Introduction

OpenViking is an open-source context database designed specifically for AI Agents. OpenViking unifies the management of context (memory, resources, and skills) that Agents need through a file system paradigm, enabling hierarchical context delivery and self-iteration. The ultimate goal is to lower the barrier for Agent development, allowing developers to focus on business innovation rather than underlying context management.

Why OpenViking

In the AI era, data is abundant, but high-quality context is scarce. When building AI Agents, developers often face these challenges:

• Context Fragmentation: Memory in code, resources in vector databases, skills scattered everywhere — difficult to manage uniformly
• Context Explosion: Long-running Agent tasks generate context with each execution; simple truncation or compression leads to information loss
• Poor Retrieval Quality: Traditional RAG uses flat storage, lacking global perspective and struggling to understand complete context
• Context Opacity: Traditional RAG's implicit retrieval pipeline is like a black box, making debugging difficult
• Limited Memory Iteration: Current memory systems only record user memories, lacking Agent-related task memories

OpenViking is designed to solve these pain points.

Core Features

1. File System Management Paradigm

Moving away from traditional flat database thinking, all context is organized as a virtual file system. Agents no longer rely solely on vector search to find data — they can locate and browse data through deterministic paths and standard file system commands.

Unified URI Identification: Each context is assigned a unique viking:// URI, enabling precise location and access to resources stored in different locations.

viking://
├── resources/              # Resources: project docs, code repos, web pages
│   └── my_project/
├── user/
│   └── {user_id}/          # Private context for the current user
│       ├── memories/       # User memories
│       ├── resources/      # Private user resources
│       ├── skills/         # Private user skills (default)
│       ├── peers/
│       │   └── {peer_id}/
│       │       ├── memories/
│       │       └── resources/
│       └── sessions/
└── agent/
    └── skills/             # Optional account-wide shared skills

Three Context Types:

Unix-like API: Familiar command-style operations

client.find("user authentication")       # Semantic search
client.ls("viking://resources/")         # List directory
client.read("viking://resources/doc")    # Read content
client.abstract("viking://...")          # Get L0 abstract
client.overview("viking://...")          # Get L1 overview

2. Hierarchical Context On-Demand Loading

Stuffing massive context into prompts all at once is not only expensive but also risks exceeding model windows and introducing noise. OpenViking automatically processes context into three levels upon ingestion:

viking://resources/my_project/
├── .abstract.md               # L0 layer: abstract
├── .overview.md               # L1 layer: overview
├── docs/
│   ├── .abstract.md          # Each directory has L0/L1 layers
│   ├── .overview.md
│   └── api.md                # L2 layer: full content
└── src/

3. Directory Recursive Retrieval

Single vector retrieval struggles with complex query intents. OpenViking implements an innovative directory recursive retrieval strategy:

1. Intent Analysis: Generate multiple retrieval conditions through intent analysis
2. Initial Positioning: Use vector retrieval to quickly locate high-scoring directories
3. Fine Exploration: Perform secondary retrieval within directories, updating candidate sets with high-scoring results
4. Recursive Descent: If subdirectories exist, recursively repeat the secondary retrieval
5. Result Aggregation: Return the most relevant context

This "lock onto high-scoring directories first, then explore content in detail" strategy not only finds semantically matching fragments but also understands the complete context of information.

4. Visualized Retrieval Traces

OpenViking's organization uses a hierarchical virtual file system structure, with all context integrated in a unified format and each entry corresponding to a unique URI, breaking away from traditional flat black-box management.

The retrieval process uses directory recursive strategy, with complete traces of directory browsing and file positioning preserved for each retrieval, enabling clear observation of problem sources and guiding retrieval logic optimization.

5. Automatic Session Management

OpenViking includes a memory self-iteration loop. After a session is committed, the system asynchronously analyzes task outcomes and user feedback, then updates memory for the current user or Peer according to the active memory policy.

Built-in memory types:

OpenViking lets applications extend or adjust memory types for their own needs.

Enabling Agents to become "smarter with use" through world interaction, achieving self-evolution.

Next Steps

• Quick Start - Get started in 5 minutes
• Architecture Overview - Understand system design
• Context Types - Deep dive into three context types
• Retrieval Mechanism - Learn about retrieval flow