ByteDance's OpenViking Turns Agent Memory Into a Filesystem

ByteDance's volcengine team turned agent memory into a filesystem agents can browse instead of query, per the OpenViking repo on GitHub. That distinction matters for anyone debugging a bad retrieval, since a vector store can't show its work the way a directory trail can. Accuracy across three agent integrations on the LoCoMo benchmark jumped from a 24-57% baseline to 80-83%.

The system stores memories, resources and skills as a virtual filesystem addressable through a viking:// protocol. Agents navigate it with ls, tree and find instead of firing similarity queries into an opaque index. OpenViking keeps a record of each query's browsing path, so you can trace why it retrieved what it did.

Input tokens dropped 34-91% and latency fell 58-66% across the same three integrations. On tau2-bench, a separate benchmark, adding experience memory added another 7-12 percentage points of success rate.

Content loads in three tiers instead of all at once: about 100-token abstracts, about 2k-token overviews, and full detail pulled in only as needed.

Licensing is split, and it matters for commercial use. The core ships AGPLv3, while the CLI and examples are Apache-2.0, a real difference between embedding the core and calling it as a service.

It's shipping alongside other high-star agent tooling on GitHub; a security-rules repo covering agent harnesses hit number 5 on GitHub's trending page.

The repo doesn't say how OpenViking performs on tasks outside LoCoMo and tau2-bench. It also doesn't say how the filesystem model holds up once a store holds millions of entries instead of benchmark-sized test sets.