Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-10-14. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Free (AGPL-3.0) | 2026-10-14 |
| 2 | managed service via Volcengine | 2026-10-14 |
One-liner: OpenViking is an open-source “context database” for AI agents that stores memories, resources, and skills under one virtual file system (viking:// URIs) instead of scattering them across vector stores and prompt fragments. It is built by Volcengine, the cloud arm of ByteDance.
Verified facts (checked 2026-09-03; repo stats re-read 2026-10-14)
| Attribute | Value | Source |
|---|---|---|
| GitHub stars | 39,856 | volcengine/OpenViking via GitHub API, 2026-10-14 |
| Forks / open issues | 3,158 / 954 | GitHub API, 2026-10-14 |
| License (LICENSE text) | AGPL-3.0 | raw LICENSE file, verified same day |
| Created | 2026-01-05 | GitHub API |
| Last push | 2026-10-14 | GitHub API (very active) |
| Latest release | v0.4.17.1 (2026-08-31) | GitHub Releases API |
| Python package | openviking 0.4.17.1 on PyPI |
PyPI page |
| Requires | Python 3.10+, a VLM + an embedding model | official quickstart docs |
What it actually does
Everything an agent knows — user memories, private knowledge, session context, reusable skills — is addressed like a path: viking://user/<id>/memories/, .../resources/, .../skills/. The agent can browse, read, and search its own context the way you browse files, which makes memory auditable instead of opaque. You run it either as a local Python library (uv tool install openviking) or as a self-hosted service via Docker on port 1933, and it ships integrations for Claude Code, Codex, Cursor, TRAE, Hermes (built-in provider), MCP clients, and more.
What the marketing doesn’t say
OpenViking’s homepage publishes LoCoMo benchmark charts claiming 80%+ long-term memory accuracy across host agents and “up to 63%” input-token reduction. Those numbers are real quotes from the project’s own benchmark page — vendor-reported, not independently replicated. Directionally plausible, but don’t quote them as third-party results.
The AGPL-3.0 license is the other fine print. Pulling context into your agent through the library or self-hosted server is fine for personal use; modifying OpenViking and offering it over a network triggers copyleft obligations. Commercial teams should read the license or use Volcengine’s managed service instead.
Who should use it
Teams already running coding agents (Claude Code, Codex, Hermes-style setups) that are unhappy with flat vector-store memory and want a self-hosted, auditable context layer. If you just want “make my chatbot remember things” with zero ops, this is more machinery than you need.
Repository stats, license text, package version and benchmark claims verified against primary sources on 2026-09-03, with repo stats re-read on 2026-10-14. This entry follows our verification policy: every claim here was checked against primary sources, not copied from other reviews.