Shared Memory for AI Agents in Obsidian: How Claude Obsidian and the Official Grok Bot Plugin WorkNote · digest for November 21
Isometric visualization of a note graph and shared knowledge base for AI agents
Note3 sections≈ 4 min

Shared Memory for AI Agents in Obsidian: How Claude Obsidian and the Official Grok Bot Plugin Work

How many times this week have you re-explained your project, your client tone, and yesterday's decisions to an AI? Every fresh chat opens with a blank slate. AI doesn't have a memory problem—it has a filing problem. Good breakdowns stay buried in old chats nobody ever reopens.

The fix is simple: a local folder of .md files on disk, opened in Obsidian. This Markdown vault gives humans a visual note graph and gives agents shared memory you can inspect with your own eyes.

From one-off search to Karpathy's living wiki

Most setups bolt on classic RAG—search that re-scans raw documents from scratch on every question. In April 2026, Andrej Karpathy popularized the LLM Wiki pattern: the model maintains a connected wiki of notes (raw/, entities/, topics/, plus index.md, log.md, and CLAUDE.md). Here, Obsidian is the IDE, the LLM is the programmer, and the wiki is the codebase.

Take a concrete example. At night, you drop a messy note into the folder: Mira wants six pages instead of eight with a draft by Friday, Dan's community launch moved to January 12, a Shopify ceramics shop needs 12 product cards, and three leads came in. Within minutes, the agent updates client cards, the launch plan, and the decision log.

Next morning, you need to reply to Mira. Instead of loading a huge wall of context, the agent reads five small files (her card, decision log, style rules, and templates), drafts a crisp "Two things, both short" email, and flags that "Friday" could mean two different dates. It burns a fraction of the tokens.

Local Markdown vault and Obsidian link graph (Source: obsidian.md)

That leads to one simple rule: stop feeding models giant prompts every day—give them a tidy, linked folder.

Claude Obsidian 3.0.1: source checks and a two-command setup

The leading open-source option for Claude Code is AgriciDaniel/claude-obsidian by Agrici Daniel (MIT). You'll need Claude Code, Python 3.11+, and Obsidian.

After the v2.0 overhaul and October's v3.0.1 release, the focus shifted to trust. A source/claim ledger ties every fact back to its original file. Queries and audits run in read-only mode: the agent answers questions or hunts for broken links, empty sections, and orphan notes without overwriting anything. Writes happen as crash-recoverable transactions without hidden git commit calls. Version 3.0.1 also restored /claude-obsidian:<skill> commands, fixed Unicode filenames on macOS/iOS, added hybrid search, and included an optional local layer via Ollama.

Inside are 15 skills (five core, seven extended like Canvas and changelogs, and three reference skills) that also link into Codex, OpenCode, Gemini, Cursor, and Windsurf. Setup takes a minute:

# Add the plugin marketplace and install it in Claude Code
claude plugin marketplace add AgriciDaniel/claude-obsidian
claude plugin install claude-obsidian@agricidaniel-claude-obsidian

# Move into your test vault and start a session
cd /path/to/my-test-vault
claude

From there, five commands drive the workflow: /claude-obsidian:wiki previews structure before writing, /claude-obsidian:wiki-ingest processes three to five files from Inbox, /claude-obsidian:wiki-query answers without edits, /claude-obsidian:save stores a useful breakdown, and /claude-obsidian:wiki-lint catches broken links. No magic—just a clear pipeline.

Grok Bot, Obsidian CLI, and the daily log habit

The second path connects your vault to Grok Bot via the built-in Obsidian CLI (available in Obsidian 1.12.4+ alongside the official obsidianmd/obsidian-skills pack). Enable Settings → General → Command line interface, add the tool to your PATH, and restart your terminal—the CLI acts as a remote control for the running app. In Grok Bot, open Connect Apps → Obsidian → Connect and pick a test vault like GrokBrain.

The main safety rail is the approval card: the bot previews the exact text of a new note and waits for you to click Allow. In a demo, when adding a lead named Northwind Coaching, the bot spotted a niche overlap with Mira, warned of a conflict of interest, and merged three duplicates.

To keep bots from repeating each other's work, store a daily Done / Stuck / Next log in the same folder. An SEO bot logs two published posts ("Don't write these again") plus items awaiting review, and a link-building bot starts its morning by reading yesterday's log.

One honest caveat: on launch day, seven extra skills for Grok Bot (Bases, Canvas, web clipping) failed to load in the demo—only basic reading and writing worked. And for both tools, the same rule holds: a vault is only as good as the clean inputs you feed it. Start with an empty folder and three to five working files, not five years of personal archives.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 5