OpenViking

OpenViking is an open-source context database for AI agents. Instead of treating memory, resources, skills, and retrieved documents as separate stores, it organizes agent context through a filesystem-style model. That gives agents a clearer way to load, search, and update the information they need across long-running tasks. The project is aimed at developers building agent systems that need more than a flat vector database. OpenViking describes a context layer with hierarchical organization, directory-style retrieval, tiered context loading, visualized retrieval paths, and automatic session management. The goal is to make context observable and manageable: an agent can pull the right memory or skill from a structured context tree rather than repeatedly stuffing large chunks into the prompt. OpenViking is especially relevant for agentic RAG, coding agents, and self-improving assistants that accumulate working memory over time. It can store task memories, reusable skills, and resources, then deliver them with L0/L1/L2 style context tiers to reduce token waste. The repository also supports VLM and embedding model configuration through providers such as Volcengine Doubao, OpenAI-compatible endpoints, Kimi, and GLM, so teams can connect it to their preferred model stack. Setup is more technical than a hosted SaaS tool. The Python package is available through pip, while the CLI can be installed through npm or Cargo. Developers should expect to configure model providers and review the AGPL license before commercial use. For the right user, OpenViking fills a clear infrastructure gap: it gives AI agents a structured, inspectable context database instead of relying only on ad hoc RAG chunks, chat history, or scattered prompt files. That makes it a strong fit for teams building agents that need durable memory, skill loading, and debuggable retrieval behavior. Developers should think of OpenViking as infrastructure for agent context, not as a finished chatbot. It is most useful when an agent has to carry knowledge across sessions, choose among many skills, or inspect how a retrieval decision was made. A simple assistant may not need that extra layer. A long-running engineering agent, research agent, or enterprise workflow agent often does. The filesystem metaphor also makes collaboration easier. Humans can reason about directories, resources, and skill files more easily than opaque vector chunks. That can improve debugging and governance because teams can ask what context was loaded, where it came from, and whether it should be changed. The main adoption cost is setup: teams need to configure models, embeddings, storage, and the workflow that writes useful memories back into the context database. Once that loop is in place, OpenViking can become a durable memory and retrieval substrate for agent products.