Verdict: You can give Claude a real, persistent memory system in about 10 minutes using claude-obsidian — a free, MIT-licensed Claude Code plugin that turns a local folder of Markdown files into a self-organizing knowledge graph your AI reads, links, and cites across every session. No cloud subscription, no proprietary database, and no data leaving your machine unless you choose to sync. It is the most practical answer to the "AI goldfish problem" — the one where your assistant forgets your business every Monday morning — and it gets better the more you use it.
Last verified: 2026-08-01 · Cost: $0 (plugin is MIT-licensed; Obsidian is free for personal use) · Setup time: ~10 minutes · Works with: Claude Code, plus Codex, OpenCode, and Gemini CLI via multi-agent setup · Volatile facts: Plugin version, star count, and skill list may change — last checked August 2026
What Is Claude Obsidian and Why Does Claude Need It?
Claude Obsidian is an open-source Claude Code plugin that gives your AI assistant one critical capability it was never designed for: a persistent, local knowledge base that compounds with every interaction. The plugin lives on GitHub at AgriciDaniel/claude-obsidian, has over 10,000 stars and 1,179 forks as of August 2026, and is built on Andrej Karpathy's LLM Wiki pattern — a concept the former Tesla AI director published as a public gist in April 2026 for turning LLMs into knowledge-base maintainers (source: Karpathy's llm-wiki gist).
Anthropic introduced persistent memory for Claude.ai in March 2026 — letting Claude retain your name, preferences, and project details across conversations (source: Claude memory documentation). Claude Code also has built-in auto-memory that saves notes about your projects in ~/.claude/projects/. But these built-in systems have limitations: memory is limited in scope, the storage is somewhat opaque, and — critically — nothing connects your accumulated knowledge into a navigable, source-cited graph you can audit and trust.
Claude Obsidian fixes this by making your knowledge base explicit: plain Markdown files in a folder on your computer, linked together, with every claim pointing back to its original source. You can open, edit, search, and verify everything yourself. The vault is never hidden in a plugin cache or locked in a cloud database — it is ordinary files you own.
How Does the Compounding Knowledge Loop Actually Work?
The system is built around a four-step cycle that Claude Obsidian calls the "compounding knowledge loop." Each step feeds the next, and the vault gets richer every time you use it — without you doing the manual filing work.
1. Capture with context
You drop sources — web articles, documents, notes, research papers, or even raw brain dumps — into an inbox/ folder inside your vault. Claude reads them, creates an immutable copy in a .raw/ directory (so the original source survives even if the web page disappears), and generates structured wiki pages from the content. The inbox is a visible folder, not a black box — you can see exactly what's waiting to be processed.
2. Ground every claim
This is the differentiator that makes the system trustworthy. Every claim in the vault carries a source/claim ledger that tracks its authority, freshness, support level, contradiction status, confidence, and review state. When Claude answers a question from the vault, it points back to the specific source. If the vault doesn't have evidence for something, it says so — instead of making something up. This addresses one of the biggest risks with AI-assisted knowledge work: the silent drift between what the AI confidently tells you and what is actually true.
3. Connect knowledge
Pages are linked together with standard Obsidian-style [[wikilinks]], organized into indexes, Maps of Content, and — optionally — visual Canvas boards. The system supports four filing methodologies out of the box:
| Method | Best for | Structure |
|---|---|---|
| Generic | Quick starts, no methodology preference | Flat wiki with basic indexes |
| LYT (Linking Your Thinking) | Progressive idea development | Map of Content → nested MOCs |
| PARA | Action-oriented knowledge (projects, areas, resources, archives) | Four-folder active/archive structure |
| Zettelkasten | Academic research, atomic ideas | One idea per note, dense cross-links |
You switch methodologies with a single command — wiki-mode reroutes new content into the chosen layout through a small Python router, so you are never locked into one structure.
4. Use the vault again
Once knowledge is captured, grounded, and connected, you query it. Claude searches relevant pages, reads them, and synthesizes answers with citations. Good answers get filed back into the wiki as new pages — so your explorations compound just like ingested sources do. The vault also tracks a wiki/log.md of all activity, and a wiki/hot.md cache of frequently needed answers for fast retrieval.
