OpenViking
A memory management tool for AI agents that organizes contextual information in a file-system structure, making it searchable and retrievable. It addresses the context management challenges that arise when AI agents switch across multiple projects, enabling persistent and organized memory.
Timeline (last 90 days)
OpenViking can be integrated into DeepSeek Harness either as a plugin or via MCP.
The UP host's OpenViking library contains approximately 5,270 projects across three mount points.
The proportion of zero-result queries in OpenViking depends largely on whether the user's directory structure is well-organized.
The OpenViking repository includes reproduction scripts so that users can verify the official data themselves.
OpenViking is a contextual database project open-sourced by ByteDance.
OpenViking organizes everything an AI needs to remember into a file system, with three top-level directories—Agent, Resources, and User—that can be operated using command-line tools such as ls, tree, and find.
OpenViking uses a three-tier progressive loading mechanism—L0 (one-sentence summary), L1 (overview layer), and L2 (full text)—to conserve tokens.
According to official benchmark data from OpenViking, the reduction in input tokens ranges from 34% to 91%.
OpenViking's official data shows that memory accuracy on Claude Code rose from 57% to 80%.
In the UP host's own testing, 40.3% of 1,257 retrievals returned zero results.
15 more timeline events
All Facts (20)
OpenViking is an AI agent self-evolving context database open-sourced by ByteDance's Volcano Engine, which unifies an agent's memory, knowledge RAG, and skills onto a single data foundation.
60%UnverifiedThe OpenViking repository includes reproduction scripts so that users can verify the official data themselves.
50%UnverifiedOpenViking is a contextual database project open-sourced by ByteDance.
50%UnverifiedOpenViking organizes everything an AI needs to remember into a file system, with three top-level directories—Agent, Resources, and User—that can be operated using command-line tools such as ls, tree, and find.
50%UnverifiedOpenViking uses a three-tier progressive loading mechanism—L0 (one-sentence summary), L1 (overview layer), and L2 (full text)—to conserve tokens.
50%UnverifiedAccording to official benchmark data from OpenViking, the reduction in input tokens ranges from 34% to 91%.
50%UnverifiedOpenViking's official data shows that memory accuracy on Claude Code rose from 57% to 80%.
50%UnverifiedIn the UP host's own testing, 40.3% of 1,257 retrievals returned zero results.
50%UnverifiedThe proportion of zero-result queries in OpenViking depends largely on whether the user's directory structure is well-organized.
50%UnverifiedOpenViking can be integrated into DeepSeek Harness either as a plugin or via MCP.
50%UnverifiedOfficial tests show that after integrating OpenViking, long-conversation memory accuracy for several agents exceeded 80%, and input token consumption was reduced.
50%UnverifiedThe OpenViking main project uses the AGPLv3 license.
50%UnverifiedDeploying OpenViking requires configuring a server, an embedding model, and agent integration, and after writing materials one must wait for parsing and indexing to complete.
50%UnverifiedOpenViking's retrieval logic first locates relevant directories and then drills down layer by layer, returning results while preserving surrounding structural relationships, distinguishing it from vector search which returns isolated fragments.
50%UnverifiedDuring retrieval, OpenViking saves the traversed path, making the retrieval process traceable and reviewable.
50%UnverifiedOpenViking Server supports Volcano Engine, OpenAI, Kimi, and GLM, and can also connect to local Ollama.
50%UnverifiedOpenViking stores three types of content: knowledge (project documentation, codebases, web pages), memory (user preferences, historical decisions, completed tasks), and capabilities (callable Skills).
50%UnverifiedAll content in OpenViking is assigned an address beginning with "viking" and stored in directories.
50%UnverifiedOpenViking prepares three levels of summaries for each piece of written content: L0 one-sentence summary, L1 more detailed summary, and L2 full original text.
50%UnverifiedOpenViking's self-evolution mechanism emphasizes that the database can continuously evolve as the agent operates, distilling new experiences from interactions, updating existing knowledge, and eliminating outdated information, distinguishing it from static traditional vector databases.
50%