OpenViking

// Overview

OpenViking is an open-source context database for AI agents, published by Volcengine. It stores three kinds of context — memories, resources, and skills — as entries in a single virtual filesystem under the viking:// protocol, so an agent locates and manipulates its own context with ls, tree, find, and grep instead of issuing opaque queries at a vector store. Each entry gets a URI, which makes context addressable the way a developer's files are.

Every entry is processed on write into three layers: L0, a one-sentence abstract for a quick relevance check; L1, an overview with core information and usage scenarios for planning; and L2, the full original data, read only when needed. Directories carry their own L0 and L1 layers, so relevance can be judged before any full file is opened. Retrieval works the same way: vector search first locates the highest-scoring directory and then drills down layer by layer, and each query preserves its browsing trajectory so a wrong-looking result can be traced to the path that produced it.

Sessions feed back into memory — after a session commits, OpenViking asynchronously extracts user preferences and agent experience into long-term storage. It installs as a Python server with an ov client CLI, has documented integrations for Claude Code, Codex, OpenClaw, Hermes, Cursor, TRAE, OpenCode and pi, and publishes reproduction scripts for its LoCoMo and tau2-bench evaluations in the repository.

// What it does

• One viking:// filesystem for memories, resources and skills, browsed with ls, tree, find and grep so context is addressed deterministically
• Three-tier processing (L0 abstract, L1 overview, L2 details) with on-demand loading, so a task pulls only as deep as it needs
• Directory-recursive retrieval: vector search locates the best-scoring directory first, then drills down, keeping surrounding context intact
• Observable retrieval — every query keeps its directory-browsing trajectory, so a wrong result can be traced to the path that produced it
• Session commit extracts user preferences and agent experience into long-term memory asynchronously
• Drop-in integrations for Claude Code, Codex, OpenClaw, Hermes, Cursor, TRAE, OpenCode and pi that inject recall and auto-commit session memory

// Getting started

OpenViking needs Python 3.10 or later. The pip install brings both the server and the ov client CLI; an interactive wizard handles provider setup. A hosted playground is available at openviking.ai/studio if you want to try it before installing.

pip install openviking --upgrade
openviking-server init
openviking-server doctor
openviking-server

ov status
ov add-resource https://github.com/volcengine/OpenViking --wait
ov ls viking://resources/
ov tree viking://resources/volcengine -L 2

ov find "what is openviking"
ov grep "openviking" --uri viking://resources/volcengine/OpenViking/docs/en

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Give a coding agent long-term memory of a codebase and its docs without pasting the whole repository into the context window
• Cut token spend on retrieval by loading abstracts and overviews first and full documents only when a task needs them
• Debug a bad retrieval by replaying the directory path the query actually walked, instead of guessing at embedding similarity
• Carry user preferences and accumulated agent experience across sessions and across several different agent harnesses

// How OpenViking compares

OpenViking alongside other open-source agent memory tools AI/TLDR tracks, ranked by GitHub stars.