Mem-LLM is a Python framework for building privacy-first, memory-enabled AI assistants that run 100% locally. The project combines persistent multi-user conversation history with optional knowledge bases, multiple storage backends, vector search capabilities, response quality metrics, and tight integration with Ollama and LM Studio so you can experiment locally and deploy production-ready workflows with quality monitoring and semantic understanding - completely private and offline.
- ✅ Collaborative AI Agents - Multiple specialized agents working together
- ✅ BaseAgent - Role-based agents (Researcher, Analyst, Writer, Validator, Coordinator)
- ✅ AgentRegistry - Centralized agent management and health monitoring
- ✅ CommunicationHub - Thread-safe inter-agent messaging and broadcast channels
- ✅ 29 New Tests - Comprehensive test coverage (84-98%)
- ✅ Deep Insights - Analyze user engagement, topics, and activity patterns
- ✅ Visual Reports - Export analytics to JSON, CSV, or Markdown
- ✅ Engagement Tracking - Monitor active days, session length, and interaction frequency
- ✅ Instant Setup - Initialize specialized agents with one line of code
- ✅ 8 Built-in Presets -
chatbot,code_assistant,creative_writer,tutor,analyst,translator,summarizer,researcher - ✅ Custom Presets - Save and reuse your own agent configurations
-
✅ Smart parser - Understands natural language tool calls
-
✅ Better prompts - Clear DO/DON'T examples for LLM
-
✅ More reliable - Tools execute even when LLM doesn't follow exact format
-
Function Calling (v2.0.0) – LLMs can call external Python functions
-
Memory-Aware Tools (v2.0.0) – Agents search their own conversation history
-
18+ Built-in Tools (v2.0.0) – Math, text, file, utility, memory, and async tools
-
Custom Tools (v2.0.0) – Easy
@tooldecorator for your functions -
Tool Chaining (v2.0.0) – Automatic multi-tool workflows
- 100% Local & Private (v1.3.6) – No cloud dependencies, all processing on your machine.
- Streaming Response (v1.3.3+) – Real-time ChatGPT-style streaming for Ollama and LM Studio.
- REST API Server (v1.3.3+) – FastAPI-based server with WebSocket and SSE streaming support.
- Web UI (v1.3.3+) – Modern 3-page interface (Chat, Memory Management, Metrics Dashboard).
- Persistent Memory – Store and recall conversation history across sessions for each user.
- Multi-Backend Support (v1.3.0+) – Choose between Ollama and LM Studio with unified API.
- Auto-Detection (v1.3.0+) – Automatically find and use available local LLM service.
- Response Metrics (v1.3.1+) – Track confidence, latency, KB usage, and quality analytics.
- Vector Search (v1.3.2+) – Semantic search with ChromaDB, cross-lingual support.
- Flexible Storage – Choose between lightweight JSON files or a SQLite database for production scenarios.
- Knowledge Bases – Load categorized Q&A content to augment model responses with authoritative answers.
- Dynamic Prompting – Automatically adapts prompts based on the features you enable, reducing hallucinations.
- CLI & Tools – Includes a command-line interface plus utilities for searching, exporting, and auditing stored memories.
- Security Features (v1.1.0+) – Prompt injection detection with risk-level assessment (opt-in).
- High Performance (v1.1.0+) – Thread-safe operations with 16K+ msg/s throughput, <1ms search latency.
- Conversation Summarization (v1.2.0+) – Automatic token compression (~40-60% reduction).
- Multi-Database Support (v1.2.0+) – Export/import to PostgreSQL, MongoDB, JSON, CSV, SQLite.
Memory LLM/– Core Python package (mem_llm), configuration examples, packaging metadata, and detailed module-level documentation.examples/– Sample scripts that demonstrate common usage patterns.LICENSE– MIT license for the project.
