OpenViking is a context database for AI agents that stores knowledge, memory, and skills in one virtual filesystem under viking:// URIs. Agents navigate it with file operations such as ls, tree, read, write, and grep, and a person can open any directory to inspect and edit what the agent knows. Resources hold documents and code, memories hold user preferences and experience, and skills define how to perform tasks.
Each processed directory carries generated summaries at three tiers: an L0 one-sentence abstract, an L1 overview, and the L2 full content loaded only when needed. Semantic search can be scoped to a project or memory subtree instead of a flat vector pool, and committing a session archives the conversation and extracts memories as editable Markdown. The bundled VikingBot agent framework can also compile source material into a wiki, knowledge graph, or report.
Coding agents connect through a memory plugin installer that covers Claude Code, Codex, Cursor, TRAE, OpenCode, and others through hooks and MCP, with further integrations for OpenClaw, Hermes, pi, DeerFlow, and LangChain. Python, Go, and TypeScript SDKs and an HTTP API cover custom integrations. The server is self-hosted with accounts, user isolation, and opt-in resource ACLs, or available as a hosted service from Volcengine, and a beta desktop app configures local agent integrations and syncs memories.
Published evaluations of version 0.3.22 report the following LoCoMo long-conversation memory accuracy:
| Agent | Native memory | With OpenViking |
|---|---|---|
| OpenClaw | 24.20% | 82.08% |
| Hermes | 33.38% | 82.86% |
| Claude Code | 57.21% | 80.32% |
Features
- Filesystem context model: knowledge, memory, and skills under viking:// URIs, browsed and edited like files
- Tiered loading: L0 abstracts and L1 overviews let agents judge relevance before reading L2 full content
- Directory-scoped retrieval: semantic search within a project or memory subtree, with find for direct queries and search for session-planned retrieval
- Session memory extraction: committing a session archives it and extracts memories as Markdown that can be edited and merged
- Agent memory plugins: hooks and MCP integrations for Claude Code, Codex, Cursor, TRAE, and OpenCode, detected by one installer
- Framework integrations: context engine for OpenClaw, built-in Hermes support, a native pi extension, and LangChain tools and store
- CLI and SDKs: the ov CLI plus Python, Go, and TypeScript SDKs and an HTTP API
- VikingBot: an agent framework built on OpenViking that compiles material into a wiki, knowledge graph, or report
- Multi-tenant server: accounts, user isolation, authentication, and opt-in resource ACLs
- Desktop app: a beta console for macOS and Windows that configures agent integrations and inspects recall and capture events
- Deployment options: self-hosted server, Docker image, hosted OpenViking Service on Volcengine, or a licensed self-managed edition
