Agents and editors

engram is a standard MCP server that runs over stdio:

{
  "mcpServers": {
    "engram": { "command": "npx", "args": ["-y", "@vefimov/engram", "serve"] }
  }
}

The server finds the project by walking up from its working directory to the nearest .engram/, or to the git root. Override this with --root <dir> or the ENGRAM_ROOT variable.

Claude Code

engram init --agent claude

This writes .mcp.json, copies the skills into .claude/skills/ and adds an engram section to CLAUDE.md. To register the server by hand instead:

claude mcp add engram -- npx -y @vefimov/engram serve

The skills become slash commands: /recall, /why-we-did-this, /remember-decision, /avoid-past-mistakes, /project-summary, /session-handoff. To install them for every project, run engram skills install --global.

Cursor

engram init --agent cursor

This writes .cursor/mcp.json and an always-applied rule, .cursor/rules/engram.mdc. The skills are available as MCP prompts.

Codex CLI, Windsurf, Zed and others

init has no setup for these. Each client keeps its MCP servers in its own configuration, so add the same command and args as in the JSON above, using the format your client documents. Then put the instructions in AGENTS.md:

engram init --agent agents-md

A client without a skills system still gets every skill as an MCP prompt with the same name.

Several agents, one memory

Every client points at the same .engram/ directory, so Claude Code can record a decision that Cursor recalls an hour later. SQLite runs in WAL mode, which allows several servers to use the same database at once.