Treat AI agent memory like a filesystem

What it does

OpenViking is a context database that treats an AI agent’s memories, knowledge resources, and skills as files in a unified hierarchy instead of scattering them across vector databases and code. It loads context in three demand-driven tiers (L0/L1/L2) to limit token consumption, and automatically compresses session history into long-term memory so the agent accumulates experience. The goal is to replace the usual fragmented RAG stack with a single, browsable context store.

The interesting bit

The filesystem metaphor is structural, not cosmetic: retrieval works by recursively traversing directories and blending positional directory context with semantic vector search. It also visualizes the exact retrieval trajectory, turning the typical black-box RAG pipeline into something you can debug like a file path.

Key highlights

• Filesystem paradigm unifies memory, knowledge RAG, and skills in one browsable hierarchy
• Tiered L0/L1/L2 context loading to cut token costs
• Recursive directory retrieval mixing semantic search with directory positioning
• Visualized retrieval trajectories that expose the black-box RAG chain
• Automatic session compression and long-term memory extraction

Caveats

• Requires Python 3.10+, a Rust toolchain, and a C++ compiler, so it is not a lightweight pure-Python dependency
• Depends on external VLM and embedding models; even the local Ollama path needs model downloads and a network connection for initial setup
• Licensed under AGPLv3

Verdict

Worth evaluating if you are building long-running agents and are tired of gluing vector databases, prompt compressors, and memory layers into a fragile RAG stack. Skip it if you need a simple, self-contained drop-in that runs without external model providers or compiling Rust extensions.