MCP and persistent memory

The Model Context Protocol (MCP) is a standard way for AI clients (like Cursor or Claude Desktop) to discover tools and call them over HTTP. For memory, that means the assistant can pull structured context when relevant and write new entries when you ask, instead of you pasting walls of text.

What the memory tools do

A typical MCP memory server exposes operations such as listing recent memories, searching by keyword or tag, fetching the latest entry, and creating a new memory with title, body, and tags. The LLM chooses when to invoke them; each invocation is a normal HTTP round trip to your MCP endpoint.

Private deploy and egress

In self-hosted mode you can disable model egress on the memory API so the server never calls OpenAI, Anthropic, or Google on your behalf. The MCP server continues to read and write the local database; inference still happens in the client (for example Cursor talking to Claude's API), while your memories never leave your network unless you explicitly allow it.

OAuth vs. agent keys

MCP integrations may use OAuth 2.1 (with a consent screen) or a pre-shared agent key for trusted internal pilots. Both map to the same underlying memory store; pick the mode that matches your security review.