Same memory, three tools. Claude, Cursor, Codex
What tool portability means in practice and why it changes everything.
The proof of the standard
MCP is only a real standard if it switches between tools. Let's verify that live.
Your memory server is running. You've configured Claude Code and worked with it. Now a new project comes up and you want to try Cursor.
Setting up in Cursor
Cursor has a settings menu (Cmd+,), MCP tab. Add the same memory server with the same key. Cursor connects, sees the same tools, accesses the same Learnings.
You work in Cursor, tell it what you're building, it gets written into the same memory. Switch back to Claude Desktop, it knows what Cursor did.
Setting up in Codex
OpenAI Codex's config file is ~/.codex/config.json. Same structure, same env vars. Codex calls the same tools, writes to the same database.
Three tools, one brain.
What this really means
No more lock-in. If a tool stops supporting your way of working, you switch. Memory stays.
Division of labor by tool. Claude Desktop for writing. Claude Code for CLI. Cursor for IDE. Codex for GPT-specific cases. Each tool where it's strongest.
Team capability. Multiple humans can access the same memory. You and your co-founder share project memory.
What this is not
The tools themselves aren't identical. Different strengths, different prices, different context windows.
Memory doesn't make all tools equal. It just lets you switch between them without restarting from zero.
Practical test
Save a new note in Claude Desktop. Call it up in Cursor. Continue in Codex. If that works end-to-end, MCP is real.
Closing Level 4
You have memory-first foundations. In Level 5 comes multi-agent: how multiple AI instances work together with shared memory and different roles. The leap from "I work with AI" to "I direct AI".