OpenViking

Agent context gets managed like a disk in OpenViking, an AGPL-3.0 context database from ByteDance's Volcengine team. Memories, documents, and skills all live as entries in one virtual filesystem addressed by viking:// URIs, so an agent explores what it knows with ls, find, and tree commands instead of firing opaque vector queries, and every retrieval leaves a browsable directory trajectory that can be audited afterward. Each entry carries three tiers, an abstract, an overview, and full detail, and the server hands agents the cheapest tier first, a scheme the project measures at 34 to 91 percent token savings compared with dumping raw context into the prompt. Retrieval itself is hybrid, combining directory structure with semantic search, and ingestion converts PDFs, web pages, and images into the same tree. The core is written in Rust with a Python SDK and CLI installed via pip on Python 3.10+, running as a client-server HTTP service that calls configured embedding and vision model endpoints. Open-sourced in January 2026, the repo topped GitHub trending and has passed 32,000 stars.