Claude Obsidian 2.0: How to Give Your AI Agent a Persistent Memory That Compounds (2026 Guide)

Last verified: 2026-08-09 · Free and open-source (MIT) · Requires Obsidian (free) + Claude Code (free tier available) · v2.1.0 is the latest release

TL;DR

• What it is: A Claude Code plugin that turns a local folder of Markdown files into a compounding knowledge vault your AI agents can read, write, and query — your files stay on your machine, never uploaded to a model.
• Why it matters: Every chat you have with an AI disappears when you close the tab. Claude Obsidian captures, grounds, connects, and reuses that knowledge so your agents get smarter over time instead of forgetting everything.
• Cost: Free. Obsidian is free for personal use (obsidian.md). Claude Code has a free tier; Pro is $20/month (claude.com/pricing).
• Best for: Developers, researchers, and small teams who use AI agents daily and are tired of re-explaining the same context every session.

---

What problem does Claude Obsidian 2.0 solve?

Every AI chat session starts from nothing. You explain your project, your stack, your constraints, and your preferences — and the moment you close the tab, all of that vanishes. Claude Obsidian 2.0 fixes this by giving your AI agents a persistent, local knowledge base that compounds with every use. Instead of re-deriving answers from scratch like traditional RAG (Retrieval-Augmented Generation), the vault compiles knowledge once and keeps it current — a pattern originally proposed by Andrej Karpathy in his April 2026 LLM Wiki gist.

The core insight, as Karpathy puts it: "The knowledge is compiled once and then kept current, not re-derived on every query. This is the key difference: the wiki is a persistent, compounding artifact." Claude Obsidian implements this pattern as a production-ready plugin with 15 skills, source provenance tracking, and crash-recovery — something the raw gist alone doesn't provide.

How does Claude Obsidian's knowledge loop work?

The system runs a four-step loop — capture, ground, connect, use — and each cycle makes the vault smarter. Here's what each step does:

Step 1: Capture

You drop a source file (a PDF, a text file, a web article) into a visible inbox/ folder. Before anything touches the content, Claude Obsidian makes an immutable, content-addressed copy. The original is never modified — a locked snapshot is preserved before any synthesis happens. This means even if a processing step goes wrong, your raw source is untouched.

Step 2: Ground every claim

This is what separates Claude Obsidian from a standard notes app. Every important claim in the vault keeps a provenance pointer back to where it came from. The system maintains a ledger that tracks:

• Source authority (how credible is the origin?)
• Freshness (how old is the information?)
• Support level (how many sources back this claim?)
• Contradiction state (does another source disagree?)
• Confidence and review state

If a claim isn't backed by a real source, it stays flagged. For critical claims, the system prefers two independent sources before treating something as solid. It would rather say "I don't have that" than invent a plausible-sounding citation — a direct counter to the hallucination problem that plagues LLMs.

Step 3: Connect

Instead of a flat pile of notes, you get linked pages, indexes, Maps of Content (MOCs), and Obsidian Canvas views. Every new note is automatically linked to related existing notes, building a knowledge graph that grows denser and more useful over time. You can switch between four filing modes (more on that below) to control how new notes get sorted.

Step 4: Use

You ask the vault a question and it answers only from the evidence already inside it — read-only, no guessing. The retrieval system uses contextual prefixes plus BM25 ranking (with optional cosine rereranking) to pull the right page even when you can't remember what you called it. You can also research, tidy, and fold old activity into summaries.

What are the 15 skills in Claude Obsidian 2.0?

The plugin ships with 15 skills organized into three tiers. Here's every skill and what it does, verified from the GitHub repository:

How do you install Claude Obsidian 2.0?

Installation is a two-step process via the Claude Code plugin marketplace. These steps are from the official install guide:

Step 1: Add the marketplace

claude plugin marketplace add AgriciDaniel/claude-obsidian

Step 2: Install the plugin

claude plugin install claude-obsidian@agricidaniel-claude-obsidian

Step 3: Verify

claude plugin list

Then, in any Claude Code session, type /wiki to start. Place source files in the inbox/ folder and run the wiki-ingest skill to process them.

