# MemU

> An agentic memory framework for LLMs and AI agents with persistent, self-evolving memory for proactive 24/7 autonomous agents.

MemU is an agentic memory framework designed for LLMs and AI agents that provides persistent, self-evolving memory capabilities. It enables autonomous AI agents to continuously predict user intentions, act proactively, and work around the clock with intelligent memory management. The platform achieves 92.09% average accuracy in reasoning tasks and offers sub-50ms latency with 99.9% uptime SLA.

- **Three-Layer Memory Architecture** provides a unified multimodal memory framework consisting of Resource Layer (raw data), Memory Item Layer (fine-grained memory items), and Memory Category Layer (thematic knowledge structures) with full bidirectional traceability.

- **Dual-Mode Retrieval** combines embedding search for fast semantic matching with LLM-based search that allows models to directly read and interpret memory category files for richer context understanding.

- **Response API** offers one API call for fully autonomous responses where agents retrieve memories, generate context-aware replies, and store new learnings automatically—perfect for 24/7 agents.

- **Memory API** provides full control over agent memory with semantic search, pattern queries, and bulk operations for storing strategic insights and building agents that anticipate user needs.

- **Self-Evolution Capability** enables memory structures to adapt automatically based on usage patterns and agent behavior, with intelligent forgetting mechanisms that gracefully manage memory decay.

- **Multimodal Support** handles text, images, audio, and video inputs, transforming heterogeneous data into queryable, semantically interpretable textual memory.

- **Visual Memory Console** allows real-time monitoring of memory health, decision tracing, and agent behavior debugging.

- **User Intention Prediction** continuously infers user intentions from behavior patterns, enabling agents to know what users need before they ask.

To get started, install the Python SDK with `pip install memu-py`, initialize the MemuClient with your API key, and use the memorize_conversation method to store interactions. The platform integrates with OpenAI, Anthropic, Gemini, DeepSeek, Qwen, and LangGraph, with CrewAI, N8N, and Dify integrations coming soon. MemU offers both cloud-hosted and self-hosted deployment options through its open-source components including memU-server and memU-ui.

## Features
- Three-layer memory architecture (Resource, Memory Item, Memory Category)
- Dual-mode retrieval (embedding search + LLM-based search)
- Response API for autonomous agent responses
- Memory API for granular memory control
- Self-evolving memory structures
- Multimodal memory support (text, image, audio, video)
- User intention prediction
- Cross-session continuity
- Proactive pattern recognition
- Visual memory console
- 24/7 always-on memory
- Intelligent forgetting mechanism
- Full bidirectional traceability
- Sub-50ms latency
- 99.9% uptime SLA
- SOC 2 compliant

## Integrations
OpenAI, Anthropic, Gemini, DeepSeek, Qwen, LangGraph, SillyTavern

## Platforms
LINUX, WEB, API, DEVELOPER_SDK

## Pricing
Open Source, Freemium — Free tier available with paid upgrades

## Version
1.0.0

## Links
- Website: https://memu.pro
- Documentation: https://memu.pro/docs
- Repository: https://github.com/NevaMind-AI/memU
- EveryDev.ai: https://www.everydev.ai/tools/memu