Engram

Self-curating coding preference memory for Claude Code.

Engram learns your coding preferences passively during Claude Code sessions, stores them with semantic deduplication, and automatically injects relevant preferences into future sessions — scoped to the language, framework, and repository you're working in.

Quick Start

# Install dependencies
just setup

# Start development server
just dev

# Open the web UI
open http://localhost:3777

How It Works

1. During coding sessions — when you give feedback like "don't mock the database" or "use frozen dataclasses", Claude Code stores it via the engram MCP server
2. Between sessions — engram deduplicates, resolves conflicts, and organizes preferences using semantic search and LLM-driven analysis
3. At session start — a Claude Code hook injects relevant preferences into your CLAUDE.md, scoped to the project's languages and repo

Setup

Prerequisites

• Python 3.11+
• uv
• just
• An Anthropic API key (set as ANTHROPIC_API_KEY environment variable)

Installation

git clone https://github.com/DannyMor/engram.git
cd engram
just setup

# Install as background service + Claude Code hook
just install

Claude Code MCP Configuration

Add to your Claude Code settings:

{
  "mcpServers": {
    "engram": {
      "url": "http://localhost:3777/mcp"
    }
  }
}

Commands

Command	Purpose
just dev	Run in foreground with auto-reload
just start / just stop	Manage background service
just status	Check if service is running
just logs	Tail service logs
just install / just uninstall	Install/remove service + hook
just check	Run lint + typecheck + tests

Configuration

Configuration lives at ~/.engram/config.yaml. Editable via the web UI settings page or directly.

llm:
  provider: anthropic
  model: claude-sonnet-4-6
  api_key_env: ANTHROPIC_API_KEY

embedder:
  provider: fastembed
  model: BAAI/bge-small-en-v1.5

storage:
  path: ~/.engram/data

Web UI

Open http://localhost:3777 to:

• Browse preferences with search, scope filtering, and tag filtering
• Chat with the curation agent for bulk cleanup, conflict review, and proactive suggestions
• Configure LLM provider and model settings

Architecture

Single Python process serving:

• MCP endpoint at /mcp (HTTP transport for Claude Code)
• REST API at /api/* (backend for web UI and injection hook)
• Static web UI at / (Alpine.js + Tailwind CSS)

Memory backend: Mem0 with embedded Qdrant (local vector storage) and fastembed (local embeddings).

Development

just dev        # Start dev server with auto-reload
just test       # Run tests
just lint       # Run ruff
just typecheck  # Run pyright
just check      # All three