Reporter Notes

Notes — 2026-03-15 — OpenViking context filesystem

Trend scan (GitHub Trending)

• volcengine/OpenViking trending with large daily stars; described as open-source context database for AI agents using a filesystem paradigm.
• Other trending repos: anthropics/claude-plugins-official, lightpanda-io/browser, msitarzewski/agency-agents.

Candidate angles (LLM review)

1) OpenViking as “context filesystem” OS-like substrate for agent memory (strongest).

2) Claude plugins directory as ecosystem normalization.

3) Lightpanda as AI-first headless browser primitive.

Chosen angle

OpenViking reframes agent memory as a filesystem hierarchy with explicit scopes and multi-layer context (L0/L1/L2), making context delivery deterministic, inspectable, and self-evolving.

Evidence highlights

• README: filesystem paradigm, tiered context loading, directory recursive retrieval, self-iteration (OpenViking README).
• Viking URI docs: viking:// scheme with scopes/resources/user/agent/session; context organized as filesystem.
• API docs: L0 abstract / L1 overview / L2 full content via API/CLI; explicit L0/L1/L2 surface.
• CLI code: commands for abstract/overview/read (crates/ov_cli/src/commands/content.rs).
• Memory extractor code: candidate memory has abstract/overview/content fields; categories for user/agent memories (openviking/session/memory_extractor.py).
• Git log: recent commit mentions memory extraction robustness.
• gsio search results show Codex/Gemini focusing on “context” handling (contrast: delivery structure vs injection/compaction).

Draft outline

• Hook: memory for agents is still a junk drawer; OpenViking treats it like a filesystem.
• What’s new: context database + file system paradigm + scopes.
• How it works: viking:// URIs + L0/L1/L2 layers + CLI endpoints.
• Why it matters: deterministic delivery, observability, self-evolving memory.
• Contrast: Codex/Gemini improve context injection; OpenViking changes the substrate.
• Close: question + call to action.

Sources — 2026-03-15 — OpenViking context filesystem

Web trend scan

• GitHub Trending (snapshot 2026-03-15): https://github.com/trending (OpenViking listed with description and stars today).

GitHub repo metadata (gh)

• gh repo view volcengine/OpenViking --json name,description,stargazerCount,updatedAt,url
• Description: context database for AI agents using filesystem paradigm; stars and updated time.

OpenViking README (core claims)

• daily evidence corpus/2026-03-15-openviking-context-fs/OpenViking/README.md
• Filesystem paradigm unifies memories/resources/skills; tiered context loading (L0/L1/L2); directory recursive retrieval; visualized trajectory; automatic session management (self-iteration).

Viking URI (filesystem model + scopes)

• daily evidence corpus/2026-03-15-openviking-context-fs/OpenViking/docs/en/concepts/04-viking-uri.md
• viking://{scope}/{path} format; scopes (resources/user/agent/session/queue/temp); initial directory structure.

L0/L1/L2 API surface

• daily evidence corpus/2026-03-15-openviking-context-fs/OpenViking/docs/en/api/02-resources.md
• L0 abstract, L1 overview, L2 full content via API/CLI calls.

CLI commands for layers

• daily evidence corpus/2026-03-15-openviking-context-fs/OpenViking/crates/ov_cli/src/commands/content.rs
• abstract_content, overview, read commands map to HTTP client calls.

Memory extraction code (self-evolving memory)

• daily evidence corpus/2026-03-15-openviking-context-fs/OpenViking/openviking/session/memory_extractor.py
• CandidateMemory with abstract/overview/content fields; categories for user/agent/tool memories.

Recent commit (git)

• git log -1 / git show -1 --stat in OpenViking repo
• commit d204bd04c1c7dd19ef9c290adf5d7ad2268ad0a3 fixes memory extraction response handling.

Context comparison (gsio search)

• gsio search q "context" -p openai/codex -p google-gemini/gemini-cli
• Results show Codex/Gemini focus on context injection/compaction; contrasts with OpenViking’s structured context delivery.