How a markdown vault became the operating system for the thinking machine age
There is a quiet revolution happening inside millions of local folders. Obsidian — the privacy-first, markdown-based knowledge tool that crossed 1.5 million users in early 2026 — was already beloved for its bidirectional links, graph view, and obsessive respect for data ownership. But something has shifted in the last year. The arrival of capable AI coding agents, open-source embedding models, and a new generation of “agent skills” has transformed Obsidian from a note-taking app into something closer to a cognitive infrastructure layer: a substrate on which AI agents think, remember, and reason alongside you.
This article maps that transformation. We look at what Graphify and obsidian-second-brain actually do under the hood, survey the essential plugin landscape for Zettelkasten practitioners, and ask what it means when your notes become not just a record of thinking, but a living knowledge graph an AI can traverse.
The Core Problem: Session Amnesia and the Dead Vault
Two problems have plagued knowledge workers who use LLMs:
Session amnesia. Every new conversation starts from zero. You re-explain your project, your constraints, your decisions. The AI has no memory of the brilliant architecture discussion you had last Tuesday.
The dead vault. You have hundreds or thousands of notes in Obsidian. They just sit there. Links go unmade. Connections that should emerge never do, because manually maintaining a dense Zettelkasten across years of notes is a task that defeats most humans.
AI changes both of these — but only if the tooling is right. The tools surveyed below represent the current state of the art in solving these two problems.
Graphify: From Raw Files to a Queryable Knowledge Graph
Graphify is an open-source knowledge graph skill for AI coding assistants — Claude Code, OpenAI Codex, Cursor, Gemini CLI, and others. Its premise is elegant and important.
When you feed raw files to an AI assistant, you are essentially pointing a flashlight at your data: narrow, token-heavy, inefficient. Graphify replaces that flashlight with something closer to a satellite — a persistent, queryable map of your entire knowledge space, served to the AI as compressed subgraphs on demand.
How it works
Graphify combines three techniques into a single pipeline:
Static analysis with Tree-sitter. For code files, Graphify extracts abstract syntax trees (ASTs), call graphs, and docstrings. It understands the structure of your code, not just its text.
Semantic extraction with LLMs. For prose — Markdown files, PDFs, research papers, even images and diagrams — an LLM extracts concepts, entities, and relationships. This is the layer that transforms “words on a page” into structured nodes in a graph.
Graph clustering with Leiden/NetworkX. Extracted nodes and edges are organized using Leiden community detection, grouping related concepts into clusters. The result is a graph.json file that encodes not just what your content says, but how everything relates.
When your AI assistant calls /graphify, it doesn’t re-read your files. It queries the graph and receives a compressed subgraph — the minimal relevant context for your current question. The reported token reduction is striking: up to 71.5x fewer tokens per session compared to naive file feeding.
Graphify is multimodal from the ground up. It processes .py, .js, .go, .java, Markdown, SQL schemas, R scripts, shell scripts, PDFs, and images through a unified pipeline. App code, database schema, and infrastructure documentation can all live in one graph.
Andrej Karpathy recently described his personal knowledge workflow as “dumping papers, screenshots, and tweets into a raw folder, then using an LLM to compile everything into a wiki.” He ended his description with a challenge: “I think there is room here for an incredible new product.” Graphify is a direct answer to that challenge.
Graphify + Obsidian
The natural pairing is using Graphify for structural, code-level knowledge (your codebase, schemas, infrastructure) and Obsidian for the human-readable wiki layer above it. In the claude-code-memory-setup pattern, a single centralized Obsidian vault acts as Claude Code’s persistent second brain — storing decisions, context, progress, and knowledge — while Graphify maps the underlying code structure. Obsidian notes follow Zettelkasten conventions: atomic, densely interlinked, with standardized frontmatter. Claude Code accesses the vault through CLAUDE.md and custom skills, and a git commit hook automatically rebuilds the graph after each commit.
obsidian-second-brain: The Vault That Runs While You Sleep
Where Graphify focuses on knowledge graph extraction, obsidian-second-brain by Eugeniu Ghelbur takes a different angle: it turns your Obsidian vault into an agentic system, one that can autonomously perform research, synthesize knowledge, and update itself on a schedule.
