finds.dev← search

// the find

volcengine/OpenViking

★ 26,112 · Python · AGPL-3.0 · updated Jun 2026

OpenViking is an open-source context database designed specifically for AI Agents(such as openclaw). OpenViking unifies the management of context (memory, resources, and skills) that Agents need through a file system paradigm, enabling hierarchical context delivery and self-evolving.

OpenViking is a context management layer for AI agents that organizes memory, resources, and skills into a virtual filesystem addressable via `viking://` URIs. It applies L0/L1/L2 tiered loading so agents only pull the detail level they need, plus a recursive directory retrieval strategy instead of flat vector search. Built by Volcengine (ByteDance's cloud arm) and backed by a VLDB 2026 paper.

The L0/L1/L2 tiered context loading is the most interesting idea here — abstracting content at three granularities so the agent pays only for what it needs is a real improvement over stuffing full documents into context windows. The filesystem metaphor (ls, find, grep over viking:// URIs) makes retrieval observable and debuggable in a way that black-box RAG pipelines are not. The LoCoMo benchmark numbers are striking: OpenClaw goes from 24% to 82% accuracy while dropping token usage by 91% — even if self-reported, the magnitude is hard to fake. HotpotQA RAG at 91% accuracy with 0.23s retrieval latency is genuinely fast compared to graph-based alternatives like LightRAG which take 75s.

The `openai-codex` provider is a red flag — it uses ChatGPT OAuth session credentials to call the Codex API without a real API key, which is a TOS violation and will break silently when session handling changes. AGPL license on the main project means any commercial product wrapping this has to open-source its own code; the split licensing (AGPL core, Apache CLI/examples) will create legal review headaches. Setup requires Python 3.10+, a Rust toolchain, and a C++ compiler — three separate toolchains before you can run it locally, which is friction most teams will underestimate. The 'recommended' production deployment is on Volcengine ECS, and the commercial tier runs on VikingDB (also Volcengine) — the open-source shell around a proprietary cloud funnel is a pattern worth noticing before building on it.

View on GitHub → Homepage ↗

// want more like this?

We dig through GitHub every week and send a few repos picked for what you actually care about — each with an honest take like this one.

Get finds in your inbox → Search again →