Auto-Claude Memory System

Graphiti-based persistent memory for cross-session context retention.

Overview

Auto-Claude uses Graphiti with embedded LadybugDB for memory:

• No Docker required - Embedded graph database
• Multi-provider support - OpenAI, Anthropic, Ollama, Google AI, Azure
• Semantic search - Find relevant context across sessions
• Knowledge graph - Entity relationships and facts

Architecture

Agent Session
     │
     ▼
Memory Manager
     │
     ├──▶ Add Episode (new learnings)
     ├──▶ Search Nodes (find entities)
     ├──▶ Search Facts (find relationships)
     └──▶ Get Context (relevant memories)
     │
     ▼
Graphiti (Knowledge Graph)
     │
     ▼
LadybugDB (Embedded Storage)

Configuration

Enable Memory System

In apps/backend/.env:

# Enable Graphiti memory (default: true)
GRAPHITI_ENABLED=true

Provider Selection

Choose LLM and embedding providers:

# LLM provider: openai | anthropic | azure_openai | ollama | google | openrouter
GRAPHITI_LLM_PROVIDER=openai

# Embedder provider: openai | voyage | azure_openai | ollama | google | openrouter
GRAPHITI_EMBEDDER_PROVIDER=openai

Provider Configurations

OpenAI (Simplest)

GRAPHITI_ENABLED=true
GRAPHITI_LLM_PROVIDER=openai
GRAPHITI_EMBEDDER_PROVIDER=openai
OPENAI_API_KEY=sk-xxxxxxxxxxxxxxxx
OPENAI_MODEL=gpt-4o-mini
OPENAI_EMBEDDING_MODEL=text-embedding-3-small

Anthropic + Voyage (High Quality)

GRAPHITI_ENABLED=true
GRAPHITI_LLM_PROVIDER=anthropic
GRAPHITI_EMBEDDER_PROVIDER=voyage
ANTHROPIC_API_KEY=sk-ant-xxxxxxxx
GRAPHITI_ANTHROPIC_MODEL=claude-sonnet-4-5-latest
VOYAGE_API_KEY=pa-xxxxxxxx
VOYAGE_EMBEDDING_MODEL=voyage-3

Ollama (Fully Offline)

GRAPHITI_ENABLED=true
GRAPHITI_LLM_PROVIDER=ollama
GRAPHITI_EMBEDDER_PROVIDER=ollama
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_LLM_MODEL=deepseek-r1:7b
OLLAMA_EMBEDDING_MODEL=nomic-embed-text
OLLAMA_EMBEDDING_DIM=768

Prerequisites:

# Install Ollama
curl -fsSL https://ollama.ai/install.sh | sh

# Pull models
ollama pull deepseek-r1:7b
ollama pull nomic-embed-text

Google AI (Gemini)

GRAPHITI_ENABLED=true
GRAPHITI_LLM_PROVIDER=google
GRAPHITI_EMBEDDER_PROVIDER=google
GOOGLE_API_KEY=AIzaSyxxxxxxxx
GOOGLE_LLM_MODEL=gemini-2.0-flash
GOOGLE_EMBEDDING_MODEL=text-embedding-004

Azure OpenAI (Enterprise)

GRAPHITI_ENABLED=true
GRAPHITI_LLM_PROVIDER=azure_openai
GRAPHITI_EMBEDDER_PROVIDER=azure_openai
AZURE_OPENAI_API_KEY=xxxxxxxx
AZURE_OPENAI_BASE_URL=https://your-resource.openai.azure.com/...
AZURE_OPENAI_LLM_DEPLOYMENT=gpt-4
AZURE_OPENAI_EMBEDDING_DEPLOYMENT=text-embedding-3-small

OpenRouter (Multi-Provider)

GRAPHITI_ENABLED=true
GRAPHITI_LLM_PROVIDER=openrouter
GRAPHITI_EMBEDDER_PROVIDER=openrouter
OPENROUTER_API_KEY=sk-or-xxxxxxxx
OPENROUTER_LLM_MODEL=anthropic/claude-3.5-sonnet
OPENROUTER_EMBEDDING_MODEL=openai/text-embedding-3-small