Alternative: Clone as a vault

git clone https://github.com/AgriciDaniel/claude-obsidian
cd claude-obsidian && bash bin/setup-vault.sh

What do you need to run it?

Obsidian's core app is completely free with no feature limits, no account requirement, and no ads. Optional paid add-ons include Sync ($5/month) and Publish ($10/month), but neither is needed for Claude Obsidian (Obsidian pricing page).

Which filing mode should you pick?

This is the setting most people skip, and it quietly shapes how every new note gets filed. Claude Obsidian supports four filing modes, and you switch between them with the wiki-mode skill. Switching only affects how new notes get sorted — it never tears apart what you've already built.

PARA mode explained

PARA was created by productivity expert Tiago Forte and popularized in his 2022 book Building a Second Brain. It organizes information into four categories by actionability:

• Projects — short-term efforts with a defined goal and deadline
• Areas — long-term responsibilities with no end date
• Resources — reference materials organized by topic
• Archives — completed or inactive items

The principle is simple: organize by what you're actively working on, not by topic. A project you're shipping this week is in Projects; the domain knowledge you'll reference for years goes in Resources. When a project finishes, it moves to Archives (Forte Labs).

Zettelkasten mode explained

Zettelkasten (German for "slip box") was developed by sociologist Niklas Luhmann, who produced over 70 books and 400 articles using it. The core principle is atomicity: one idea per note, uniquely identified, and heavily linked to related notes. The links create a network where new connections emerge as the graph grows — you discover relationships between ideas you wouldn't have seen in a linear outline.

How does Claude Obsidian prevent data loss?

Version 2.0's headline feature is recoverable transactions. This is the safety net that makes running AI agents on your knowledge base trustworthy.

Here's how a single operation works:

1. The system reads the target files and records their expected SHA-256 hashes.
2. Parallel agents (if you're running multiple) return drafts and evidence only — they never write directly.
3. One orchestrator merges the complete change into a single operation bundle.
4. The bundle is inspected, then applied in one atomic step.
5. If anything fails halfway through, the vault rolls back to exactly how it was before.

The core holds a process-lifetime vault lock, journals backups, uses atomic file replacement, and restores prior state if an apply fails. A changed target is treated as a conflict — never a silent overwrite. This means you can let multiple agents work on the vault simultaneously without worrying about them scrambling your notes.

Does Claude Obsidian work with other AI agents?

Yes. While it's built for Claude Code, Claude Obsidian is a local-first knowledge system compatible with any Agent Skills host. The v2.0.0 release added a host-neutral claude_obsidian standard-library core, meaning it works with:

• Claude Code (Anthropic) — primary target
• Codex (OpenAI) — confirmed compatible
• Gemini (Google) — confirmed compatible
• Open-source agents — any Agent Skills-compatible host

Your vault is just a folder of plain Markdown, JSON, and source files. It's not hidden in a plugin cache, locked in a cloud database, or uploaded to a model. If you ever remove the plugin, your notes stay exactly where they are. Network egress — sending data to an API — is always an explicit, separate decision, never automatic.

What changed in version 2.0?

The v2.0.0 release — published July 29, 2026 — was titled "Reliability and Evidence Refoundation." It was a top-to-bottom rebuild focused on three principles:

The release passed 39/39 focused tests, validated all 15 skills with zero findings, and completed a secret review (208 scanner detections classified as deterministic hashes, examples, or negative fixtures — zero real secrets). The public artifact SHA-256 is published for reproducibility.

v2.1.0 (July 31, 2026) added native Windows read-only support — writes still require WSL on Windows, but inspection, dry-runs, and retrieval work natively.

How is this different from regular Obsidian or Notion AI?

The key difference: Notion AI and most "AI notes" tools use RAG under the hood — they retrieve chunks at query time and generate answers, but nothing is compiled or accumulated. Claude Obsidian follows Karpathy's LLM Wiki pattern: knowledge is compiled once, kept current, and compounds. Each query builds on the last instead of starting from scratch.

A practical workflow: building a content knowledge base

Here's how you'd use Claude Obsidian in a real small-business scenario — say, you run a content operation and want your AI agents to know your brand, your audience, and your editorial standards.

