OpenViking: Give Your Agent a Filesystem Instead of a Black-Box Vector Store
It's called OpenViking, an open-source context database that gives agents one place to store knowledge, user memories, and reusable experience.
The trick is how it organizes things. Everything is a virtual filesystem the agent can browse and search with commands it already knows: ls, tree, read.
And it doesn't dump everything into context. It stores summaries first, so the agent decides which files are worth opening in full before it reads them.
The results:
→ Memory accuracy: 80–83%
→ Input tokens: down 34.3% to 91%
So the agent remembers more and reads less.
Link: https://lnkd.in/eESv6Efn
100% open source.
OpenViking is real, and the specific numbers in this post match the project's own published benchmark almost verbatim — but there's important context missing about whose numbers these are.
Project confirmed: OpenViking is an open-source context database built by ByteDance's Volcengine team, organizing agent context (memories, resources, skills) as a virtual filesystem under a viking:// protocol.
"Everything is a virtual filesystem... ls, tree, read" — confirmed exactly. The project's own description states agents can operate on it like files: use ls, tree, read, and write to browse directories, read, create, and edit content, or search within a directory.
"Stores summaries first, agent decides what's worth opening" — confirmed, and matches the project's tiered-loading design precisely. Content is processed into three tiers — L0 abstract, L1 overview, L2 details — and loaded on demand, so an agent can scan cheap summaries before pulling full content.
The numbers — "Memory accuracy: 80-83%" and "input tokens down 34.3% to 91%" — are quoted essentially verbatim from OpenViking's own README benchmark section, which states: with OpenViking, all three agent integrations land at 80-83% accuracy — up from 24-57% on their native memory — while input tokens drop by 34.3-91.0% and query latency by 58.45-66.10%.
One thing the post leaves out that actually strengthens its own claim: the 80-83% figure is specifically a lift over baseline — native memory systems (OpenClaw, Hermes, Claude Code's own auto-memory) scored only 24-57% on the same LoCoMo benchmark without OpenViking. That comparison makes the number more meaningful than it reads standalone, and it's a shame to leave it out.
What's worth flagging as context, not correction: these are OpenViking's own self-reported numbers, from their own benchmark setup (Doubao 2.0 Pro as the evaluation model), not independently reproduced results. A third-party technical review explicitly notes: the project reports large memory-accuracy and token reductions on its chosen integrations, models and benchmark setup... they are not independent evidence for your corpus, and that the reviewer did not independently reproduce the project benchmarks. Not a red flag — this is normal for a new project's README — but worth treating as "vendor-reported, methodologically documented" rather than "independently confirmed."
"100% open source" is accurate but licensing detail is worth knowing: OpenViking is AGPLv3-licensed — genuinely open source, but AGPL is a stricter copyleft license (with network-use obligations) that some companies specifically avoid for commercial deployments, unlike a permissive license (MIT/Apache).