Guides

Persistent memory for Claude and Cursor

The repository ships an MCP server so Cursor, Claude and Codex can call memory_remember and memory_recall on a brain directory on your machine. These steps are the ones on the MCP integration page, in order. Claude and Cursor were not attached, and a server process was not left listening.

Install the MCP extra

Install with pip install "fluctlightdb[native,mcp]". The repository documents that line for Cursor, Claude and Codex. The GitHub release v0.5.22, published 1 October 2026, says the extra pins mcp below 2 so the server module can import FastMCP. PyPI published fluctlightdb 0.5.22 the same day, and its metadata lists that pin. This site does not name a version to install.

Unpinned. The v0.5.22 note says the mcp pin is in that tag.

pip install "fluctlightdb[native,mcp]"

Point Cursor at the server

The integration page shows the Cursor and Codex template: command python3, args ["-m", "fluctlightdb.mcp_server"], and env FLUCTLIGHT_AGENT set to the assistant’s name. Codex uses the same shape. That template was not started as a listening process in the check this site ran.

Shape of the Cursor and Codex template, from the MCP page

{
  "mcpServers": {
    "fluctlight": {
      "command": "python3",
      "args": ["-m", "fluctlightdb.mcp_server"],
      "env": { "FLUCTLIGHT_AGENT": "cursor" }
    }
  }
}

Point Claude at the same server

Claude gets the same server merged into its settings, with FLUCTLIGHT_AGENT set to its own name. The command and arguments are the same python3 module. This site did not attach Claude or Cursor and did not record a session in either one.

Remember, then recall from a new process

On a re-run from main, memory_remember stored a sentence and called checkpoint(), and memory_recall in a second process printed that sentence. The v0.5.22 release note says that checkpoint shipped. Offline, a cue has to share tokens with what was stored unless you bring an embedder. Nothing is durable until checkpoint() returns.