{"type":"blog_post","title":"memU MCP Server: Structured Memory for Proactive AI Agents","description":"memU is an MCP Server designed as a memory harness for proactive AI agents. It offers multimodal ingestion and structured memory graphs, significantly reducing token usage by providing scoped, ranked context. Developers building advanced AI agent systems should consider memU for its efficient memory management.","content":"# memU: The Memory Harness for Proactive AI Agents\n\nmemU, an MCP Server, functions as a memory harness specifically built for proactive AI agents. It focuses on structured storage, proactive intent capture, and a claimed 10x token reduction by providing highly relevant context. With over 14,000 GitHub stars, this Python-based server offers a robust solution for managing complex agent memory.\n\n## Multimodal Ingestion and Structured Memory\n\nOne of memU's core strengths is its ability to ingest a wide array of data types. It handles conversations, documents, images, video, audio, URLs, logs, and local files, consolidating diverse information streams into a unified memory. This ingested data is then organized into a structured memory graph, persisting resources, memory items, categories, relations, summaries, and embeddings. This architecture supports detailed and interconnected memory representation for agents.\n\n## Typed Memory Extraction and Automatic Organization\n\nmemU goes beyond simple storage by performing typed memory extraction from raw sources. It can identify and extract specific memory types such as profile, event, knowledge, behavior, skill, and tool memories. A key capability is its automatic organization feature, which builds categories, relations, summaries, and embeddings without requiring manual tagging. This automates a significant portion of the memory management overhead, ensuring agents always have access to well-structured context.\n\n## Agent-Ready Context Retrieval and Pluggable Storage\n\nFor agents, memU provides agent-ready retrieval, returning scoped and ranked context that can be injected into any agent workflow. This direct utility for agent operations is central to its design. Developers also benefit from pluggable storage options, allowing them to use in-memory, SQLite, or Postgres backends with the same repository contracts, offering flexibility based on project scale and persistence needs.\n\n## Configurable LLM Routing for Context\n\nmemU incorporates profile-based LLM routing, enabling chat, embedding, vision, and transcription work to be routed through configurable LLM profiles. This allows for tailored interactions with various language models. The retrieval strategy is configured on the `MemoryService` instance via `retrieve_config`. For instance, to prioritize vector-first recall, the configuration would be:\n\n```python\nMemoryService(retrieve_config={\"method\": \"rag\"})   # vector-first recall\n```\n\nAlternatively, for LLM-ranked recall, the configuration is:\n\n```python\nMemoryService(retrieve_config={\"method\": \"llm\"})   # LLM-ranked recall\n```\n\nQueries can then be scoped using filters like `user_id`:\n\n```python\nresult = await service.retrieve(\n    queries=[{\"role\": \"user\", \"content\": {\"text\": \"What are their preferences?\"}}],\n    where={\"user_id\": \"123\"},   # scope filter\n)\n```\n\nThis granular control over retrieval methods and scoping ensures that agents receive the most relevant and efficient context for their tasks.\n\n## References\n- [memU on GitHub](https://github.com/NevaMind-AI/memU)\n- [Model Context Protocol Documentation](https://modelcontextprotocol.io/introduction)\n- [memU on model-context-protocol.com](https://model-context-protocol.com/servers/)","keywords":["memu","mcp-server","nevamind-aimemu","ai-memory-harness"],"published_at":"2026-07-09T12:00:33.969+00:00","related_repository":{"slug":"memu","type":"Server","url":"https://model-context-protocol.com/servers/memu"},"source_url":"https://model-context-protocol.com/blog/memu-mcp-server-structured-memory-for-proactive-ai-agents-mcp-server-guide"}