cursor¶
connect Cursor to your existing Engram store over local MCP stdio. Cursor starts the Python process and uses its tools from chat; no separate HTTP service is needed. your Cursor model choice is independent of Engram's model settings.
complete installation and the quick start
first. keep the absolute Python executable and config path printed by engram init.
choose where to register it¶
use ~/.cursor/mcp.json for your personal configuration across projects, or
.cursor/mcp.json inside one project for project-specific availability.
these locations and the stdio fields are documented in
Cursor's MCP reference.
a project entry controls where the tools appear. it does not isolate memories: clients using the same Engram config connect to the same store. keep personal paths and credentials out of a shared project configuration.
add the server¶
merge this entry into the existing mcpServers object, replacing both paths:
{
"mcpServers": {
"engram": {
"type": "stdio",
"command": "/absolute/path/to/venv/bin/python",
"args": [
"-m", "engram",
"--config", "/absolute/path/to/config.yaml",
"serve", "--mcp"
]
}
}
}
command is the executable, not a whole shell command. keep arguments as
separate array elements, including paths containing spaces. on Windows, use
your environment's Scripts\\python.exe and escape backslashes in JSON.
environment and paths¶
an editor launched from the desktop may not inherit your terminal's PATH or
activated virtual environment. the absolute executable avoids that ambiguity.
use the same explicit --config path for Cursor and terminal checks.
if your Engram setup needs environment credentials, add an env object inside
the engram entry. Cursor supports environment interpolation, for example:
this is an optional fragment, not another server. the variable must exist in Cursor's environment; omit it for a local SQLite setup. see config interpolation.
check the connection¶
open Customize in Cursor, find engram, and enable it. its tools should
appear under Available Tools. ask in chat:
use Engram's
statustool, thenconfig_show, and show which store is connected.
inspect the actual tool result. config_show reports effective settings with
secrets redacted; a chat response without a tool call is not a connection check.
these tool-discovery controls are described in
Cursor's chat integration.
if it does not connect¶
- restart Cursor after changing the command or updating Engram.
- open Output → MCP Logs for process and connection errors.
- check that the selected Python can import Engram and read the config file.
- for missing tools, check that the server and required tools are enabled.
run the local check with the exact executable from the entry:
the doctor reference explains model downloads and
isolated checks. doctor tests the local endpoint; repeat the Cursor tool call
to verify the client connection. if the server works but recall is empty, use
recall_explain to inspect retrieval gates.
source checked: official Cursor MCP documentation, 20 september 2026. configuration reviewed against the docs; no Cursor session is claimed as tested by this guide.