Persistent memory management and pre-action gates for AI coding agents.
MCP Memory Gateway
Pre-action gates that physically block AI coding agents from repeating known mistakes. Dual-memory recall (MemAlign-inspired principles + episodic context). Captures feedback, auto-promotes failures into prevention rules, and enforces them via PreToolUse hooks. Works with Claude Code, Codex, Gemini, Amp, Cursor.
Honest disclaimer: This is a context injection system, not RLHF. LLM weights are not updated by thumbs-up/down signals. What actually happens: feedback is validated, promoted to searchable memory, and recalled at session start so agents have project history they'd otherwise lose. That's genuinely valuable — but it's context engineering, not reinforcement learning.
Works with any MCP-compatible agent: Claude, Codex, Gemini, Amp, Cursor, OpenCode.
Verification evidence for shipped features lives in docs/VERIFICATION_EVIDENCE.md.
Repo-local operator guides:
MCP Memory Gateway keeps one sharp agent on task. Continuity tools help you resume work. The resumed session stays sharper with recall, reliability rules, pre-action gates, session handoff primers, and verification layered on top of that continuity workflow without another planner or swarm.
Claude Workflow Hardening
If you are selling or deploying Claude-first delivery, the cleanest commercial wedge is not "AI employee" hype. It is a Workflow Hardening Sprint for one workflow with enough memory, gates, and proof to ship safely.
Use that motion when a buyer already has:
- one workflow owner
- one repeated failure pattern or rollout blocker
- one buyer who needs proof before broader rollout
That maps cleanly to three offers:
- Workflow Hardening Sprint for one production workflow with business value
- code modernization guardrails for long-running migration and refactor sessions
- hosted Pro at
$49 one-timewhen the team only needs synced memory, gates, and usage analytics
Use these assets in sales and partner conversations:
Claude Desktop Extensions
This repo already ships a Claude Desktop extension lane:
- Claude metadata:
.claude-plugin/plugin.json - Claude marketplace metadata:
.claude-plugin/marketplace.json - Claude extension install and support guide:
.claude-plugin/README.md - Claude Desktop bundle builder:
npm run build:claude-mcpb - Claude Desktop bundle launcher:
.claude-plugin/bundle/server/index.js - Claude Desktop bundle icon:
.claude-plugin/bundle/icon.png - Internal submission packet: docs/CLAUDE_DESKTOP_EXTENSION.md
Install locally today with:
claude mcp add rlhf -- npx -y mcp-memory-gateway serve
Build a submission-ready .mcpb locally with:
npm run build:claude-mcpb
Treat Anthropic directory inclusion as a discoverability and trust lane, not as revenue proof or partner proof.
For paired phone + desktop workflows, keep Dispatch in a constrained remote-ops lane:
RLHF_MCP_PROFILE=dispatch claude mcp add rlhf -- npx -y mcp-memory-gateway serve
npx mcp-memory-gateway dispatch
That profile stays read-only: metrics, gates, diagnostics, planning, and recall. Use a dedicated worktree plus RLHF_MCP_PROFILE=default when the task graduates into code edits or memory writes. Guide: docs/guides/dispatch-ops.md.
Cursor Marketplace
This repo now ships a submission-ready Cursor plugin bundle:
- Root marketplace manifest:
.cursor-plugin/marketplace.json - Plugin direc
Environment Variables
RLHF_MCP_PROFILESets the operational profile for the gateway (e.g., 'dispatch' or 'default')Configuration
{"mcpServers": {"rlhf": {"command": "npx", "args": ["-y", "mcp-memory-gateway", "serve"]}}}