All TutorialsOVERVIEW
The Memory MCP Server is an official TypeScript reference server in the modelcontextprotocol/servers repository. The main README describes it as a "Knowledge graph-based persistent memory system" that lets AI assistants remember entities and relations across conversations instead of forgetting everything when a chat ends.
This guide uses only verified commands from https://github.com/modelcontextprotocol/servers. The canonical one-line install is documented in the Getting Started section of that README.
WHAT IT DOES
Memory exposes tools to create, query, and update a local knowledge graph. Entities might represent people, projects, preferences, or technical decisions; relations connect them ("uses", "depends on", "prefers"). Unlike stuffing facts into a long system prompt, Memory persists structured data the model can retrieve through MCP tool calls in future sessions.
The directory listing in this site cites nine Memory tools. The server runs as a long-lived stdio process started by your MCP client — you do not run a separate database server.
PREREQUISITES
- Node.js 18 or newer
- An MCP-compatible client
- Awareness of what you allow the assistant to remember (avoid secrets)
Verify: node --version
STANDALONE INSTALL COMMAND (OFFICIAL)
From the modelcontextprotocol/servers README Getting Started section:
npx -y @modelcontextprotocol/server-memory
Running this in a terminal starts the server on stdio for debugging. In daily use, your MCP client launches the same command automatically.
CLAUDE DESKTOP CONFIGURATION (OFFICIAL EXAMPLE)
The README shows this exact configuration:
{
"mcpServers": {
"memory": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-memory"]
}
}
}
Config file path:
- macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
- Windows: %APPDATA%\Claude\claude_desktop_config.json
- Linux: ~/.config/Claude/claude_desktop_config.json
WINDOWS NOTE (OFFICIAL)
Wrap npx with cmd /c on Windows per the same README:
{
"mcpServers": {
"memory": {
"command": "cmd",
"args": ["/c", "npx", "-y", "@modelcontextprotocol/server-memory"]
}
}
}
Leave uvx-based servers unchanged — only npx entries need this wrapper.
CURSOR NOTES
Add Memory to ~/.cursor/mcp.json or project .cursor/mcp.json using the identical JSON structure. Memory requires no path arguments, making it one of the simplest servers to demo MCP. Restart Cursor after editing config. In Agent mode, verify whether memory writes require approval in your MCP settings.
CLINE NOTES
Configure Memory in Cline's MCP Servers panel with the same block. Cline agents can store milestones during long tasks — for example remembering that you chose PostgreSQL over SQLite. Review Cline's autonomy settings so memory updates align with your workflow.
SECURITY NOTES
Official reference servers are not audited for regulated data. Do not ask the assistant to remember passwords, API keys, or personal data you would not store in plain local files. Periodically inspect what the knowledge graph contains via Memory tools. The servers README emphasizes evaluating your own security requirements.
Memory complements but does not replace encrypted secret stores or enterprise knowledge bases.
FAQ
Q: Where is data stored?
A: Consult the Memory server README in the modelcontextprotocol/servers repo for persistence details on disk.
Q: Does Memory sync across devices?
A: No — it is local to the machine running the MCP server unless you add your own sync layer.
Q: Memory vs Filesystem notes?
A: Memory provides structured graph queries. Filesystem markdown files are human-editable but unstructured.
Q: Command source?
A: Verified from https://github.com/modelcontextprotocol/servers README.
TROUBLESHOOTING
If Memory does not appear, confirm valid JSON and full Claude Desktop restart. Test npx -y @modelcontextprotocol/server-memory in a terminal. Ensure Node 18+ is on the PATH visible to GUI apps on macOS (sometimes requires shell profile fixes).
USAGE TIPS
Ask the assistant explicitly to remember stable facts: coding conventions, deployment URLs (non-secret), team timezone preferences. Pair with Sequential Thinking for complex planning sessions where conclusions should be persisted after reasoning completes.
KNOWLEDGE GRAPH CONCEPTS
Entities are nodes — a project, a person, a library version. Relations are edges — "project uses library", "person owns service". The Memory server tools let the model create and query this graph without you hand-maintaining JSON. This differs from RAG vector databases: Memory is structured and lightweight, not designed for embedding thousands of PDF pages.
MULTI-CLIENT DEPLOYMENT
The same Memory server config works in Claude Desktop, Cursor mcp.json, and Cline MCP settings because all speak standard MCP over stdio. You typically run one Memory process per client instance — they do not automatically sync graphs across machines unless they share the same on-disk store by coincidence of paths.
EVERYTHING SERVER CONTRAST
npx -y @modelcontextprotocol/server-everything is for protocol testing, not persistence. Install Memory when you want durable cross-chat recall; install Everything when validating that tools, prompts, and resources render correctly in a new MCP client build.
EXTENDED FAQ
Q: Will Memory slow Claude down?
A: Each tool call adds latency, but Memory reads are usually small compared to large file reads.
Q: Can I reset Memory?
A: Consult the Memory README for how to clear stored graph data on disk.
Q: Is Memory HIPAA-compliant?
A: No certification is implied — reference server only.
Q: Package name verification?
A: @modelcontextprotocol/server-memory per official README — not third-party memory packages.
How to Install the Memory MCP Server
Give your AI assistant persistent memory across conversations with the Memory MCP server.