Inspect and edit agent context like a file system
When you need to know what an agent remembers, you no longer have to guess. OpenViking bundles an agent's knowledge, memory, and skills into a single virtual file system, allowing you to view and modify contents as easily as opening a folder. This provides a foundation for transparently managing the context an agent has accumulated and efficiently retrieving only the necessary information.
A file system, not a black box
Most agent memories operate like a black box where text goes in and only embeddings come out. OpenViking places all context under a virtual path called viking://, enabling users to directly inspect what the agent knows using commands like ls, tree, and read. Each directory includes an automatically generated summary, allowing the agent to assess relevance before reading the full content. This makes it possible to trace which information the agent used for its decisions and to edit it directly when needed.
Hierarchical information loading and precise search
OpenViking loads information in three stages. L0 provides a one-sentence summary for quick relevance checks, L1 offers key information and usage scenarios for planning, and L2 contains the full original content for detailed analysis. Because the agent selectively reads only the necessary information along this hierarchy, it reduces unnecessary token consumption. Additionally, the search scope can be limited to specific projects or memory subdirectories, improving accuracy compared to scanning the entire index.
Performance verified by benchmarks
OpenViking version 0.3.22 has been verified for performance on the LoCoMo and tau2-bench benchmarks. In user memory evaluation, the agent integration environment with OpenViking recorded an accuracy of 80–83%, a significant improvement over the 24–57% of existing native memory methods. At the same time, input tokens decreased by 34.3–91.0%, and query latency was shortened by 58.45–66.10%. In agent experience evaluation, having memory functionality increased task success rates by 6.87 percentage points in the retail domain and 11.87 percentage points in the aviation domain compared to not having it.
Installation and agent integration
Running the server requires uv and Python 3.10 or higher, along with a model provider that supplies an embedding model and a VLM. After installation, restarting the agent activates memory sharing between sessions.
Pre-deployment checks
OpenViking is distributed under the AGPLv3 license, so commercial distribution requires considering Volcengine's managed service or a self-hosted edition. Before exposing the server outside the local host, authentication settings must be completed, and security can be ensured through user isolation and resource ACLs. The desktop app is in beta and supports macOS and Windows x64, providing local agent integration settings and memory synchronization features. It is recommended to review policies regarding API key management for specific model providers and network access permissions in advance.
