Key points
- The problem it solves: Most AI assistants forget everything between sessions, forcing users to repeat context every time. memU provides persistent, cross-session memory so agents behave like companions with long-term recall.
- What memU is: An open-source memory framework from NevaMind AI, designed for always-on (24/7) proactive AI agents. It has gained more than 10,800 GitHub stars and is positioned as memory infrastructure for the AGI era.
- Three-layer hierarchical architecture (file-system style, not a flat vector DB):
- Resource: raw materials such as conversations, documents, images, audio, and video.
- Item: structured knowledge points extracted from resources (preferences, facts, learned skills).
- Category: auto-organized topic folders forming a dynamic memory tree.
- Why a filesystem model: memories are explainable, navigable from broad themes down to specific facts, and portable (exportable and backed up like folders).
- Core capabilities:
- 24/7 proactive agent that learns and updates in the background.
- User-intent capture — e.g., mentions "business trip to Paris next week" trigger inferences about flights, hotels, restaurants, and prior food dislikes.
- Cost efficiency via memory caching and selective context injection, avoiding linear token growth.
- Dual-mode retrieval: fast RAG vector search for real-time responses vs. slower LLM-based reasoning for complex intent prediction.
- Memory lifecycle: Monitor inputs → extract insights and update profile → predict user intent → proactively pre-fetch context and prepare suggestions.
- Deployment options:
- Cloud at memu.so (REST API): endpoints such as
/api/v3/memory/memorizeand/api/v3/memory/retrieve, with RAG-style queries. - Self-hosted via
pip install memuand theMemUServicePython client, suitable when data privacy is required. - Ecosystem integrations: originally built for OpenClaw; now used with the TEN Framework (real-time voice agents), n8n (workflow automation), and the official n8n plugin.
- Comparison vs. alternatives:
- memU — filesystem architecture, native proactivity, high explainability, non-embedding retrieval, self-evolving prioritization.
- Mem0 — vector + graph, limited proactivity, medium explainability.
- Zep — temporal graph, proactivity supported, high explainability.
- MemOS — neural tensor, proactivity supported, lower explainability.
- Use cases: AI companion/emotional support, personal assistant, personalized tutoring, financial advisory, and consistent customer service.
- Philosophy: If the model represents reasoning and tools represent action, memory represents personality — identity, history, preferences, values, time awareness, behavior patterns, and a world model.
- GitHub: https://github.com/NevaMind-AI/memU
- Cloud service: https://app.memu.so
- Documentation: https://memu.pro/docs
- Community: Discord