What Are the 15 Skills and How Do They Work Together?
Claude Obsidian ships as a Claude Code plugin with 15 skills — discoverable instruction files (each a SKILL.md with YAML frontmatter and operational guidance) that Claude invokes based on the task. The skills are namespaced under the /claude-obsidian: prefix, so they never collide with other plugins you may have installed (source: Claude Code plugin documentation).
Build and use the wiki (5 core skills)
| Skill | Command | What it does |
|---|---|---|
wiki |
/claude-obsidian:wiki |
Diagnose vault health, check readiness, route work to the right workflow |
wiki-ingest |
/claude-obsidian:wiki-ingest |
Read sources from inbox/, create linked pages + provenance records |
wiki-query |
/claude-obsidian:wiki-query |
Answer questions from vault evidence with citations (read-only) |
save |
/claude-obsidian:save |
Explicitly save one scoped answer or insight (never automatic — you approve) |
wiki-lint |
/claude-obsidian:wiki-lint |
Reports dead links, orphaned pages, metadata gaps, and stale indexes |
Extend the workflow (7 skills)
| Skill | What it adds |
|---|---|
autoresearch |
Bounded web research with explicit egress tracking and canonical merge into the vault |
canvas |
Create and maintain Obsidian Canvas boards — visual whiteboard maps of topics, projects, or clients |
defuddle |
Clean web content (remove boilerplate/ads) before ingestion |
wiki-fold |
Extractive, traceable rollups of the operation log into weekly/monthly summaries |
wiki-mode |
Switch between Generic, LYT, PARA, or Zettelkasten filing methodologies |
wiki-retrieve |
Contextual prefix retrieval, BM25 ranking, and optional cosine reranking for better recall |
wiki-cli |
Obsidian CLI integration for reads/search and transaction-safe writes |
How to Set Up Claude Obsidian in 10 Minutes
What you need before starting
- Claude Code installed with an active Anthropic subscription (Pro, Max, or Team — starting at $20/month). If you are looking for free alternatives to access Claude Code, see our guide on how to use Claude Code for free.
- Obsidian (optional but recommended) — the free app is your visual interface to the vault. Download it from obsidian.md. The core app is free for personal use; optional Sync costs $5/month and Publish costs $10/month (source: Obsidian pricing). You do not need any paid plan or community plugin for this setup.
- Python 3 on your machine (the plugin's scripts are Python-based).
- Git (the vault is git-friendly, so you can version-control your entire knowledge base).
Step 1: Clone the plugin
git clone https://github.com/AgriciDaniel/claude-obsidian.git
cd claude-obsidian
Step 2: Initialize your vault
Claude Obsidian uses a two-step approval process for all write operations — it generates a plan, shows it to you, and waits for your explicit approval before touching any files. This is a safety feature: the vault is never modified without your review.
export GENERATED_AT="$(date -u +%Y-%m-%dT%H:%M:%SZ)"
export OPERATION_ID="init-reviewed"
python3 scripts/claude-obsidian.py init "$HOME/Documents/MyKnowledgeVault" \
--generated-at "$GENERATED_AT" --operation-id "$OPERATION_ID"
Review the JSON plan it outputs, copy the approved_plan_sha256 value, then apply:
python3 scripts/claude-obsidian.py init "$HOME/Documents/MyKnowledgeVault" \
--generated-at "$GENERATED_AT" --operation-id "$OPERATION_ID" \
--approved-plan-sha256 "<sha256-from-the-plan>" --apply
Step 3: Start Claude from the vault
cd "$HOME/Documents/MyKnowledgeVault"
claude --plugin-dir /absolute/path/to/claude-obsidian
That is it. Claude now has access to all 15 skills. Try these commands to get started:
/claude-obsidian:wiki— check your vault's health and get routed to the right workflow/claude-obsidian:wiki-ingest— process sources you dropped ininbox//claude-obsidian:save— save a valuable answer to the vault/claude-obsidian:wiki-query— ask the vault a question
Step 4 (optional): Set up multi-agent access
Claude Obsidian is not limited to Claude Code. The plugin's AGENTS.md documents that skills are compatible with any Agent Skills-aware host — including Codex, OpenCode, and Gemini CLI. To wire it up:
bash bin/setup-multi-agent.sh --host codex
bash bin/setup-multi-agent.sh --host codex --apply
Repeat with --host opencode or --host gemini for other agents. This means your knowledge base works across multiple AI tools — not just one vendor's product.
