mem0llama 0.1.2

A specialized fork of Mem0 optimized for small, local Large Language Models (LLMs) running through Ollama.

Mem0llama

A specialized fork of Mem0 optimized for small local Large Language Models (LLMs) running through Ollama.

Overview

Mem0llama enhances the compatibility of Mem0's graph memory system with small local LLMs while preserving its integration with Qdrant (for vector storage/RAG) and Neo4j (for graph relationships). This implementation ensures structured and predictable outputs from LLMs using Ollama's format argument and Pydantic models.

Key Features

• Structured Output Support: Implemented Pydantic models to standardize LLM outputs for reliable entity and relationship extraction
• LLM Formatting Utilities: Created utilities for structured prompts and response parsing with robust error handling
• Ollama Integration: Optimized for local LLMs running through Ollama with format parameter configuration
• Neo4j Community Edition Support: Enhanced compatibility with Neo4j Community Edition
• Preserved Functionality: Maintains all existing Mem0 features including graph memory, search, and retrieval mechanisms

Components

Core Enhancements

1. Pydantic Models (models.py)

   • Standardized models for entities, relationships, and memory operations
   • Ensures consistent data structures for LLM outputs

2. LLM Formatting Utilities (llm_formatter.py)

   • Functions to create structured prompts for various operations
   • Response parsing with comprehensive error handling
   • Ollama format parameter configuration

3. Graph Memory Improvements (graph_memory.py)

   • Modified node retrieval and relationship establishment
   • Enhanced entity and relationship extraction
   • Added LLM configuration adaptation

4. Bug Fixes

   • Fixed "relatationship" typo in multiple files
   • Improved error handling for inconsistent LLM responses

Getting Started

Prerequisites

• Python 3.8+
• Ollama (for local LLMs)
• Neo4j (Community Edition compatible)
• Qdrant (for vector storage)

Installation

Docker servers

qdrant

cd servers/qdrant
./qdrant_run_docker.bat

neo4j

cd servers/neo4j
./neo4j_run_docker.bat

pip install -r requirements.txt

Configuration

Create a .env file with the following variables:

LLM_BASE_URL=http://localhost:11434
LLM_API_KEY=
LLM_MODEL=llama3
EMBEDDER_MODEL=nomic-embed-text
QDRANT_HOST=localhost:6333
QDRANT_API_KEY=your_qdrant_api_key
NEO4J_URL=bolt://localhost:7687
NEO4J_USER=neo4j
NEO4J_PASSWORD=your_password

Basic Usage

from mem0 import Memory

# Initialize Mem0 client
memory_config = {
    "vector_store": {
        "provider": "qdrant",
        "config": {
            "collection_name": "mem0",
            "host": "localhost",
            "port": 6333,
            "embedding_model_dims": 1024,
            "api_key": qdrant_api_key,
        },
    },
    "llm": {
        "provider": "ollama",
        "config": {
            "ollama_base_url": "http://localhost:11434",
            "model": "llama3",
        }
    },
    "embedder": {
        "provider": "ollama",
        "config": {
            "model": "nomic-embed-text",
            "ollama_base_url": "http://localhost:11434",
        },
    },
    "graph_store": {
        "provider": "neo4j",
        "config": {
            "url": "bolt://localhost:7687",
            "username": "neo4j",
            "password": "password",
        },
    },
}

memory = Memory.from_config(memory_config)

# Add memories
messages = [
    {"role": "user", "content": "I like to program in Python"},
    {"role": "assistant", "content": "That's great! Python is a versatile language."}
]
memory.add(messages, user_id="user123")

# Search memories
results = memory.search(query="What programming languages do I know?", user_id="user123")

Testing

Run the test script to validate the structured output capabilities:

python ./mem0llama_structured_output.py

For a more interactive experience, try the terminal chatbot:

python ./mem0llama_chat_cli.py

Add the --debug flag to see detailed memory and relation information:

python ./mem0llama_chat_cli.py --debug

License

This project is licensed under the same terms as the original Mem0 project.

Acknowledgements

• Mem0 - The original memory layer for AI agents
• Ollama - For running local LLMs
• Neo4j - Graph database for relationship storage
• Qdrant - Vector database for RAG