Player not loading? Watch on YouTube
OpenViking is an open source context database for AI agents from Volcengine, ByteDance's cloud arm. The review explains its virtual filesystem under viking://, where resources and memories have paths agents can browse. Entries have three levels: an abstract of roughly 100 tokens, an overview of around 2,000 tokens, and the full content. Retrieval follows directories and records its route so users can inspect how it selected a file.
The speaker reports benchmark results for version 0.3.22, including Claude Code memory accuracy rising from 57.21% to 80.32%. The headline 91% input-token reduction is the largest reported saving, not an average. Claude Code's reported token totals imply a reduction closer to 63%. The tests used Doubao 2.0 Pro and a Doubao embedding model; these results do not establish equivalent gains with every provider.
The setup walkthrough requires Python 3.10 or higher and covers installation, an interactive configuration wizard, validation, and server startup. Ollama is an option for local AI, with runtime detection and model selection based on hardware. Agent connections use different plugins, hooks, or MCP configurations, with a self-hosted server available for the MCP route.
OpenViking Helper is a beta desktop app for macOS and Windows x64. The speaker recommends this memory layer for repeated agent work over shared knowledge and flags the AGPL v3 license for review before commercial use.
