OpenViking

Why we included this project

OpenViking treats an agent's memory, knowledge, and skills as one browsable filesystem under the viking:// protocol. Instead of querying a black-box vector store, the agent inspects its own context with familiar commands like ls and tree. Content is stored in three tiers (abstract, overview, detail) and loaded only as deep as a task requires, so retrieval stays cheap and long-running sessions don't inflate token costs. Because memory, RAG, and reusable skills share one store, it fits production agents that juggle many tools. The database also reorganizes itself as the agent works, and each retrieval leaves a debuggable trajectory, which helps when you want context that stays relevant without constant manual curation.