude-code|codex|gemini
• Browser automation / ops: playwright-core
• Social analytics store: better-sqlite3

Billing and hosting

• Billing: Stripe
• Hosted API / landing page: Railway
• Worker lane: Cloudflare Workers in workers/

Agent Integration Guides

• Claude Desktop extension
• Cursor plugin operations
• Continuity tools integration
• OpenCode integration

Operator Contract

For autonomous agent runs against this or any repo using this workflow:

• WORKFLOW.md — scope, proof-of-work, hard stops, done criteria
• .github/ISSUE_TEMPLATE/ready-for-agent.yml — bounded intake template
• .github/pull_request_template.md — proof-first PR handoff

Pro Pack

$49 one-time — hosted dashboard, priority support, commercial license.

• Commercial Truth
• Verification Evidence
• Pitch
• Anthropic Marketplace Strategy

License

MIT. See LICENSE.