The project is a cross-CLI skill — it works across Claude Code, Codex CLI, Gemini CLI, and OpenCode — distributed as a set of 32 slash commands plus optional Python scripts. The distinction from a conventional plugin is architectural: an Obsidian plugin is constrained by Obsidian’s TypeScript API. A Claude Code skill is constrained only by what Claude can do in your shell. That is a much larger surface area.
What it enables
Vault-first research. Commands like /x-read, /x-pulse, /research, /research-deep, and /youtube pull live content — X posts via Grok, web research via Perplexity Sonar, YouTube transcripts — directly into your vault as AI-first notes, complete with preambles, frontmatter, recency markers, and verbatim sources. Crucially, the research is vault-first: before fetching anything new, it scans your existing notes and fills only genuine gaps.
Cross-temporal synthesis. Claude scans your daily notes, project logs, and decision records across the full history of your vault. It surfaces patterns you never named: “You mentioned ‘onboarding friction’ in four unrelated projects. You never connected them. Onboarding is your actual bottleneck.”
Memory as a decision audit trail. When you propose an architectural change, the system finds your own past notes about why a similar approach failed. It says: “Your 2025 post-mortem says the Rust rewrite failed. Your decision log commits to TypeScript for two years. Still want to proceed?” This is knowledge management as active governance, not passive filing.
Scheduled agents. Agents can be set to run autonomously — updating notes, synthesizing new research, flagging stale information — while you work on other things.
The installation is a single line that clones the repo to ~/.claude/skills/obsidian-second-brain and symlinks the slash commands into ~/.claude/commands/. Restart Claude Code and the skill loads automatically on any session that touches an Obsidian vault.
The Core Plugin Stack: AI Inside the Vault
Beyond the agent-layer tools, a robust set of plugins brings AI directly into the Obsidian editing experience. Here is what matters for serious Zettelkasten practitioners.
Smart Connections — Semantic links over keyword matching
Smart Connections is arguably the most important AI plugin for Zettelkasten work. Instead of keyword search, it embeds your notes as vectors and surfaces semantically related content: ideas that mean similar things, even when they share no common words. For anyone maintaining a Zettelkasten across domains — technical notes alongside philosophical reading, project logs alongside theoretical frameworks — this is transformative. The plugin supports local models via Ollama as well as cloud APIs, making it viable for privacy-conscious setups. It has accumulated over 4,300 stars on GitHub and is one of the most active projects in the Obsidian ecosystem.
Copilot for Obsidian — Conversational AI grounded in your vault
Copilot brings a ChatGPT-style chat interface into Obsidian, with a critical difference: it can read your vault as context. Instead of copy-pasting notes into an external chat window, you ask questions like “What did I conclude about embedding models last month?” and get answers drawn from your actual content. It supports Claude, GPT-4, Gemini, and local models through Ollama, with over 5,700 GitHub stars and active development.
Text Generator — LLM-powered writing inside the editor
Text Generator supports OpenAI, Anthropic, Google, and local models, and operates directly in the editor context. The practical uses for Zettelkasten are concrete: expanding a bullet-point fleeting note into a permanent note, generating a synthesis from multiple linked notes, creating a literature note from a rough summary. It turns the Zettelkasten workflow’s most friction-heavy step — the transition from raw capture to polished atomic note — into something much faster.
InsightA — Atomic notes from long-form content
InsightA directly addresses the Zettelkasten practitioner’s perennial challenge: you have a long article or paper, and you need to distill it into atomic, interlinked permanent notes with a Map of Content. InsightA uses an LLM to perform this transformation automatically — breaking long-form content into structured, interconnected notes explicitly inspired by the Zettelkasten method. For researchers processing large volumes of literature, this is a significant time multiplier.
Notemd — Auto-generated wiki-links and concept notes
Notemd integrates with multiple LLMs to process notes and automatically generate wiki-links for key concepts, creating corresponding concept notes in the process. It can also perform web research to enrich those concept notes. For a Zettelkasten, this addresses the link-creation bottleneck: identifying which concepts in a new note deserve their own permanent note entries and which existing notes they connect to.