Looking for API docs or more detailed examples? Start with
Memory LLM/README.md, which includes extensive usage guides, configuration options, and advanced workflows.
pip install mem-llm
# Or with optional features
pip install mem-llm[databases] # PostgreSQL + MongoDB
pip install mem-llm[postgresql] # PostgreSQL only
pip install mem-llm[mongodb] # MongoDB only
# Vector search support (v1.3.2+)
pip install chromadb sentence-transformersOption A: Ollama (Local, Free)
# Install Ollama from https://ollama.ai
ollama pull granite4:3b
ollama serveOption B: LM Studio (Local, GUI)
# Download from https://lmstudio.ai
# Load a model and start serverfrom mem_llm import MemAgent
# Option A: Ollama
agent = MemAgent(backend='ollama', model="granite4:3b")
# Option B: LM Studio
agent = MemAgent(backend='lmstudio', model="local-model")
# Option C: Auto-detect
agent = MemAgent(auto_detect_backend=True)
# Use it!
agent.set_user("alice")
print(agent.chat("My name is Alice and I love Python!"))
print(agent.chat("What do I love?")) # Agent remembers!
# Streaming response (v1.3.3+)
for chunk in agent.chat_stream("Tell me a story"):
print(chunk, end="", flush=True)
# NEW in v2.0.0: Function calling with tools
agent = MemAgent(enable_tools=True)
agent.set_user("alice")
agent.chat("Calculate (25 * 4) + 10") # Uses built-in calculator
agent.chat("Search my memory for 'Python'") # Uses memory tool
# NEW in v2.1.0: Async tools & validation
from mem_llm import tool
@tool(
name="send_email",
pattern={"email": r'^[\w\.-]+@[\w\.-]+\.\w+$'} # Email validation
)
def send_email(email: str) -> str:
return f"Email sent to {email}"# Install with API support
pip install mem-llm[api]
# Start API server (serves Web UI automatically)
python -m mem_llm.api_server
# Or use dedicated launcher
mem-llm-web
# Access Web UI at:
# http://localhost:8000 - Chat interface
# http://localhost:8000/memory - Memory management
# http://localhost:8000/metrics - Metrics dashboard
# http://localhost:8000/docs - API documentationfrom mem_llm import MemAgent
# LM Studio - Fast local inference with GUI
agent = MemAgent(
backend='lmstudio',
model='local-model',
base_url='http://localhost:1234'
)
# Auto-detect - Use any available local backend
agent = MemAgent(auto_detect_backend=True)
# Advanced features still work!
agent = MemAgent(
backend='ollama', # NEW in v1.3.0
model="granite4:3b",
use_sql=True, # Thread-safe SQLite storage
enable_security=True # Prompt injection protection
)For advanced configuration (SQL storage, knowledge base support, business mode, etc.), copy config.yaml.example from the package directory and adjust it for your environment.
- ✅ 20+ examples demonstrating all features
- ✅ Function Calling (3 examples - basic, memory tools, async+validation)
- ✅ Ollama and LM Studio backends (14 tests)
- ✅ Conversation Summarization (5 tests)
- ✅ Data Export/Import (11 tests - JSON, CSV, SQLite, PostgreSQL, MongoDB)
- ✅ Core MemAgent functionality (5 tests)
- ✅ Factory pattern and auto-detection (4 tests)
- Write Throughput: 16,666+ records/sec
- Search Latency: <1ms for 500+ conversations
- Token Compression: 40-60% reduction with summarization (v1.2.0+)
- Thread-Safe: Full RLock protection on all SQLite operations
- Multi-Database: Seamless export/import across 5 formats (v1.2.0+)
Contributions, bug reports, and feature requests are welcome! Please open an issue or submit a pull request describing your changes. Make sure to include test coverage and follow the formatting guidelines enforced by the existing codebase.
- PyPI: https://pypi.org/project/mem-llm/
- Documentation: Memory LLM/README.md
- Changelog: Memory LLM/CHANGELOG.md
- Issues: https://github.com/emredeveloper/Mem-LLM/issues
Mem-LLM is released under the MIT License.
- PyPI: https://pypi.org/project/mem-llm/
- Documentation: Memory LLM/README.md
- Changelog: Memory LLM/CHANGELOG.md
- Issues: https://github.com/emredeveloper/Mem-LLM/issues
Mem-LLM is released under the MIT License.