How Does Claude Obsidian Compare to Built-in Memory Options?
| Feature | Claude.ai Memory | Claude Code Auto-Memory | Claude Obsidian |
|---|---|---|---|
| Storage | Cloud (Anthropic servers) | ~/.claude/projects/ (local files) |
Your own Obsidian vault (local Markdown) |
| Transparency | Shows when memory is used, with citations | Some visibility into saved notes | Fully inspectable — every file, link, and source is visible |
| User editability | Full direct edit of project memory | Editable in .claude/ directories |
Full direct edit in any text editor or Obsidian |
| File format | Proprietary | Markdown files | Plain Markdown + Obsidian Flavored |
| Linking | None — isolated memory entries | Basic file storage | Bidirectional wikilinks, graph view, Canvas |
| Source tracking | Memory includes provenance tag | No explicit sourcing | Source/claim ledgers with authority, freshness, support, contradiction, confidence, and review state |
| Multi-agent | No — Claude only | No — Claude Code only | Yes — Claude Code, Codex, OpenCode, Gemini |
| Visual interface | Web UI | None | Obsidian graph, Canvas boards, Dataview |
| Cost | Included in Claude subscription | Included in Claude Code | Free (MIT-licensed plugin + free Obsidian app) |
The key distinction: built-in memory writes about you behind the scenes. Claude Obsidian writes for you, in files you own, in a structure you can navigate and verify. Think of it as the difference between a diary your assistant keeps in a drawer versus one that sits open on your desk with footnotes.
If you are already managing Claude Code token consumption, note that a structured vault can reduce the context you need to paste into each session — see our guide on free tools to cut Claude Code tokens.
The Trust Architecture: How Does It Prevent Data Loss?
Claude Obsidian is designed for people who treat their knowledge base as infrastructure. The trust model has several layers worth understanding before you commit your data to it:
Explicit vault selection
The plugin will not guess where your vault is. It selects the vault using CLAUDE_OBSIDIAN_VAULT, the nearest .claude-obsidian.json config file, or one unambiguous initialized ancestor directory. If selection is uncertain, the plugin exits without writing — a fail-safe default that prevents accidental writes to the wrong directory.
One operation = one recoverable transaction
Every write operation — ingesting a source, saving an answer, updating the index — is treated as a single logical unit. An orchestrator applies one transaction at a time. Parallel agents return drafts rather than writing directly, and one orchestrator applies the recoverable transaction. This prevents race conditions when multiple agents are working in the same vault simultaneously.
Immutable source preservation
When the system ingests a source, it creates a content-addressed copy in .raw/ before generating any wiki pages. The original source material survives even after synthesis — so you can always go back and verify what the AI actually read versus what it wrote.
Network egress is explicit
The default is local. If a skill needs to make a web request (like autoresearch), it declares that egress explicitly. The plugin does not silently upload your vault contents to a model or external service unless you configure it to do so.
Claude Obsidian vs. Other AI Memory Approaches: Which Should You Pick?
The "AI persistent memory" space has several approaches in 2026. Here is how they compare for practical use:
| Approach | What it is | Best for | Key limitation |
|---|---|---|---|
| Claude Obsidian | Local Markdown vault + Claude Code plugin | People who want full ownership, source-cited knowledge, and multi-agent support | Requires Claude Code (or compatible agent); no mobile-native experience (sync via Obsidian Sync or Git) |
| Claude.ai Projects + Memory | Anthropic's built-in persistent workspaces | Quick start, no setup, cloud convenience | Locked to Anthropic's platform; limited file capacity (~200k tokens); memory is somewhat opaque |
| Custom GPTs + Obsidian (manual) | Obsidian vault + Claude via copy-paste or MCP | Coders who want fine-grained control | Manual or semi-automated; no built-in source tracking or ingestion pipeline |
| NotebookLM | Google's source-grounded notebook tool | Document Q&A with citation | Not a persistent knowledge base — queries don't accumulate; no local ownership |
If the overriding priority is speed and cloud convenience, Claude.ai's built-in Projects + Memory is the path of least resistance. If the priority is owning your knowledge, citing sources explicitly, and having it work across multiple AI tools, Claude Obsidian is the stronger long-term system.
