OpenViking brings filesystem model to AI agent memory

Most agent memory solutions are either flat vector stores that scale poorly or hand-rolled retrieval hacks — OpenViking's filesystem metaphor is a genuinely different take that maps to how developers already think about organizing information.

• –Tiered context loading (L0/L1/L2 at ~100 / ~2k / full tokens) is the key innovation: agents spend tokens on summaries first and drill down only when needed, which is why token counts drop so dramatically
• –The `viking://resources/`, `viking://user/`, `viking://agent/` namespace split cleanly separates project context, user preferences, and agent self-knowledge — a sensible schema that most memory systems leave undefined
• –Self-evolving memory at session end (async analysis of task results to update memory directories) is the feature that could matter most long-term: agents that improve without explicit human curation
• –Built-in support for LiteLLM means it's not locked to Doubao/Volcengine models — Claude, DeepSeek, Gemini, and local models via Ollama all work out of the box
• –The OpenClaw integration is the growth driver: OpenClaw went viral in China in early 2026 and OpenViking is the recommended memory backend, which explains the 1,610 stars in a single day