OpenViking

OpenViking is an open-source context database from Volcengine. It unifies agent memory, knowledge/RAG resources, and skills as one virtual filesystem under the viking:// protocol. Each entry is written into three tiers — L0 abstract, L1 overview, L2 full detail — so agents load only as deep as the task needs. Sessions can commit into long-term preferences and experience memory. There is a hosted Studio playground, server + CLI install, and published LoCoMo / tau2-bench numbers on memory accuracy and token savings.

Agent memory is usually a vague retrieval bag that burns tokens and is hard to debug. Founders shipping multi-session agents need a real store they can inspect, tier, and evolve. OpenViking turns context into something you can ls, tree, and audit with a path history.

pip install openviking, run openviking-server init and doctor, then start the server. Point agents at the ov client or APIs and address context with viking:// URIs for resources, memories, and skills. Retrieval first finds a high-scoring directory, then drills down layer by layer while keeping the browse trajectory for debugging. After a session ends, async extraction can fold user preferences and agent experience back into durable memory.

This is not another coding harness and not a single-repo memory MCP. It is a context database product with a filesystem metaphor, tiered loading to cut tokens, observable directory trajectories, and a live Studio demo. You get one place for memory, RAG resources, and skills instead of three disconnected stores.

• Public demo available
• API / SDK surface

• codebase-memory-mcp ↗
• Paritok ↗