Mem-LLM 🧠💻

Mem-LLM is a privacy-first, local Python framework for building memory-enabled AI assistants. By running entirely on your local machine, it combines persistent multi-user conversation history with configurable knowledge bases, robust storage backends, and seamless multi-model support.

Perfect for privacy-first, production-ready workflows!

🔗 Quick Links

• GitHub Repository: github.com/emredeveloper/Mem-LLM
• PyPI Package: pypi.org/project/mem-llm/
• Issue Tracker: Report a Bug or Request a Feature

What's New in v2.5.1

Dependency fixes. Graph memory needs pydantic and the API server's upload endpoint needs python-multipart; neither was declared, so on a clean install graph memory silently degraded to a no-op and pip install mem-llm[api] could not import the API server. Both are now declared, so a plain pip install mem-llm gets working graph memory.

What's New in v2.5.0

This release expands local backend support and upgrades long-term memory:

• OpenAI-Compatible Backends: Use any /v1/chat/completions compatible server.
• llama.cpp Support: Connect directly to llama-server with backend="llamacpp".
• MemoryRouter: Unified core memory, archival memory, recall, graph context, and KB retrieval.
• Temporal Graph Memory: Track current facts and historical facts with validity windows.
• Cleanup: Reduced duplicated backend alias and chat context-building logic.

✨ Core Highlights

• Persistent Multi-User Memory: Keep context across sessions. Supports lightweight JSON or robust SQLite databases.
• Advanced Tool Calling: Endow your agent with superpowers! Use built-in tools or easily create your own with the @tool decorator.
• Multi-Backend Flexible Support: Switch between Ollama, LM Studio, OpenAI-compatible APIs, and llama.cpp.
• Long-Term Memory Routing: Combine core memory, archival memory, recall, knowledge base, and graph context.
• Temporal Graph Memory: Preserve changing facts without losing history.
• Knowledge Base (RAG) & Vector Stores: Empower your agent with your own documents and databases organically.
• Conversation Analytics: Track interactions, model performance, and agent behavior systematically.
• REST API + Web UI: Deploy your local agent instantly with the built-in, ready-to-use API server and slick web interface.
• Real-Time Streaming: Stream chat responses chunk by chunk for ultra-low latency experiences.

📦 Installation

Get up and running in seconds.

pip install mem-llm

Optional Power-ups:

# Add API server dependencies (FastAPI, Uvicorn)
pip install mem-llm[api]

# Add advanced database support
pip install mem-llm[databases]

⚡ Quick Start

Using Ollama 🦙

Make sure your Ollama instance is running, then try this simple example:

from mem_llm import MemAgent

# Initialize the agent
agent = MemAgent(backend="ollama", model="granite4:3b")

# Set the active user (memory will be uniquely tied to this ID)
agent.set_user("alice")

# Chat and watch it remember!
print(agent.chat("Hi! My name is Alice and I am a Software Engineer."))
print(agent.chat("What was my name and what do I do?")) 

Using LM Studio 🛠️

Ensure LM Studio's local server is running on its default port:

from mem_llm import MemAgent

agent = MemAgent(backend="lmstudio", model="qwen3.5-2b")
agent.set_user("bob")

print(agent.chat("Explain Python memory management in 2 sentences."))

Using llama.cpp

Start llama-server with an OpenAI-compatible endpoint:

llama-server.exe -m C:\path\to\model.gguf --alias local-model --host 127.0.0.1 --port 8080

Connect Mem-LLM:

from mem_llm import MemAgent

agent = MemAgent(
    backend="llamacpp",
    model="local-model",
    base_url="http://localhost:8080",
)
agent.set_user("carol")

print(agent.chat("Remember that I prefer concise answers."))

📄 License

Mem-LLM is proudly open-source and released under the MIT License. Build away!