// Agent profile
OpenViking is a context database for AI agents developed by Volcengine (ByteDance's cloud platform) that stores memories, resources, and skills as a virtual filesystem under the viking:// protocol. Instead of opaque vector queries, agents retrieve context using familiar filesystem commands — ls, tree, find — with content organized into three processing tiers: L0 abstracts (~100 tokens for quick relevance assessment), L1 overviews (~2k tokens for core information), and L2 full details (original data loaded on demand). This tiered architecture reduces token consumption while maintaining context quality, and retrieval trajectories are fully observable for debugging. OpenViking supports automatic session-to-memory conversion, extracting structured knowledge from agent conversations asynchronously after commits. Built primarily in Rust with Python support for CLI and integrations, it integrates with Claude Code, Codex, OpenClaw, Hermes, Cursor, Trae, OpenCode, pi, MCP clients, and LangChain/LangGraph. Licensed under AGPLv3 (core) and Apache-2.0 (CLI/examples), with a hosted personal tier offering a free trial up to 50 files.
OpenViking rethinks how AI agents store and retrieve context. Developed by Volcengine (ByteDance’s cloud infrastructure division), it replaces the typical vector-database-and-hope approach with a structured virtual filesystem that agents navigate using commands they already understand — ls to list, tree to browse hierarchies, find to search. Every piece of context lives under the viking:// protocol, organized into directories that map to memories, resources, and skills. The key insight is that filesystem navigation is deterministic and debuggable, while vector similarity search is neither.
Tiered content loading: OpenViking organizes all content into three processing tiers. L0 provides abstracts of roughly 100 tokens — enough for an agent to assess relevance without loading full content. L1 delivers core overviews at around 2,000 tokens. L2 contains the complete original data. Agents load deeper tiers only when needed, dramatically reducing token consumption for context-heavy workflows. This tiered architecture means an agent can browse thousands of memories efficiently, drilling into details only where relevant.
Observable retrieval: Unlike black-box embedding stores where you cannot see why a particular chunk was retrieved (or missed), OpenViking’s filesystem-based approach produces visible retrieval trajectories. You can trace exactly which directories an agent browsed, which files it opened, and at which tier — making it straightforward to debug context quality issues.
Session-to-memory extraction: OpenViking automatically converts agent conversation sessions into structured memories. After a session commits, an asynchronous process extracts key information and stores it in the filesystem hierarchy, building an ever-growing knowledge base without manual curation.
Broad integration: The platform connects to a wide range of agent frameworks and coding tools: Claude Code, Codex, OpenClaw, Hermes, Cursor, Trae, OpenCode, pi, MCP clients, and LangChain/LangGraph. Multi-provider support for the underlying LLM (Volcengine, OpenAI, Ollama) means teams can run it with their existing infrastructure.
Development teams use OpenViking to give their AI coding agents persistent memory across sessions — decisions, patterns, and project context survive between conversations instead of being lost. Agent framework developers integrate it as the context layer for multi-agent systems where agents need shared, structured knowledge. Teams building custom AI workflows use the MCP client integration to plug OpenViking into their existing tool chains. The tiered loading is particularly valuable for agents working with large codebases or documentation sets where loading everything into context is impractical.
The core project is licensed under AGPLv3, which requires derivative works to be open-sourced — the CLI and examples use the more permissive Apache-2.0 license. Self-hosting is free; a hosted personal tier (OpenViking Personal) offers a free trial with up to 50 files, with premium tiers available for larger workloads. The Rust-based core is performant but requires a Rust toolchain for building from source. The project is relatively new and evolving quickly.
OpenViking is for developers and teams building AI agent systems that need structured, persistent context management. It suits anyone frustrated with the opacity and imprecision of vector-only retrieval, and teams that want their agents to accumulate knowledge over time rather than starting fresh every session. Agent framework developers benefit from its broad integration support and the filesystem metaphor that maps naturally to how developers already think about organizing information.
Builder.io's open-source framework for building agent-native applications — shared actions, SQL-backed state, identity, tools, skills, jobs, observability, and UI surfaces that all work together.
CLI orchestration tool that installs and configures platform connectors for 16+ social and web sources — Twitter/X, Reddit, YouTube, GitHub, Bilibili, and more. Zero API fees.
Persistent memory layer for AI coding agents — benchmark-backed (95.2% on LongMemEval-S), 92% fewer tokens per session vs full-context pasting, zero manual memory.add() calls.