Coding Horizon

Your AI Agent Is Forgetting The Most Expensive Part

Sources for every figure, version, licence and benchmark this video puts on screen. Checked 2026-08-25.

What OpenViking is

OpenViking is an open source context database for AI agents, published by ByteDance’s Volcengine team at github.com/volcengine/OpenViking. It stores memory, resources and skills as one virtual filesystem addressed by a viking:// URI, so an agent browses its own context rather than querying an opaque vector store.

The three loading tiers

The project documents three levels of context, loaded on demand:

Level What it holds
L0 the abstract, for deciding whether a directory is relevant at all
L1 the overview, enough to understand what is inside without opening every file
L2 the detail, the original source, loaded only when the task earns it

L0 and L1 are directory sidecars written alongside the content as .abstract.md and .overview.md. Parent summaries are generated bottom up: the semantic processor walks the tree upward so a directory’s summary incorporates its children.

Version tested

The published evaluations were run on 0.3.22.

LoCoMo, long conversation memory accuracy

Native baseline against the same agent with OpenViking attached:

Integration Baseline With OpenViking Change
OpenClaw 24.20% 82.08% up 57.88 points
Hermes 33.38% 82.86% up 49.48 points
Claude Code 57.21% 80.32% up 23.11 points

Input tokens and latency

Across the three agent integrations the project reports input tokens falling by 34.3% to 91.0% and query latency improving by 58.45% to 66.10%. The documentation separately reports an 83% token cost reduction for OpenClaw specifically (4.3M tokens against a 24.6M baseline).

tau2 bench, task success on multi turn work

Domain Baseline With OpenViking Change
retail 70.94% 77.81% up 6.87 points
airline 54.38% 66.25% up 11.87 points

Maturity and licence

The published package classifier is Development Status :: 3 - Alpha, and the licence is AGPL-3.0. The repository notes exceptions: crates/ov_cli and the examples are Apache 2.0. AGPL section 13 covers remote network interaction, which is the clause that makes it a question for commercial codebases rather than a formality.

Hosting

Alongside self hosting, a managed SaaS is offered on ByteDance’s Volcano Engine, in personal and enterprise tiers.

Observability

The project lists visualised retrieval trajectories as its answer to context opacity: the route a retrieval took, the scope and path at each hop, and which level was loaded.

Mem0

Mem0 exposes a small memory API, add() and search(), embedding content into a vector store for semantic retrieval. Its Pro tier additionally builds a knowledge graph of entities and relationships for multi hop queries. It is framework agnostic.

Zep

Zep builds a temporal knowledge graph, which its documentation calls the Context Graph, out of messages and documents. It extracts entities, relationships and facts, and produces episodes, thread summaries and user summaries. Facts carry validity dates: as new data arrives, outdated facts are invalidated with a timestamp rather than deleted, so history is preserved.

Letta

Letta is a stateful agent framework rather than a memory store. Agents run inside the Letta runtime, which owns the agent loop, tool execution, state persistence and tiered memory, with core memory held in the context window.

Not verified