DEV Community

Thomas Jumper
Thomas Jumper

Posted on • Edited on • Originally published at stremai.com

How to give Claude Code, Cursor, and Codex shared memory over MCP

#ai #mcp #productivity #claudecode

I use more than one coding agent in a normal week: Claude Code for most work, Cursor when I'm living in the editor, Codex CLI for scripted runs. They're all good. They also share nothing.

A lesson learned in one session ("staging DB resets Sunday nights", "that test is flaky, ignore it", "we chose pgvector on purpose, stop suggesting alternatives") is gone in the next session, and was never visible to the other tools at all.

The built-in memory options don't cross that gap, and their docs say so: Claude Code's auto memory is per-repository and machine-local; Codex's memories are stored locally and off by default; Cursor's documented persistence layer is Rules, which are files you write yourself.

Instruction files like CLAUDE.md, AGENTS.md, and .cursor/rules are genuinely useful for the context you write. Keep them. What they don't hold is the context your agents learn while working.

The fix is a shared memory layer over MCP: one server that any MCP client can connect to, so a memory stored by one connected agent can be recalled by another.

Here's the full setup with StremAI. Full disclosure: I'm the founder. There are also open-source options like OpenMemory, agentmemory, Memorix, and others if you'd rather self-host, and the verification steps below work with any shared-memory MCP server.

1. Claude Code 

claude mcp add --transport http stremai https://stremai.com/api/mcp --scope user
Enter fullscreen mode Exit fullscreen mode

Run it, then approve the browser sign-in. No API key to paste. --scope user makes the connection available in every project, not just the current directory.

2. Cursor

Add one server entry to ~/.cursor/mcp.json for global use, or .cursor/mcp.json for one project:

{
  "mcpServers": {
    "stremai": {
      "url": "https://stremai.com/api/mcp"
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Cursor should prompt a browser sign-in. Honest note: OAuth verification for Cursor is still in progress on our side. If the prompt doesn't appear in your build, use an API key fallback:

{
  "mcpServers": {
    "stremai": {
      "url": "https://stremai.com/api/mcp",
      "headers": {
        "Authorization": "Bearer ab_live_YOUR_KEY"
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Existing "agentbay" entries keep working; use "stremai" for new installs so docs and troubleshooting match.

3. Codex CLI 

curl -sSL https://stremai.com/install-codex.sh | bash
Enter fullscreen mode Exit fullscreen mode

Or manually in ~/.codex/config.toml:

[mcp_servers.stremai]
url = "https://stremai.com/api/mcp"
Enter fullscreen mode Exit fullscreen mode

4. Verify it actually works

Most setup guides skip the proof step. Don't.

In any connected agent:

store a test memory titled "setup check" with content "shared memory is wired up"

Then open a different tool, or a fresh session, and ask:

recall the test memory called "setup check"

If the second agent returns it without you repeating anything, you have working shared memory. That's the whole product in one round trip: what one connected agent learns, another connected agent can recall.

5. Make it useful

First, tell your agents when to use memory. Put a few lines in your existing CLAUDE.md or AGENTS.md:

- Before answering questions about this codebase, recall relevant project memory.
- After completing a real change, store durable lessons: decisions, pitfalls, gotchas.
- Don't store secrets or raw code; those belong in git.
Enter fullscreen mode Exit fullscreen mode

Second, seed it. Store the five things you're most tired of re-explaining: deploy quirks, flaky tests, architecture decisions, setup gotchas, or the rule everyone keeps rediscovering.

Recall quality is ranked, so the relevant entry surfaces when it matters instead of everything flooding the context window.

What to expect

Memory is user-controlled. Entries are human-readable, attributed to the agent that stored them, and you can export, archive, or erase them from the dashboard.

On teams, one engineer's agent's lesson can be recalled by teammates' connected agents, which is the part that changes onboarding. The point is not that agents magically know everything. The point is that useful working context stops dying inside one tool's session.

StremAI is free to start with no credit card. It is verified with Claude Code and Claude Desktop today, with Cursor, Codex, Windsurf, and OpenClaw supported or in active verification.

Docs: stremai.com/docs/mcp-memory

Try a real recall without an account: stremai.com/playground

Questions or a setup that didn't work? Comment. I read all of them, and broken-setup reports are worth more to me than praise.

Top comments (0)