# What is an MCP memory server? Durable codebase memory for any AI tool

An MCP memory server lets any MCP-compatible AI tool read a persistent memory of your codebase instead of re-scanning it every session. Here's what it is and how to set one up.

An **MCP memory server** is a Model Context Protocol server that gives AI tools a persistent, shared memory — so any MCP-compatible tool (Claude Code, Cursor, Windsurf, VS Code, and hosted agents) reads it through one standard interface instead of rebuilding context every session. For a codebase, that memory is a structure-aware map of your repo.

## What MCP is, briefly

The Model Context Protocol is an open standard for connecting AI tools to external context. A tool that "speaks MCP" can plug into any MCP server without custom integration. That's what makes a memory server valuable: write once, read from every tool.

## What an MCP memory server does for a codebase

Instead of each tool scanning your files from scratch, it exposes:

- a structural map of your repo (symbols, signatures, routes)
- a resolved call graph (who calls what, and the blast radius of a change)
- search over the codebase
- your recorded decisions and conventions

Your AI tool queries these over MCP and gets accurate, compact context — often [several times fewer tokens](/benchmark) than reading the files cold.

## Why "memory", not just "search"

A plain search server finds text. A memory server stays *correct*: it's anchored to your code's structure, so when you rename a function or change the schema, the memory updates and stale notes are flagged. Search finds; memory remembers, and knows when it's wrong.

## How to set one up

With SecondOS:

1. Run **bunx @secondos/cli push .** to build and sync a structure-aware map of your repo (your source stays local).
2. Run **bunx @secondos/cli mcp-config** to print ready-to-paste MCP config for Claude Code, Cursor, Windsurf and VS Code.
3. Ask your tool a question — it reads the map over MCP.

For hosted agents (claude.ai, ChatGPT) that can't run a local process, connect over HTTP to the remote MCP endpoint.

[MCP docs](/docs/mcp) · [quickstart](/docs/quickstart) · [remote MCP](/docs/remote-mcp).