OpenViking guides
What is a context database for AI agents?
A context database stores and retrieves the information an AI agent needs to work: reference material, persistent memory, and reusable skills. OpenViking is an open-source implementation that exposes this context as a navigable virtual filesystem with viking:// addresses.
What does it store?
Resources include documents, code repositories, and other reference material. Memories capture preferences, events, entities, and experience extracted from interactions. Skills describe reusable capabilities. These have different lifecycles, but an agent can address them through the same context interface.
Context database vs vector database
A vector index finds similar items. A context database adds organization and a lifecycle around retrieval: named resources, summaries, sessions, memory updates, and access boundaries. OpenViking uses vector retrieval as part of its implementation; it does not make embeddings obsolete. Compare the complete retrieval workflow, not just the storage label.
How OpenViking retrieves context
An agent can browse directories or search semantically. Layered content lets it inspect an L0 abstract, read an L1 overview, and open L2 source material when it needs the details. Hierarchical retrieval narrows the search through directory context. The right layer depends on the question and the evidence needed to answer it.
Where does the agent harness fit?
The harness runs the model, tools, and task loop. OpenViking provides a context backend that the harness can call through an integration, SDK, CLI, HTTP API, or MCP. Connecting storage does not automatically replace a harness's prompt assembly or compaction policy; that behavior depends on the integration.
When should you evaluate it?
Evaluate a context database when agents repeatedly rediscover project information, lose useful task history, or need shared resources with distinct user scopes. A small application with a fixed prompt may not need another service. Test retrieval quality, latency, operating cost, and isolation using your own tasks before adopting it.