Database Settings

# Database name (default: auto_claude_memory)
GRAPHITI_DATABASE=auto_claude_memory

# Storage path (default: ~/.auto-claude/memories)
GRAPHITI_DB_PATH=~/.auto-claude/memories

Memory Operations

How Memory Works

1. During Build

   • Agent discovers patterns, gotchas, solutions
   • Memory Manager extracts insights
   • Insights stored as episodes in knowledge graph

2. New Session

   • Agent queries for relevant context
   • Memory returns related insights
   • Agent builds on previous learnings

MCP Tools

When GRAPHITI_MCP_URL is set, agents can use:

Python API

from integrations.graphiti.memory import get_graphiti_memory

# Get memory instance
memory = get_graphiti_memory(spec_dir, project_dir)

# Get context for session
context = memory.get_context_for_session("Implementing feature X")

# Add insight from session
memory.add_session_insight("Pattern: use React hooks for state")

# Search for relevant memories
results = memory.search("authentication patterns")

Memory Storage

Location

~/.auto-claude/memories/
├── auto_claude_memory/     # Main database
│   ├── nodes/              # Entity nodes
│   ├── edges/              # Relationships
│   └── episodes/           # Session insights
└── embeddings/             # Vector embeddings

Per-Spec Memory

.auto-claude/specs/001-feature/
└── graphiti/               # Spec-specific memory
    ├── insights.json       # Extracted insights
    └── context.json        # Session context

Querying Memory

Command Line

cd apps/backend

# Query memory
python query_memory.py --search "authentication"

# List recent episodes
python query_memory.py --recent 10

# Get entity details
python query_memory.py --entity "UserService"

Memory in Action

Example session:

Session 1:
  Agent: "Implemented OAuth login, discovered need to handle token refresh"
  Memory: Stores insight about token refresh pattern

Session 2:
  Agent: "Implementing user profile..."
  Memory: "Previously learned about token refresh in OAuth implementation"
  Agent: Uses learned pattern for profile API calls

Best Practices

Effective Memory Use

1. Let agents learn naturally

   • Don't force memory storage
   • Agents automatically extract insights

2. Use semantic search

   • Query with natural language
   • Memory finds related concepts

3. Clean up periodically

   • Remove outdated insights
   • Update incorrect information

Provider Selection

Performance Tips

1. Embedding model selection

   • text-embedding-3-small: Fast, good quality
   • text-embedding-3-large: Better quality, slower

2. LLM model selection

   • gpt-4o-mini: Fast, cost-effective
   • claude-sonnet: High quality reasoning

3. Ollama optimization

   # Use smaller models for speed
   OLLAMA_LLM_MODEL=llama3.2:3b
   OLLAMA_EMBEDDING_MODEL=all-minilm
   OLLAMA_EMBEDDING_DIM=384
   

Troubleshooting

Memory Not Working

# Check if enabled
grep GRAPHITI apps/backend/.env

# Verify provider credentials
python -c "from integrations.graphiti.memory import get_graphiti_memory; print('OK')"

Provider Errors

# OpenAI
curl -H "Authorization: Bearer $OPENAI_API_KEY" https://api.openai.com/v1/models

# Ollama
curl http://localhost:11434/api/tags

# Check logs
DEBUG=true python query_memory.py --search "test"

Database Corruption

# Backup and reset
mv ~/.auto-claude/memories ~/.auto-claude/memories.backup
python query_memory.py --search "test"  # Creates fresh DB

Embedding Dimension Mismatch

If changing embedding models:

# Clear existing embeddings
rm -rf ~/.auto-claude/memories/embeddings

# Restart to re-embed
python run.py --spec 001

Advanced Usage

Custom Memory Integration

from integrations.graphiti.queries_pkg.graphiti import GraphitiMemory

# Create custom memory instance
memory = GraphitiMemory(
    database="custom_db",
    db_path="/path/to/storage",
    llm_provider="anthropic",
    embedder_provider="voyage"
)

# Custom operations
memory.add_entity("UserService", {"type": "service", "purpose": "auth"})
memory.add_relationship("UserService", "uses", "Database")

Memory MCP Server

Run standalone memory server:

# Start Graphiti MCP server
GRAPHITI_MCP_URL=http://localhost:8000/mcp/ python -m integrations.graphiti.server

Related Skills

• auto-claude-setup: Initial configuration
• auto-claude-optimization: Performance tuning
• auto-claude-troubleshooting: Debugging