Engrm — Docs

Quick Start (5 minutes)

1

Sign up at engrm.dev

Create your free account at engrm.dev. No credit card required.

2

Install Engrm

Run the init command to connect this machine to your account:

npx engrm init

This opens your browser to authenticate, registers the device, and configures your supported agent hooks automatically.

3

Verify the connection

Check that everything is working:

npx engrm status

You should see your account email, sync status, and observation count.

4

Start coding with Engrm

That's it. Engrm captures observations automatically as you work in Claude Code, Codex, or any supported MCP agent. Discoveries, decisions, bugfixes, and patterns are saved and synced across all your devices.

How It Works

Engrm is a memory layer for AI coding agents. During your Claude Code or Codex sessions, it automatically captures observations — discoveries, decisions, bugfixes, and patterns — and stores them locally in SQLite.

Observations sync across all your devices via Candengo Vector, a hybrid search backend that combines full-text and semantic search. When you start a new session, Engrm injects relevant context from past sessions so your AI assistant remembers what matters.

Everything works offline-first. Your local database is the source of truth, and sync happens in the background when connectivity is available.

CLI Commands

engrm init	Set up Engrm on this machine. Opens browser for authentication, registers the device, and installs the right hooks for your agent.
engrm init --token=cmt_xxx	Set up from a provisioning token (shown after signup at engrm.dev).
engrm init --url=https://...	Connect to a self-hosted Candengo Vector instance.
engrm init --manual	Interactive manual setup for air-gapped environments.
engrm status	Show account, plan, sync health, observation counts, Sentinel status, and security findings.
engrm packs	List available starter packs for your tech stack.
engrm install-pack <name>	Install a starter pack (typescript-patterns, node-security, react-gotchas, python-django, etc.).
engrm sentinel rules	List available Sentinel rule packs (security, auth, api, react, database).
engrm sentinel init-rules	Install all Sentinel rule packs for real-time code audit.

MCP Tools

These tools are available to supported agents during your coding sessions. They call them automatically when relevant.

search	Find relevant observations from memory. Combines local FTS5 + sqlite-vec with remote Candengo Vector semantic search.
timeline	Get chronological context around an observation — what happened before and after.
get_observations	Fetch full details of observations by ID.
save_observation	Manually save something worth remembering — a decision, discovery, or pattern.
install_pack	Load a curated knowledge pack for your detected tech stack.
pin_observation	Pin an observation so it's always included in session context and never ages out.
check_messages	Check for unread team messages sent through Engrm.
session_context	Get the current session's context — what memory was injected at startup.

Sentinel — Real-Time Code Audit

Sentinel reviews every Edit and Write operation before it executes, checking for security issues, anti-patterns, and drift from team decisions.

How it works: When Claude proposes a code change, Sentinel sends the change to an LLM endpoint (your own or platform-provided) for review. The verdict is one of: PASS, WARN, BLOCK, or DRIFT (contradicts a previous decision).

Modes: In advisory mode (Vibe/Pro plans), warnings are shown but changes are allowed. In blocking mode (Team plan), flagged changes are prevented from executing.

Rule packs: 5 built-in packs cover common patterns — security (SQL injection, XSS, hardcoded secrets), auth, api, react, and database.

npx engrm sentinel init-rules

Agent Hooks

Engrm registers Claude Code lifecycle hooks and Codex startup/stop integration automatically after engrm init. Hook coverage varies a little by agent, but the setup flow is the same.

SessionStart	Injects relevant project memory into the agent's context at session start.
PreToolUse	Sentinel audits Edit/Write operations before they execute.
PostToolUse	Extracts observations from tool results (the main capture path on Claude Code).
PreCompact	Re-injects memory before conversation compression so context survives.
ElicitationResult	Captures user responses to MCP server form dialogs.
Stop	Generates session digest, flushes sync queue, sends telemetry beacon.

Troubleshooting

"engrm: command not found"

Use npx engrm to run without a global install, or install globally with npm install -g engrm.

"Connection failed"

Check that your ENGRM_TOKEN environment variable is set correctly. Verify internet connectivity. Run npx engrm status for diagnostics.

"No observations"

Make sure your agent hooks are registered. Run npx engrm init again to reinstall hooks. Observations are captured automatically during supported agent sessions — try starting a new session and making a few changes.