engramdocs

Using Engram in your editor

Engram speaks the Model Context Protocol (MCP), the open standard for connecting tools to AI coding assistants. That means one server plugs into most AI coding CLIs — the same memory, in every terminal you work in.

The server is always the same command: engram mcp. What changes between tools is only where you register it. Whatever the tool, make sure you have installed the engram engine first (see Getting started) so the command is on your PATH.

What the MCP server exposes

  • A recall tool, engram_ask — index-first recall over your store. Your assistant can call it on its own to pull the relevant pages mid-conversation, each surfaced with a sealed record that it was used. It is the same recall path as the engram ask command; there is only one, so a query behaves identically however it arrives.

That single tool is the universal adapter. Any MCP-capable assistant that can reach it gets your memory.

Claude Code (the plugin)

Claude Code has a first-class plugin that wires more than just recall — it also adds the automatic session hooks (surface memory at session start, distill the conversation at session end) and a /engram slash command.

1. Add and install the plugin — inside Claude Code:

/plugin marketplace add ad-astra26/engram
/plugin install engram

2. Make it operational — in a normal terminal. The bootstrap installs the engine on your PATH and self-tests:

# macOS / Linux
bash setup.sh
# Windows (PowerShell)
powershell -ExecutionPolicy Bypass -File setup.ps1

Then run /reload-plugins back in Claude Code.

Once installed you get:

  • engram_ask — the recall tool Claude can call itself.
  • /engram <query> — a slash-command shortcut for on-demand recall.
  • Session hooks — relevant pages surfaced automatically at the start of a session; the conversation distilled into a new page at the end. Every hook fails quietly, so a missing or broken Engram never blocks your session.

If Claude Code shows the Engram server as failed before you have run the bootstrap, that is the expected signal that the engine is not on your PATH yet — not a bug. Finish step 2 and reload.

Everything else (the same MCP server)

For any other MCP-capable assistant, register the exact same server — command: engram, args: ["mcp"]. Here are the common ones.

Codex

In ~/.codex/config.toml:

[mcp_servers.engram]
command = "engram"
args = ["mcp"]

Gemini CLI

In your MCP servers config:

{ "mcpServers": { "engram": { "command": "engram", "args": ["mcp"] } } }

Cursor

In .cursor/mcp.json:

{ "mcpServers": { "engram": { "command": "engram", "args": ["mcp"] } } }

Cline, Windsurf, Zed, Continue, and other MCP hosts

These all follow the same pattern — register a server with:

  • command: engram
  • args: ["mcp"]

Consult your tool's MCP configuration docs for exactly where its server list lives, then drop that server in. Because they all talk to the one engram mcp adapter, they all share the same store and the same memory.

One store, many tools

Every editor above connects to the same local store. Learn something in Cursor today, recall it in Claude Code tomorrow — it is one memory, following you across tools and projects. And it never leaves your machine to do it.

Next: the FAQ answers the questions this raises — privacy, API keys, what "sealed" actually proves.

← all docs · engram home