MCP Memory
MCP Memory is a MCP Server that gives MCP Clients (Cursor, Claude, Windsurf and more) the ability to remember information about users (preferences, behaviors) across conversations. It uses vector search technology to find relevant memories based on meaning, not just keywords. It's built with Cloudflare Workers, D1, Vectorize (RAG), Durable Objects, Workers AI and Agents.
📺 Video
🚀 Try It Out
https://memory.mcpgenerator.com/
🛠️ How to Deploy Your Own MCP Memory
Option 1: One-Click Deploy Your Own MCP Memory to Cloudflare
In Create Vectorize section choose:
- Dimensions: 1024
- Metric: cosine
Click button "Create and Deploy"
In Cloudflare dashboard, go to "Workers & Pages" and click on Visit
Option 2: Use this template
- Click the "Use this template" button at the top of this repository
- Clone your new repository
- Follow the setup instructions below
Option 3: Create with CloudFlare CLI
npm create cloudflare@latest --git https://github.com/puliczek/mcp-memory
🔧 Setup (Only Option 2 & 3)
- Install dependencies:
npm install
- Create a Vectorize index:
npx wrangler vectorize create mcp-memory-vectorize --dimensions 1024 --metric cosine
- Install Wrangler:
npm run dev
- Deploy the worker:
npm run deploy
🧠 How It Works
-
Storing Memories:
- Your text is processed by Cloudflare Workers AI using the open-source
@cf/baai/bge-m3model to generate embeddings - The text and its vector embedding are stored in two places:
- Cloudflare Vectorize: Stores the vector embeddings for similarity search
- Cloudflare D1: Stores the original text and metadata for persistence
- A Durable Object (MyMCP) manages the state and ensures consistency
- The Agents framework handles the MCP protocol communication
- Your text is processed by Cloudflare Workers AI using the open-source
-
Retrieving Memories:
- Your query is converted to a vector using Workers AI with the same
@cf/baai/bge-m3model - Vectorize performs similarity search to find relevant memories
- Results are ranked by similarity score
- The D1 database provides the original text for matched vectors
- The Durable Object coordinates the retrieval process
- Your query is converted to a vector using Workers AI with the same
This architecture enables:
- Fast vector similarity search through Vectorize
- Persistent storage with D1
- Stateful operations via Durable Objects
- Standardized AI interactions through Workers AI
- Protocol compliance via the Agents framework
The system finds conceptually related information even when the exact words don't match.
🔒 Security
MCP Memory implements several security measures to protect user data:
- Each user's memories are stored in isolated namespaces within Vectorize for data separation
- Built-in rate limiting prevents abuse (100 req/min - you can change it in wrangler.jsonc)
- Authentication is based on userId only
- While this is sufficient for basic protection due to rate limiting
- Additional authentication layers (like API keys or OAuth) can be easily added if needed
- All data is stored in Cloudflare's secure infrastructure
- All communications are secured with industry-standard TLS encryption (automatically provided by Cloudflare's SSL/TLS certification)
💰 Cost Information - FREE for Most Users
MCP Memory is free to use for normal usage levels:
- Free tier allows 1,000 memories with ~28,000 queries per month
- Uses Cloudflare's free quota for Workers, Vectorize, Worker AI and D1 database
For more details on Cloudflare pricing, see:
📧 Email API
In addition to storing arbitrary text, MCP Memory can capture emails and related metadata for improved context retrieval. The following endpoints are available:
POST /:userId/emails – Store a new email. Body fields:
subject, body, sender, optional recipients array, date,
messageId, inReplyTo, and company.
GET /:userId/emails – List stored emails for the user.
GET /:userId/emails/search?q=your+query[&company=acme] –
Search stored emails semantically. The optional company
parameter filters results by sender domain.
DELETE /:userId/emails/:memoryId – Remove an email from storage.
❓ FAQ
-
Can I use memory.mcpgenerator.com to store my memories?
- Yes, you can use memory.mcpgenerator.com to store and retrieve your memories
- The service is free
- Your memories are securely stored and accessible only to you
- I cannot guarantee that the service will always be available
-
Can I host it?
- Yes, you can host your own instance of MCP Memory for free on Cloudflare
- You'll need a Cloudflare account and the following services:
- Workers
- Vectorize
- D1 Database
- Workers AI
-
Can I run it locally?
- Yes, you can run MCP Memory locally for development
- Use
wrangler devto run the worker locally - You'll need to set up local development credentials for Cloudflare services
- Note that some features like vector search or workers AI requires a connection to Cloudflare's services
-
Can I use different hosting?
- No, MCP Memory is specifically designed for Cloudflare's infrastructure
-
Why did you build it?
- I wanted an open-source solution
- Control over my own data was important to me
-
Can I use it for more than one person?
- Yes, MCP Memory can be integrated into your app to serve all your users
- Each user gets their own isolated memory space
-
Can I use it to store things other than memories?
- Yes, MCP Memory can store any type of text-based information
- Some practical examples:
- Knowledge Base: Store technical documentation, procedures, and troubleshooting guides
- User Behaviors: Track how users interact with features and common usage patterns
- Project Notes: decisions and project updates
- The vector search will help find related items regardless of content type