Documentation

Everything you need to get started with AutoMem

Architecture Overview

🗄️ Storage Layer

FalkorDB: Graph database for memory relationships and knowledge graphs

Qdrant: Vector database for semantic similarity search with OpenAI embeddings

🔌 Integration Layer

MCP Server: Model Context Protocol server exposing 6 tools (store, recall, update, delete, associate, health check)

🤖 Client Support

  • Claude Desktop (Anthropic)
  • Cursor IDE
  • GitHub Copilot (via MCP)
  • Any future MCP-compatible LLM

☁️ Deployment Options

  • Railway: One-click deploy (~$5/month)
  • Docker: Self-hosted with docker-compose
  • Local: Python development server

Key Concepts

Memory Types

AutoMem categorizes memories into 7 semantic types for better organization and retrieval:

Decision Pattern Insight Preference Style Habit Context

Importance Scoring

Memories are scored 0.0-1.0 based on relevance. Importance automatically decays over time unless reinforced, mimicking human memory.

Knowledge Graphs

Memories can be linked with 11 relationship types:

RELATES_TO, LEADS_TO, EVOLVED_INTO, DERIVED_FROM, EXEMPLIFIES, CONTRADICTS, REINFORCES, INVALIDATED_BY, OCCURRED_BEFORE, PART_OF, PREFERS_OVER

Automatic Consolidation

Background scheduler automatically:

  • Decay: Reduces importance of old, unused memories (hourly)
  • Creative clustering: Groups related memories (every 6 hours)
  • Forgetting: Removes low-importance memories (daily)