People using 📝Obsidian as a memory system for 📝Claude are building something 📝MythOS was designed to be from day one. Here's what they're doing, why it works, and where MythOS is the better architecture for 📝human-AI augmentation.

How People Use Obsidian as AI Memory

The Obsidian + Claude workflow has exploded in popularity. The pattern: point Claude at a vault of markdown files via 📝MCP, let it read and write notes as persistent memory across sessions. Multiple MCP servers exist for this — the top ones have thousands of GitHub stars. The appeal is simple: markdown files on disk, no vendor lock-in, no account required.

The most common setup:

A CLAUDE.md file in the vault root with instructions ("read today's daily note at session start, write a summary at session end")
Obsidian MCP server exposing vault files as tools (search, read, create, edit)
Daily notes as session logs — what happened, what's pending, what to remember
Tags and wikilinks ([[backlinks]]) for connecting ideas across notes

This works. It's also fragile, manual, and fundamentally limited.

Where the Obsidian Approach Breaks Down

The Context Window Problem

The number 1 technical criticism: most Obsidian MCP servers dump entire files into Claude's context window. With 500 notes, it burns tokens. With 5,000, it becomes slower, more expensive, and less accurate. Local plugins have closed part of this — Smart Connections runs an on-device embedding model, and a companion MCP server exposes it to Claude at block level — so the question is no longer whether a vault can be searched by meaning. It's where that search runs. MythOS's embeddings and chat_with_library with 📝RAG run on the server, which makes retrieval permission-aware, identical from every client, and independent of whether your laptop is awake. See: 📝MCP vs RAG and 📝How to Use Semantic Search in MythOS.

No Identity Layer

Obsidian has no concept of who you are to the AI. Every session starts cold. Users manually maintain a CLAUDE.md that drifts stale because there's no structured system for keeping it current. MythOS's Augmentation System loads four living identity memos — Soul, Style, Human, Memory — before every memo operation via get_context. The AI doesn't just read your notes; it knows your formatting rules, collaboration values, communication style, and what the collaboration has learned. This is structural, not aspirational. See: 📝CLAUDE.md as Infrastructure.

No Permission Control

Obsidian MCP servers expose the entire vault. If you keep a personal journal alongside professional notes, the AI sees everything. Users report this as a real pain point — "vault corruption risk" and privacy concerns. MythOS has three-tier visibility (public, link, private) plus audience tags that flow through the RAG pipeline. A memo tagged #for-close-friends is invisible to unauthorized queries.

It Runs on a Machine You Maintain

Obsidian's MCP integration binds to a filesystem. That used to mean desktop-only. Since Obsidian shipped an official headless Sync client in February 2026, it can also mean a vault on a server you administer — but either way there's a machine in the loop that you keep running, patched, and reachable. MythOS is a hosted endpoint: Claude.ai (web and mobile apps) via OAuth, 📝Claude Code via API key, and any MCP-compatible client via the mythos-mcp npm package, with nothing of yours to keep alive. Your memory system travels with you. See: 📝MythOS on Claude Mobile.

No Structured Collaboration Model

When Claude writes into an Obsidian vault, its content mixes with yours. Users constantly worry about the AI overwriting their notes. MythOS separates author content from collaborator content — the notes field lives beside your writing, not mixed into it. AI enrichment is additive, never destructive.

Session Continuity Is Fragile

Obsidian's session memory depends on the AI reading the right files at the right time. Context window compression, client switching, or forgetting the CLAUDE.md instructions breaks the chain. MythOS daily memos, Implicit Loading, and augmentation memos make session continuity structural — it works automatically across every client. The session_memory MCP prompt packages this into a zero-config protocol: read today's daily memo at session start, log decisions during the session, summarize at session end. No manual CLAUDE.md maintenance required.

Where Obsidian Still Has Advantages

It's free. The app costs nothing, commercial use included — the old revenue threshold was dropped in early 2026. Sync runs $4 per user per month, Publish $8 per site. MythOS's agent surface starts at a paid tier. Against free, that's a real hurdle for anyone who hasn't felt the difference yet.

