We built MemorySync, a hosted memory API for AI agents. In Cursor it runs as a remote MCP server: Cursor can save the decisions, conventions and constraints you tell it, and search them in later sessions instead of relying on the chat history.
To add it, put this in .cursor/mcp.json. The first time, Cursor asks you to sign in to MemorySync.
{
"mcpServers": {
"memorysync": { "url": "https://mcp.memorysync.io/mcp" }
}
}
To try it before adding anything, Playground – Persistent AI Agent Memory in Your Browser | MemorySync runs the whole loop in the browser with no account: save a fact, reload the page, ask for it back. The starter repo at GitHub - memorysyncio/memorysync-cursor-starter: Starter template for Cursor AI with persistent memory and live docs over Model Context Protocol (MCP). · GitHub has a working mcp.json and .cursorrules.
One thing it doesn’t handle yet: when a fact changes, and it isn’t a one-value fact, like “we moved from MySQL to Postgres”, the old and new facts both come back. How do you deal with stale context in Cursor today?
Happy to answer questions here.