Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/xantorres/engram/llms.txt

Use this file to discover all available pages before exploring further.

Engram ships a built-in MCP server built on FastMCP that runs over standard input/output. Any MCP-capable agent — Claude Code, Codex, opencode, or a custom client — can connect to it with a single config entry and immediately start reading from and writing to your local memory store.

​
Starting the server

The server can be started in two equivalent ways: 
# Dedicated entry point (recommended for MCP config files)
engram-mcp

# Alias via the main CLI
engram serve
Both commands block and communicate over stdio. There is no TCP port, no authentication header, and no cloud endpoint — the process speaks the MCP stdio transport protocol directly with its parent client.
Run engram init <harness> to print a ready-to-paste config snippet for your specific client (Claude Code, Codex, or opencode). See the Wiring Agents guide for full walkthrough instructions.

​
What the server exposes

The Engram MCP server registers two tools and one resource:
NameTypePurpose
rememberToolStage a durable fact about the user into memory (pending review)
recallToolReturn promoted memories, optionally filtered by a query string
memory://recallResourceCompact memory context block for clients to auto-load each session
Detailed documentation for each surface lives on the following pages:

Tools

Full parameter docs for remember and recall.

Resources

The memory://recall URI and rendered block format.

​
Why review and promotion are CLI-only

Agents can write new facts via remember, but they cannot approve or promote those facts into the durable store. Promotion, rejection, and forgetting are operations that only the human operator can perform through the CLI (engram review, engram approve, engram reject).
This is a deliberate security boundary. An agent that can both write and approve memories could silently overwrite sensitive information — fiscal details, health data, real-name identity — without your knowledge. The CLI is the human gate; the MCP server is the agent gate.
Facts written through remember land in a pending queue with a status of pending. They become visible to recall only after a human promotes them. Low-risk kinds (preference, tooling, project, infra) may be auto-promoted if autopromote = true is set in your config; sensitive kinds (identity, fiscal, people, constraint, location, health) always require explicit CLI review regardless of that setting.

​
Integration snippets

engram init <harness> prints the correct config block for your client. The three supported harnesses are shown below. 
// Place in your project root as .mcp.json, or in ~/.claude.json for global scope
{
  "mcpServers": {
    "engram": { "command": "engram-mcp" }
  }
}
All three harnesses discover engram-mcp on your PATH. Make sure the package is installed — which engram-mcp should return a path — before adding the config entry.

Build docs developers (and LLMs) love

Get started for freeTalk to us