Eml — AI skill for Claude Code

EML — Engram Protocol

**A typed memory for AI agents, on your own machine.** EML (Epistemic Memory Layer) stores what an agent learns while working with you as typed *engrams*: decisions, rules, tasks, status, notes, references, events, logs and profiles. They are linked in a graph, carry their provenance and have a lifecycle (active, superseded, resolved). An agent loads its identity, rules and open work at the start of a session, recalls what it needs, and writes back what it decided or learned.

This repository is the local product: the **EML Cockpit**, a web app that runs isolated Claude Code sessions next to your brain, with everything it is built from.

Quick start

You need macOS or Linux, Node ≥ 22 and pnpm 10 (`corepack enable`).

git clone https://github.com/EpistemicLabs/eml.git && cd eml
pnpm install
pnpm run cockpit

This builds everything, starts the cockpit on `127.0.0.1` and opens it in your browser. A fresh brain is created under `~/.eml-cockpit/`.

To run a release as an app that updates itself (login service, update banner), use `scripts/install-app.sh latest` instead. Details, updates and the development setup are in [`cockpit/README.md`](cockpit/README.md).

What's in here

Path	What it is
cockpit/host	The eml-cockpit package: the local host. It runs Claude Code sessions in PTYs, launches the brain server and serves the UI.
cockpit/frontend	The cockpit UI (Vite + React): desks with terminal and brain panes, sidebar, brain dashboard, graph and search.
packages/core	The engram store: SQLite schema and migrations, typed validation, the graph, recall, boot.
packages/server	The HTTP API over core, one SQLite file per brain, with token routing.
packages/mcp-server	@engramprotocol/mcp-server: the brain as MCP tools for Claude Code, plus the session hooks and skills.

Development

pnpm build          # TypeScript project build (core, server, mcp-server)
pnpm test