Share one memory across Claude Code, Cursor, Codex, and Gemini CLI

If you use more than one AI coding tool — Claude Code for the heavy lifting, Cursor in the editor, maybe Codex or Gemini CLI on the side — you have several memory silos. Claude's built-in memory stays with Claude. Cursor's rules stay with Cursor. Explain a constraint to one agent, and the next one violates it an hour later.

The fix is architectural, not heroic: keep memory outside the agents, in a neutral store they all read. linksee-memory does this with one SQLite file at ~/.linksee-memory/memory.db, exposed to every agent through the Model Context Protocol. Declare once, recall anywhere.

How it works

Each client is configured to spawn the same MCP server — npx -y linksee-memory — over stdio. The server owns the SQLite file; the agents just call remember / recall / read_smart tools. Because the store is a single local file, there is no sync service, no account, and no network call involved.

Setup per client

The interactive way is npx -y linksee-memory setup, which detects installed clients. Manual configs:

Claude Code

claude mcp add -s user linksee -- npx -y linksee-memory

Cursor

Add to ~/.cursor/mcp.json, then restart Cursor:

{
  "mcpServers": {
    "linksee": {
      "command": "npx",
      "args": ["-y", "linksee-memory"]
    }
  }
}

OpenAI Codex

codex mcp add linksee -- npx -y linksee-memory

Or add to ~/.codex/config.toml:

[mcp_servers.linksee]
command = "npx"
args = ["-y", "linksee-memory"]

Gemini CLI

Add the same mcpServers block to ~/.gemini/settings.json.

Claude Desktop / Windsurf

Same JSON block in claude_desktop_config.json (macOS ~/Library/Application Support/Claude/, Windows %APPDATA%\Claude\) or Windsurf's MCP settings.

Try the round-trip

1. In Claude Code: “linksee this — we use FTS5, not vector search, because our corpus is small and we want zero infra.”
2. In Cursor (same machine, any time later): “use linksee — what did we decide about search?”

The second agent answers with the decision and the reasoning, because memories carry six layers (goal / context / emotion / implementation / caveat / learning), not just flat facts.

What about drift between agents?

Shared memory creates a new possibility: the store knows what you decided, so it can notice when any agent's work quietly contradicts it. Declare decisions as anchors, and ask what's drifting? from any client to reconcile reality against them. That layer is what separates a shared notebook from a shared brain — see the persistent-memory guide for the full setup.

Limits, honestly

• One machine. The free core shares memory between agents on the same computer. Cross-device sync is planned for the paid tier.
• Consumer ChatGPT apps can't connect — they don't support stdio MCP servers. Codex CLI works.
• Concurrent writes are serialized by SQLite. Fine for interactive use; not designed as a team-wide server (yet).