volcengine/OpenViking

OpenViking treats agent context as a filesystem problem, not a retrieval problem. Every memory, resource, and skill gets a viking:// URI, and agents navigate it with ls, tree, and find — the same mental model a developer uses with files — rather than firing off queries to an opaque vector index.

The architecture's central bet is tiered loading. On write, every entry is decomposed into three layers: a one-sentence abstract (L0), a structured overview covering key points and usage scenarios (L1), and the full original content (L2). Directories carry their own L0 and L1 layers, so a retrieval pass can judge relevance at the folder level before ever touching a full document. Vector search locates the highest-scoring directory first, then drills down only as far as the task demands. The practical effect is that tokens spent on context scale with actual need rather than with corpus size.

The benchmark numbers are striking. On the LoCoMo long-conversation memory benchmark, three different agents — OpenClaw, Hermes, and Claude Code — all landed between 80 and 83% accuracy when backed by OpenViking, up from a range of roughly 24 to 57% on their native memory. Input tokens dropped by between 34 and 91% depending on the agent, and query latency fell by 58 to 66%. On tau2-bench multi-turn agent tasks, experience memory lifted task success by nearly 7 percentage points on retail scenarios and nearly 12 on airline scenarios. Those gains come from OpenViking asynchronously extracting agent experience and user preferences after each session commits — turning conversation history into structured long-term memory rather than discarding it.

The observability angle is underrated. Every retrieval preserves the directory-browsing trajectory that produced it, so when a result is wrong you can trace exactly which path was followed. That's a meaningful debugging affordance in a space where most retrieval failures are invisible.

The open-source edition is AGPLv3 with no feature gates and no account requirement. The CLI Rust crate ships under Apache 2.0. Commercial tiers exist — a managed SaaS on Volcano Engine and a self-managed enterprise option — but they differ only in who operates the infrastructure, not in what the software can do. The underlying research is formalized in the VikingMem paper, accepted to VLDB 2026.

The main practical constraint is that the benchmark evaluations used Volcengine's own Doubao models for both the VLM and the embedding layer, so performance on other providers may differ. The project supports OpenAI, Kimi, GLM, and local Ollama alongside Volcengine, but independent benchmarks on those backends don't appear in the README.