RLM Memory MCP Server

Persistent memory + semantic file discovery for ANY AI coding agent.

Works with Claude Code, OpenAI Codex, Gemini CLI, Cursor, Windsurf — anything that speaks MCP.

What is this?

AI agents forget everything between sessions. This MCP server fixes that:

• 🧠 Memory — after every task, the agent records what it changed and why. Next session, it remembers.
• 🗺️ File map — a semantic index of your codebase ("this file is the login form", "this is the checkout API"), so the agent finds the right files without grepping the whole repo.
• 🔄 Bi-directional — the agent asks the MCP ("user wants to fix the submit button — which files?") and the MCP answers with files, history, and suggestions.

The core idea (Recursive Large Model): the agent stays intentionally "blind" to the filesystem and uses the MCP as its eyes and memory — making it faster, cheaper, and more focused.

┌────────────────────────────────────────────────────────┐
│  AI Agent (Claude Code, Codex, Gemini CLI, Cursor...)  │
│                                                        │
│  "User wants to fix login" ──► rlm_query               │
│                            ◄── relevant files+history  │
│  [does the work]                                       │
│  "Here's what I changed"   ──► rlm_smart_memory        │
└────────────────────────────────────────────────────────┘
                          │ stdio (MCP)
┌────────────────────────────────────────────────────────┐
│  RLM Memory MCP Server                                 │
│  • JSON storage per project (projects/<name>/.rlm/)    │
│  • AI-powered matching via OpenRouter or Gemini        │
│  • Web UI for you at http://localhost:3848             │
└────────────────────────────────────────────────────────┘

Setup in 3 Steps

1. Install & build

git clone https://github.com/jumpino27/RLM-Memory-MCP-Server.git
cd RLM-Memory-MCP-Server
npm install
npm run build

2. Add an API key

Copy .env.example to .env and set one key:

# Option A (recommended): OpenRouter — one key, any model
# https://openrouter.ai/keys
OPENROUTER_API_KEY="sk-or-..."

# Option B: Google Gemini direct — https://aistudio.google.com/
# GEMINI_API_KEY="..."

• With OpenRouter the server uses google/gemini-3.5-flash by default — fast, cheap, near-Pro quality.
• With Gemini direct it uses gemini-3.5-flash.
• No key at all? Everything still works using keyword matching (just less smart).

Want a different model? Set LLM_MODEL (e.g. anthropic/claude-haiku-4.5 or openai/gpt-4o-mini on OpenRouter). See .env.example for all options.

3. Connect your AI agent

Replace C:\\path\\to with where you cloned the repo.

Claude Code (one command):

claude mcp add rlm-memory -- node C:\\path\\to\\RLM-Memory-MCP-Server\\dist\\index.js

OpenAI Codex CLI — add to ~/.codex/config.toml:

[mcp_servers.rlm-memory]
command = "node"
args = ["C:\\path\\to\\RLM-Memory-MCP-Server\\dist\\index.js"]

Gemini CLI — add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "rlm-memory": {
      "command": "node",
      "args": ["C:\\path\\to\\RLM-Memory-MCP-Server\\dist\\index.js"]
    }
  }
}

Any other MCP client: launch node dist/index.js over stdio.

💡 Tell your agent how to use it: copy the rules from example_agents.md into your agent's instructions file (CLAUDE.md, AGENTS.md, .cursorrules, …).

The Workflow

First time on a project — index it once:

rlm_init → rlm_index_codebase → rlm_verify_index

Every task after that — three steps:

1. rlm_query          "User wants X — which files?"   → files + history + tips
2. (agent does the actual work)
3. rlm_smart_memory   "Here's what I changed"         → remembered forever

That's it. The more the agent works, the smarter the memor

... View full README on GitHub