For a deeper look at how persistent context changes developer workflows, including how Claude's built-in project memory shapes responses, check our Hermes Agent pro tips and configuration tricks.
What This Means for You
For solopreneurs and small business owners: This system lets your AI agent maintain a knowledge base of your business context — your products, pricing, client notes, and decisions — without you re-explaining it every session. Drop a doc in the inbox, and Claude files it, links it, and remembers it. After a few weeks, asking "What are some personalized content ideas for my business?" yields answers grounded in your actual accumulated context, not generic suggestions. If you are dealing with session sprawl — context fragmenting across dozens of disconnected chats — our guide on the one AI agent mistake that sabotages your business covers the broader problem.
For developers and AI builders: Claude Obsidian is one of the most complete open-source implementations of Karpathy's LLM Wiki pattern — and it's genuinely useful as a reference for building your own agent memory architecture. The source/claim ledger model, the progressive disclosure query path (hot cache → index → pages), and the BM25 + optional cosine reranking pipeline are all worth studying. The plugin's multi-agent setup means you can maintain one knowledge base across Claude Code, Codex, and other tools, rather than fragmenting context across vendor-specific silos.
For researchers and knowledge workers: The PARA or Zettelkasten filing methodologies, combined with the autoresearch skill (bounded web research with explicit egress tracking), make this a legitimate research tool — not just a note-taking app with AI bolted on. The wiki-fold skill rolls weeks of vault activity into traceable summaries, so your knowledge base writes its own changelog. This is the kind of compounding value that manual PKM systems lose when the maintenance burden overwhelms the user.
FAQ
Q: Is Claude Obsidian actually free?
A: Yes. The plugin is MIT-licensed and lives on GitHub at AgriciDaniel/claude-obsidian — free to clone, use, and modify. Obsidian's core app is free for personal use (no sign-up required, no feature limitations). The only cost is a Claude Code subscription (starting at $20/month) or a compatible agent CLI. Optional Obsidian Sync ($5/month) and Publish ($10/month) are add-ons you do not need for this setup.
Q: Do I need Obsidian installed to use the plugin?
A: No. Obsidian is optional. The vault is just a directory of Markdown files — you can read, edit, and search them in any text editor or terminal. Obsidian adds the graph view, Canvas whiteboard, and a polished visual interface, but the plugin fully works without it. Transport detection in the setup scripts prefers the Obsidian CLI where available and falls back to direct filesystem access.
Q: Can I use Claude Obsidian with ChatGPT or Gemini instead of Claude?
A: Claude Obsidian is primarily a Claude Code plugin, but the skills format is compatible with any Agent Skills-aware host. The repo includes setup-multi-agent.sh for Codex, OpenCode, and Gemini CLI. You get one knowledge base that works across multiple AI tools, rather than being locked to one vendor.
Q: How is this different from just using Claude Code's /init command?
A: The /init command generates a CLAUDE.md file based on a one-time scan of your codebase. Claude Obsidian creates a persistent, evolving knowledge system that grows with every interaction — capturing sources, linking pages, tracking source/claim provenance, and compounding answers back into the vault. /init is a snapshot; Claude Obsidian is a growing library.
Q: What happens if the plugin becomes unmaintained?
A: Your vault is safe. It is plain Markdown files — no proprietary format, no lock-in. You can continue reading, editing, and using the knowledge base with any text editor or AI tool even if the plugin is abandoned. The .raw/ sources and wiki/ pages are ordinary files that outlive any specific tool.
Q: How much manual work does the system require?
A: Minimal. You drop sources into the inbox/ folder and Claude reads, links, and files them. You approve saves explicitly (memory is never written automatically without your say-so). The main human tasks are curating good sources, directing the analysis, and periodically running wiki-lint to check for dead links, orphaned pages, and stale indexes. The bookkeeping — the part humans abandon manual wikis over — is automated.
Discussion
0 comments