obsidian-knowledge

Obsidian Knowledge

A Claude Code plugin marketplace with skills and hooks for maintaining Obsidian knowledge base vaults.

Skills

vault-organizer

Maintains vault organization through a single-pass pipeline:

• Sync indexes — creates and updates index.md files with thin pointer entries for managed zones
• Organize files — moves misplaced files to appropriate locations using the Obsidian CLI
• Rename ambiguous files — detects files with non-descriptive names (device-generated, hash-based, generic labels), reads their content, and renames them following the vault's naming conventions from CLAUDE.md
• Fix broken links — detects unresolved links and resolves them when possible, flags ambiguous cases for human review
• Report issues — maintains a needs-attention.md worklist for issues requiring human judgment
• Regenerate reports — rewrites reports/open-questions.md from > [!question] callouts across wiki/, giving agents and humans a single place to see unresolved questions flagged in prose

remember-conversations

Files valuable conversation outputs as permanent vault notes and updates the changelog, so insights compound rather than disappearing into chat history:

• Session notes — two types: -diary for narrative accounts (what happened, what was tried) and -convo for analytical synthesis (comparisons, decision rationales, research summaries)
• Changelog updates — creates a terse per-session file in changelog/ summarizing actions taken
• Automatic placement — notes filed in sessions/ subfolders within the relevant subtree, preserving progressive disclosure
• Stop hook integration — a reminder nudges the agent to file sessions at the end of each conversation

improve-harness

Disabled as of 3.22.15. The command shim and skill index are retained as .disabled files for reference, but they are no longer discoverable by the plugin runtime.

deploy-harness

Single-purpose skill: merge an approved improve/<slug> branch into main, bump patch version, push to origin. Invokable manually for approved harness changes.

Hooks

Vault protection (PreToolUse)

protect-vault.py runs before every Bash, Write, and Edit tool call. It provides four layers of safety:

Read-only _sources/ directories. Folders named _sources/ anywhere in the vault tree are protected from agent writes. These typically hold irreplaceable originals (tax records, legal filings, vital docs, property deeds). Agents can read them to generate summaries and indexes, but cannot create, modify, rename, move, or delete files inside them.

Destructive command guards. Recursive rm and mv targeting paths that appear to be inside an Obsidian vault are blocked.

Published file guard. Write and Edit to any vault file with dg-publish: true in its frontmatter are blocked — edits to published files go live on the website and require explicit user confirmation.

Auto-memory redirect. Agents are blocked from writing operational knowledge (feedback_*.md, project_*.md, reference_*.md) to their per-project auto-memory. Auto-memory is a silo invisible to other sessions, other tools, and vault search. The hook redirects this knowledge to a standardized vault location, scoped to the agent's current context:

• Inside a git repo → <vault>/wiki/repos/<owner>/<repo>/memory/ (parsed from git remote get-url origin; portable across hosts).
• Outside any repo → <vault>/wiki/systems/machines/<hostname>/memory/ (host-scoped fallback).

Each memory dir mirrors Claude's native layout (MEMORY.md index plus one feedback_*.md / project_*.md / reference_*.md per fact), so the workflow is identical — only the storage location changes. The SessionStart primer surfaces the resolved path so agents know where to read and write at session start. Migrate existing per-project memories with scripts/migrate_claude_memory.py.

Escape hatch. Prefix a Bash command with I_AM_BEING_CAREFUL=1 to bypass the _sources/ and published-file guards after the user explicitly confirms. The auto-memory redirect has no escape hatch — write to the wiki instead.

For agents: add this line to the user's primary CLAUDE.md

Built-in Claude auto-memory under ~/.claude/projects/*/memory/ is replaced by the obsidian-knowledge plugin. Agents should add a single pointer to the user's primary memory file (~/.claude/CLAUDE.md) so sessions on hosts without this plugin still know where to look:

Agent memory is managed by the obsidian-knowledge plugin; refer to it
for memory location, read/write conventions, and session-start recall.

Stop hooks

Stop hooks fire at the end of each Claude Code turn. Each checks whether the working directory is inside a configured vault root. They have a 5-minute cooldown per session to avoid being noisy in long conversations.