Absorbed into Qumio as part of its memory layer. While building Qumio's Obsidian integration, the overlap became clear. Vault access through kepano's agent skills and the Obsidian CLI covered the core of what Archivian was going to do.

Original idea (March 2026)

The idea

My Obsidian vault is a mess. Notes from different projects, scattered tags, inconsistent naming, files in the wrong folders. I reorganize it manually every few weeks and it drifts back within days.

I wanted an AI agent to do it. Not a plugin that runs inside Obsidian, but a skill file that teaches any compatible agent how to scan, analyze, and restructure an entire vault.

What it would do

Scan a vault's notes, frontmatter, tags, and internal wikilinks
Reorganize files into folders based on user-defined rules and AI inference
Rename files to consistent conventions
Suggest or apply tags based on content
Fix broken wikilinks after moves
Generate index notes (Maps of Content) for folders
Dry-run mode to preview everything before applying

Why an Agent Skill

Archivian follows the Agent Skills specification. Agent Skills can range from simple prompt files to full CLI tools with APIs. Archivian will include TypeScript tooling for vault scanning, wikilink processing, and change planning.

Any compatible agent (Claude Code, OpenCode, etc.) can invoke it. No Obsidian plugin install. No Obsidian API dependency. It works directly on the vault's markdown files through the filesystem.

How it will be built

Same process as Qumio. Claude Code with the GSD (Get Shit Done) framework handling the discuss/research/plan/execute/verify cycle. The research phase is already done: stack decisions, architecture patterns, and pitfalls are documented.

The planned architecture is a four-phase pipeline: scan the vault into an in-memory graph, analyze notes against rules (with AI inference as fallback), build an immutable change plan, then execute or dry-run. Rules get applied first. The agent's own LLM only gets called for notes that don't match any explicit rule.

All AI inference runs through the invoking agent's model. No separate model embedded. If you run it from Claude Code, it uses Claude. If Qumio picks it up as an OpenClaw skill, it runs on the local model through Ollama, same as everything else.