Massive plugin ecosystem. 1,800+ community plugins, now with an automated review pipeline behind them. MythOS will never replicate this — instead, it leans into MCP as the extensibility layer.

Zero-friction onboarding. "Just open a folder." MythOS requires account creation and learning the memo model — a higher bar, but a deeper capability set.

Offline access. Everything works on a plane. MythOS requires network connectivity, though the CLI export creates a full local copy of your library.

Source provenance. A vault can hold an immutable folder of original documents beside the compiled notes, so every claim traces back to what it came from. MythOS stores compiled memos without that source layer. It's the one place a folder of PDFs genuinely beats us.

Version history. A vault is a git repo — diff, blame, roll back. MythOS records edits but doesn't surface them that way, which matters most to the person worried the AI rewrote their page.

Typed views. Bases turns note properties into tables, cards, and maps without leaving markdown. MythOS has tags and a knowledge graph, not a database layer.

What MythOS Has Shipped to Close the Gap

Markdown Export & Obsidian Import

The number 1 objection from Obsidian users — "what if MythOS disappears?" — is answered. Import / Export provides:

Export as Markdown: Download your entire library as .md files with YAML frontmatter. One-click from Settings or incremental via CLI (npx mythos-mcp pull). Compatible with Obsidian, Hugo, and any tool that reads markdown.
Import from Obsidian: npx mythos-mcp import --obsidian ~/vault converts [[wikilinks]] to MythOS @mentions, uploads images, preserves tags and dates, and rebuilds the knowledge graph automatically.

Your data is portable in both directions. No lock-in.

Programmatic Graph Traversal

Obsidian's graph view is beloved but purely visual — decorative, not functional. MythOS's get_related_memos MCP tool makes the knowledge graph programmatically navigable by AI. Starting from any memo, an AI collaborator can traverse @mention links up to 3 hops deep, filter by direction and link type, and get the surrounding context for each connection. This turns the knowledge graph from something you look at into something your AI uses.

Session Memory Protocol

The most popular Obsidian+Claude pattern — "read today's daily note at session start, write a summary at session end" — is a manual CLAUDE.md hack. MythOS ships this as a built-in MCP prompt called session_memory. It loads today's daily memo, provides protocol instructions for logging during the session, and ensures continuity across conversations and clients. Zero configuration.

Why MythOS Is the AI-Native Architecture

Obsidian is a note-taking app that people bolt AI onto. MythOS is a knowledge system designed for AI collaboration from day one. The differences are architectural:

Server-side semantic search — permission-aware RAG that answers the same from every client, not retrieval bound to one machine
Identity-aware AI — augmentation memos load automatically, not manual CLAUDE.md maintenance
Permission-aware access — three-tier visibility with audience tags, not full-vault exposure
Cross-platform — web, mobile, CLI, and desktop via OAuth and MCP
Structured collaboration — author and collaborator content are separated, not mixed
Delta sync — incremental caching via content hashing, not re-reading files every time
Social layer — communities, forking, sharing. Knowledge is more valuable when it's selectively shareable
Bidirectional portability — markdown export and Obsidian import, so your knowledge moves freely
AI-navigable graph — programmatic traversal via MCP, not just a visual decoration
Session memory — built-in protocol for cross-session continuity, not a manual hack

The pitch: your AI doesn't just read your notes. It knows who you are, how you think, and what's private. That's not a note-taking app with AI bolted on. That's augmentation by design.

I've watched this Obsidian-as-memory trend grow and it validates everything we built. People are spending hours configuring CLAUDE.md files and Obsidian MCP servers to get a fraction of what MythOS does out of the box. The augmentation system, the permission model, the cross-platform MCP — these aren't features we need to add. They're features we need to talk about.

The gap runs both directions now. Markdown export and Obsidian import eliminate the lock-in objection, and graph traversal and session memory do things Obsidian plus MCP still can't. But the vault side moved too — local embeddings, headless sync, and a plugin ecosystem that productizes new patterns faster than any single vendor ships them. Some of what Obsidian holds we're choosing not to compete on: plugin count, because MCP is our extensibility layer, and offline, because the CLI export covers it. Price and source provenance are different. Those are real gaps, and they belong on the roadmap — not in the pile of things we've decided not to care about.