Smart Composer — Context-aware writing with vault references
Smart Composer provides AI-powered writing assistance that is aware of your vault’s context. It allows you to reference specific notes as context while composing, bringing relevant existing knowledge into the generation process. The result is writing that builds coherently on what is already in your system.
The Agent-Skill Ecosystem: Beyond Plugins
A new category has emerged alongside the plugin ecosystem: standalone agent skills that treat Obsidian as infrastructure rather than application.
claude-obsidian (Agrici Daniel) implements the full Karpathy LLM Wiki pattern with ten specialized skills, two parallel agents, and a hot-cache system. The hot cache is stored as wiki/hot.md — roughly 500 words of recent session context that is silently loaded at the start of every new conversation. The result is zero re-explanation overhead. Entities, concepts, and sources are automatically classified and cross-referenced, with contradictions flagged via [!contradiction] callouts.
COG-second-brain takes a self-evolving approach inspired by Garry Tan’s gstack and gbrain patterns. It ships with 17 Claude Code skills, six worker agents (Sonnet for I/O, Opus for reasoning), seven role packs for personalized recommendations, and a people CRM layer. The underlying philosophy is explicit: just .md files, any AI agent, zero maintenance. The Zettelkasten method, PARA organization, and GTD capture are all baked into the folder conventions.
obsidian-ai-second-brain (jamesmcroft) is a structured starter template built around the CODE/PARA methodology and GitHub Copilot skills. It enforces machine-readable vault conventions — consistent templates, typed frontmatter, predictable section layouts — so that AI agents can reliably traverse and extend the knowledge graph. The core insight is that structure enables AI, and AI in turn strengthens structure: every skill-driven operation adds cross-links and fills gaps.
Obsidian’s Own Strategic Direction
It would be a mistake to treat all of this as purely community-driven. Since February 2026, Obsidian ships with an official CLI that exposes over 100 commands — search, note creation, daily note management, content appending — all accessible from the terminal. Any AI agent with shell access can now operate on your vault directly, without MCP servers or additional plugins:
obsidian search query="authentication decision"
obsidian daily:append content="Pending task"
obsidian create --title="ADR: New API" --vault=my-vault
Steph Ango (CEO of Obsidian) has released Obsidian Skills, a set of official agent skill definitions. This is not a community hack. It is a product direction signal: Obsidian is evolving from a note-taking app into a lightweight operating system for knowledge work, with AI as the coordination layer.
A Note on Zettelkasten Specifically
The Zettelkasten method demands things that are hard to do at scale by hand: strict atomicity (one idea per note), dense interlinking, consistent metadata, and regular synthesis. These are precisely the things AI excels at maintaining. The emerging pattern is a division of labor:
- Human responsibility: choosing what to capture, what matters, what to push back on, what the AI got wrong
- AI responsibility: atomization, link suggestion, metadata enforcement, gap detection, cross-temporal synthesis
This is not AI replacing the Zettelkasten. It is AI making the Zettelkasten’s demanding discipline actually sustainable for mortal humans with finite time and attention.
The tools surveyed here — Graphify, obsidian-second-brain, Smart Connections, InsightA, Notemd, and the agent-skill frameworks — all converge on this same division of labor. They are not trying to think for you. They are trying to maintain the infrastructure of thinking so you can actually do it.
Where This Goes
The trajectory is clear. Obsidian has become the de facto standard for AI-adjacent personal knowledge management because it resolves a fundamental tension: you want AI to know your context, but you do not want your context locked in someone else’s database. Local markdown files, owned entirely by you, readable by any tool, versioned with git, are the right substrate.
The next step — already visible in tools like obsidian-second-brain and COG — is vaults that are not just readable by AI but actively maintained by it: agents that run on a schedule, fill gaps in your knowledge graph, surface contradictions in your decision history, and update your notes with new research while you sleep.
The vault is becoming a living system. The question is no longer whether to bring AI into your Obsidian workflow, but how to structure the collaboration.
More for AI powered folks :
Clarity Engineer : Code Is the Side Effect
Tools referenced: Graphify · obsidian-second-brain · Smart Connections · Copilot for Obsidian · Text Generator · InsightA · Notemd · claude-obsidian · COG-second-brain · obsidian-ai-second-brain
Comments
Loading comments…