Agent memory has felt like a junk drawer: we throw context in, pray the model finds it, and hope the glue holds. OpenViking’s bet is bolder — treat memory like an operating system, not a shoebox.
Trending on GitHub today, OpenViking describes itself as a “context database” for AI agents. The twist is the storage model: memories, resources, and skills become a virtual filesystem with deterministic paths. Instead of “search and pray,” agents can navigate context like directories.
The design is explicit. A Viking URI looks like viking://{scope}/{path} with scopes for resources, user, agent, and session. That gives agents a map, not just a bag of embeddings.
OpenViking also bakes in layered context delivery. The API surfaces three tiers:
- L0 abstracts (one‑sentence summaries)
- L1 overviews (mid‑detail snapshots)
- L2 full content (deep reads on demand)
The CLI mirrors that structure with abstract, overview, and read commands. Even the memory extractor code stores each memory as abstract/overview/content, so “self‑evolving memory” isn’t just a slogan — it’s a schema.
Why this matters: Codex and Gemini have been tightening how they inject or compact context. OpenViking shifts the substrate itself. It makes context inspectable, debuggable, and routable — a filesystem you can traverse, not a black‑box retrieval gamble.
Takeaway: OpenViking treats context like an OS does files — organized, layered, and addressable — which could be the missing primitive for reliable long‑running agents.
What would change for your team if agent memory had real paths and predictable layers instead of fuzzy retrieval? If that question sparks ideas, go explore OpenViking’s repo and bring your feedback to the community — the best context systems will be built in the open.