memU is a Python-based personal memory system designed for AI agents that compiles conversations, documents, code, images, audio, video, URLs, and tool traces into human-readable Markdown files organized across three layers: an Index that maps memories and their sources, a Memory layer containing extracted personal facts and preferences, and a Skill layer that auto-extracts and refines reusable tool patterns from execution traces. The system addresses a core problem in agent development by enabling fast, scoped context retrieval instead of rescanning entire histories or injecting long prompts into every request, resulting in faster retrieval, higher accuracy, and lower operational costs.
The repository is classified across multiple domains including LLM Memory, Language Models, Memory Management, Context Retrieval, Knowledge Storage, Semantic Search, Vector Database, Persistent Memory, AI Systems, and Information Retrieval.
Core features include multimodal ingestion supporting conversations, documents, images, video, audio, URLs, and logs; a compiled memory workspace that persists Index, Skill, and Memory layers with folders, files, source artifacts, links, summaries, and embeddings; typed memory extraction for profile, event, knowledge, behavior, skill, and tool memories; self-evolving skills that auto-extract reusable tool patterns from traces and refine them on every memorize operation; self-organizing folders that auto-build categories, links, summaries, and embeddings without manual tagging; agent-ready retrieval for scoped and ranked context; pluggable storage supporting in-memory, SQLite, and Postgres backends; and profile-based LLM routing for chat, embedding, vision, and transcription work.
The system offers both cloud and self-hosted deployment options. The cloud version at memu.so provides a managed API with endpoints for memorize operations, status checking, category listing, and memory retrieval. Self-hosted deployment requires Python 3.13 or higher and supports custom LLM and embedding providers including OpenRouter integration. The retrieval system offers two methods: RAG mode using vector-first category and item recall with optional LLM routing and sufficiency checks, and LLM mode using LLM-ranked recall at each tier.
According to the repository documentation, memU achieves 92.09 percent average accuracy on the Locomo benchmark across all reasoning tasks, with detailed experimental results available in the memU-experiment repository. The ecosystem includes complementary projects such as memU-server for backend operations with real-time sync and webhook triggers, and memU-ui for visual dashboard browsing and memory monitoring. The project maintains partnerships with frameworks and platforms including TEN-framework, OpenAgents, Milvus, xRoute, Jazz, Buddie, Bytebase, LazyLLM, and Clawdchat.
