Connect MemoryCode to Windsurf (MCP)
Windsurf (Codeium) reads MCP servers from a JSON config file. After you add MemoryCode, Cascade can use the same tools as other MCP-aware clients, subject to Windsurf's own limits and startup behavior.
Before you start
- Node.js 18+ and a valid export path for memorycode-mcp.json.
- Patience on first
npxrun: downloading the package can take a minute or two. - If UI differs from this guide, follow Windsurf's MCP documentation.
Config file location
- macOS / Linux: ~/.codeium/windsurf/mcp_config.json
- Windows: %USERPROFILE%\.codeium\windsurf\mcp_config.json
Add the server entry
Merge the MemoryCode block into the MCP servers section Windsurf expects (often mcpServers). Replace the file path with your absolute path:
{
"mcpServers": {
"memorycode": {
"command": "npx",
"args": [
"-y",
"@memorycode/mcp-server",
"--file",
"/absolute/path/to/memorycode-mcp.json"
]
}
}
}Restart and verify
- Quit and reopen Windsurf.
- Check Cascade / MCP settings for server status.
- If you see initialization timeouts, inspect MCP logs and network access to npm; as a next step, consider the local
node+dist/index.jsfallback described in the full manual. - Windsurf may enforce a tool limit across servers. If MemoryCode does not appear, try disabling other MCP servers temporarily.
After setup — profiles & cognitive chips
MemoryCode MCP 1.1 separates profiles (identity base, switched infrequently) from cognitive chips (thinking/output mode, switched frequently in the AI). You do not need to change profiles to switch chips.
Verify tools (8 total)
- Profiles:
list_configs,load_config - Chips (1.1.0+):
list_chips,load_chip,unload_chip,get_cognitive_chip - Identity:
get_user_profile,get_expertise
Daily use (natural language)
- "List my memory profiles" →
list_configs - "Switch to my work profile" →
load_config - "What cognitive chips are available?" →
list_chips - "Switch to Rigorous Analysis chip" →
load_chip(profile must be loaded first)
When to re-export
Re-export memorycode-mcp.json from MemoryCode when you create or edit profiles or chip content. You do not need to re-export when you only switch chips inside the AI (load_chip writes to local mcp.runtime).
MCP 1.0 compatibility
Legacy 1.0 exports still work for load_config. Chip tools will prompt you to re-export an MCP 1.1 file. Website chip selection affects QuickCopy; MCP chip state in mcp.runtime does not sync back to the browser.
Full reference: MCP setup guide — daily use