1. Install the plugin (two commands, above).
2. Set your mode — run /wiki-mode and pick PARA. Your notes will auto-file into Projects, Areas, Resources, and Archives.
3. Drop sources into inbox/ — your brand guide, past articles, audience research, competitor analysis, style guide. Run wiki-ingest to process them.
4. Ask questions — run wiki-query and ask "What's our tone for technical tutorials?" The vault answers from your own materials, with source pointers.
5. Save insights — when you get a good answer, run save to keep it as a permanent page. Don't dump your whole chat history; save the one insight that matters.
6. Lint periodically — run wiki-lint to catch dead links, orphans, and metadata gaps before your vault quietly rots.
7. Fold old activity — run wiki-fold to roll weeks of vault activity into clean summary pages. Your knowledge base writes its own changelog.

Over time, every agent you run — for writing, research, analysis — draws from this grounded vault instead of guessing. You stop re-explaining your business every morning. The context is just there.

What this means for you

If you use AI agents for real work — writing, coding, research, business operations — the biggest productivity tax is the context you re-explain every single session. Claude Obsidian 2.0 eliminates that tax. Your vault becomes a shared brain that every agent reads from, and it only gets smarter.

For small businesses specifically, the compounding knowledge loop means your AI agents learn your business once and build on it forever. No subscription fees for the plugin, no data lock-in, and no cloud dependency. If you're already using an Agent Operating System to run an AI team, Claude Obsidian is the memory layer that makes it actually remember things between sessions.

Related reading

• build a company brain where knowledge compounds over time

FAQ

Q: Is Claude Obsidian free? A: Yes. The plugin is MIT-licensed and free. Obsidian (the notes app) is free for personal use. Claude Code has a free tier, with Pro at $20/month. You can run the entire system at zero cost on the free tiers.

Q: Do I need an Obsidian subscription to use this? A: No. Obsidian's core app is completely free with no feature limits. Optional paid add-ons (Sync at $5/month, Publish at $10/month) are not required. Your vault is just a local folder of Markdown files — you can open it in any text editor.

Q: Can Claude Obsidian work without Claude Code? A: Yes. It's compatible with any Agent Skills host, including Codex, Gemini, and open-source agents. The v2.0.0 release added a host-neutral core library specifically for this. However, Claude Code is the primary target and has the best integration.

Q: Will the AI mess up my existing notes? A: No. Claude Obsidian never modifies your original sources — it creates immutable copies before processing. All vault writes go through recoverable transactions with SHA-256 verification and automatic rollback. If anything fails, your vault reverts to its prior state.

Q: What happens if I uninstall the plugin? A: Your notes stay exactly where they are. The vault is a normal folder of plain Markdown files on your machine. Nothing walks out the door with the plugin. You can continue using the files in Obsidian, any text editor, or a different agent system.

Q: What's the difference between Claude Obsidian and Karpathy's LLM Wiki gist? A: Karpathy's gist is an idea file — a prompt pattern you copy-paste to your AI agent. Claude Obsidian is a production-ready implementation of that pattern: 15 structured skills, source provenance tracking, crash recovery, multi-agent safety, and filing-mode support. The gist gets you started; Claude Obsidian is the finished product.

Sources

1. AgriciDaniel/claude-obsidian — GitHub repository (10,594 stars, MIT license, v2.1.0)
2. Claude Obsidian releases page — v2.0.0 (July 29, 2026), v2.1.0 (July 31, 2026)
3. Karpathy's LLM Wiki gist — the original pattern (April 2026)
4. Obsidian pricing page — free personal use, Sync $5/mo, Publish $10/mo
5. Claude pricing page — free tier, Pro $20/mo, Max $100–$200/mo
6. Forte Labs — The PARA Method — Tiago Forte's original framework
7. Claude Code plugin marketplace docs — installation reference

Updates & corrections

• 2026-08-09 — Initial publication. All facts verified against primary sources (GitHub repo, releases, Obsidian pricing, Claude pricing, Karpathy gist, Forte Labs). Pricing and version information flagged as volatile — re-check monthly.

---

Every claim here is traced to a primary source, dated, and listed under Sources. Research and drafting are AI-assisted; editing, verification and publication are human decisions, and a person is accountable for what appears on this page. How we work →