What can you do with this server?

The MCP Memory Gateway is a context engineering server that captures agent feedback, enforces pre-action gates to block known mistakes, and injects relevant past context into AI coding agent sessions for improved reliability and continuity. Feedback & Memory * Capture feedback (capture_feedback, capture_memory_feedback): Record up/down signals with context, reasoning, and rubric scores; vague feedback is rejected with a clarification prompt * Recall past context (recall, commerce_recall): Vector-search relevant past feedback, memories, and prevention rules for the current task * View summaries & analytics (feedback_summary, feedback_stats, dashboard): Approval rate trends, gate enforcement stats, and prevention impact overviews * Generate prevention rules (prevention_rules, get_reliability_rules): Auto-generate blocking rules from repeated failure patterns Pre-Action Gates & Safety * Satisfy gates (satisfy_gate): Record evidence that a gate condition is met (e.g., PR threads checked) with a 5-minute TTL * Gate statistics (gate_stats): See blocked/warned counts and top triggered gates Session Continuity * Session handoff (session_handoff): Write a primer capturing git state, last task, next step, and blockers * Session primer (session_primer): Restore context at session start from the most recent handoff Workflow Planning & Diagnosis * List & plan intents (list_intents, plan_intent): View available workflows and generate checkpointed execution plans with policy gates * Diagnose failures (diagnose_failure): Root-cause analysis for failed or suspect workflow steps * Bootstrap agents (bootstrap_internal_agent): Normalize GitHub/Slack/Linear triggers into startup context with recall packs and worktree sandboxes * Delegation handoffs (start_handoff, complete_handoff): Manage sequential agent delegation with verification outcomes Context Engineering * Context packs (construct_context_pack, evaluate_context_pack): Build and evaluate bounded context packs for large projects, closing the retrieval learning loop * Context provenance (context_provenance): Audit trail of recent context and retrieval decisions * Estimate uncertainty (estimate_uncertainty): Bayesian uncertainty estimates for risky tags before acting Business Metrics * Business metrics (get_business_metrics): Retrieve Revenue, Conversion, and Customer metrics from the Semantic Layer * Semantic entity descriptions (describe_semantic_entity, describe_reliability_entity): Canonical definitions and state of Customer, Revenue, or Funnel entities Export & Fine-Tuning * Export DPO pairs (export_dpo_pairs): Build preference pairs from promoted memories for model fine-tuning * Export Databricks bundle (export_databricks_bundle): Export RLHF logs and proof artifacts as a Databricks-ready analytics bundle * Generate skills (generate_skill): Auto-generate Claude skill files (SKILL.md) from clustered failure patterns

The MCP Memory Gateway is a context engineering server that captures agent feedback, enforces pre-action gates to block known mistakes, and injects relevant past context into AI coding agent sessions for improved reliability and continuity.

Feedback & Memory

Capture feedback (capture_feedback, capture_memory_feedback): Record up/down signals with context, reasoning, and rubric scores; vague feedback is rejected with a clarification prompt
Recall past context (recall, commerce_recall): Vector-search relevant past feedback, memories, and prevention rules for the current task
View summaries & analytics (feedback_summary, feedback_stats, dashboard): Approval rate trends, gate enforcement stats, and prevention impact overviews
Generate prevention rules (prevention_rules, get_reliability_rules): Auto-generate blocking rules from repeated failure patterns

Pre-Action Gates & Safety

Satisfy gates (satisfy_gate): Record evidence that a gate condition is met (e.g., PR threads checked) with a 5-minute TTL
Gate statistics (gate_stats): See blocked/warned counts and top triggered gates

Session Continuity

Session handoff (session_handoff): Write a primer capturing git state, last task, next step, and blockers
Session primer (session_primer): Restore context at session start from the most recent handoff

Workflow Planning & Diagnosis

List & plan intents (list_intents, plan_intent): View available workflows and generate checkpointed execution plans with policy gates
Diagnose failures (diagnose_failure): Root-cause analysis for failed or suspect workflow steps
Bootstrap agents (bootstrap_internal_agent): Normalize GitHub/Slack/Linear triggers into startup context with recall packs and worktree sandboxes
Delegation handoffs (start_handoff, complete_handoff): Manage sequential agent delegation with verification outcomes

Context Engineering

Context packs (construct_context_pack, evaluate_context_pack): Build and evaluate bounded context packs for large projects, closing the retrieval learning loop
Context provenance (context_provenance): Audit trail of recent context and retrieval decisions
Estimate uncertainty (estimate_uncertainty): Bayesian uncertainty estimates for risky tags before acting

Business Metrics

Business metrics (get_business_metrics): Retrieve Revenue, Conversion, and Customer metrics from the Semantic Layer
Semantic entity descriptions (describe_semantic_entity, describe_reliability_entity): Canonical definitions and state of Customer, Revenue, or Funnel entities

Export & Fine-Tuning

Export DPO pairs (export_dpo_pairs): Build preference pairs from promoted memories for model fine-tuning
Export Databricks bundle (export_databricks_bundle): Export RLHF logs and proof artifacts as a Databricks-ready analytics bundle
Generate skills (generate_skill): Auto-generate Claude skill files (SKILL.md) from clustered failure patterns

ThumbGate

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

Self-Healing npm License: MIT Node Sponsor Buy Me a Coffee Pro Pack — 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

Context Engineering Architecture

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

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

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.

License

MIT. See LICENSE.

Install Server

security – no known vulnerabilities

license - permissive license

quality - confirmed to work

How are these scores calculated?

Resources

Need Help?

Report Issue

Related Servers

Tools

View all tools

mcp-memory-gateway