volcengine/OpenViking

OpenViking treats agent context as a filesystem problem, not a search problem. Every piece of memory, every resource, every skill gets a viking:// URI and lives in a directory tree the agent can traverse with ls, tree, and find. That design choice is the whole thesis: instead of firing a query into a vector store and hoping the top-k results are coherent, an agent navigates its own context the way a developer navigates a codebase.

The tiered loading scheme is where the engineering gets interesting. On write, every entry is processed into three layers: L0 is a one-sentence abstract for quick relevance checks, L1 is a structured overview covering core information and usage scenarios, and L2 is the full original content. Directories carry their own L0 and L1 layers, so the retrieval path can decide whether to drill deeper before loading anything large. The README quotes token reductions of 34 to 91 percent across three agent integrations on the LoCoMo long-conversation benchmark — a range that wide suggests the gains depend heavily on how verbose the baseline agent already is, but even the low end is meaningful at scale.

The benchmark numbers are the most concrete thing in the README. On LoCoMo user-memory accuracy, three agents (OpenClaw, Hermes, Claude Code) go from a range of roughly 24 to 57 percent natively to a tight cluster of 80 to 83 percent with OpenViking. On tau2-bench multi-turn agent tasks, task success improves by about 7 percentage points on retail and nearly 12 on airline scenarios. The paper behind this — VikingMem, accepted to VLDB 2026 — is cited directly, so the claims have a peer-reviewed anchor even if the README itself is marketing material.

The observability angle is underplayed but practically important. Every retrieval preserves its directory-browsing trajectory, so when a result is wrong you can see which path produced it. That is a real debugging affordance that vector-only stores do not offer.

The open-source edition is AGPLv3 with no feature gates, which matters: the commercial tiers are about who operates the infrastructure, not about which features you can access. The CLI supports Volcengine, OpenAI, Kimi, GLM, and local Ollama, and the Ollama path can detect and install the runtime automatically — a detail that suggests the team is serious about local-first deployment, not just cloud upsell.

The main thing to watch: the filesystem metaphor is elegant but it shifts complexity onto whoever structures the context tree. A poorly organized viking:// hierarchy will produce the same retrieval failures as a poorly chunked vector store, just with different error modes.