OpenViking
An open-source context database for AI agents that organizes knowledge, memory and skills as a browsable virtual filesystem instead of an opaque vector index.
OpenViking exposes an agent's context under a viking:// filesystem that can be navigated with familiar file operations such as ls, tree, read, write and grep, so a person can open any directory to inspect or edit what the agent knows, rather than treating memory as an opaque store of embeddings. Every directory carries a generated summary, and content loads in three tiers: a one-sentence abstract, a fuller overview, and the full original data, so an agent can judge relevance before reading further.
Semantic search can be scoped to a specific project or memory subtree instead of scanning the whole index. Completed conversation sessions are archived as Markdown that can be inspected, edited and merged, and an optional feature can organize source material into a wiki, knowledge graph or report. In the project's own evaluation on long-conversation memory and multi-turn agent benchmarks, adding OpenViking's memory raised accuracy and task success over the agents' native memory while reducing input tokens and latency.
It is written in Python, installs via pip, and is licensed under AGPL-3.0.