Mythos

If you've spent any time in AI coding communities over the past year, you've seen the pattern: someone drops a CLAUDE.md in their 📝Obsidian vault, points an 📝Model Context Protocol (MCP) server at the folder, and suddenly their AI assistant remembers them. It's a real unlock. And for a lot of people, it's the first time they've experienced persistent AI memory that they actually own.

Obsidian didn't set out to be an AI memory system. It's a note-taking app — a very good one — built around local Markdown files, bidirectional links, and an ecosystem of 1,800+ community plugins. But when the 📝Model Context Protocol (MCP) made it possible to connect any tool to 📝Claude Code, 📝Cursor, and other AI agents, Obsidian vaults were the obvious first thing to plug in. People already had their thinking there. Why not let the AI read it?

That instinct was right. The pattern took off. And then it started showing its edges.

How the pattern works

There are really three flavors of "Obsidian as AI memory," and most people use some combination:

The 📝CLAUDE.md file. The simplest version. A single Markdown file in your vault root that Claude Code reads automatically at session start. You put your preferences, current projects, working style — whatever you want Claude to carry as baseline context. Zero setup, zero dependencies. The limitation is that it's manually maintained, so it goes stale the moment your life changes and you forget to update it.

MCP server retrieval. A step up: an MCP server exposes your vault as a set of tools — read_note, search_vault, create_note — so the AI can actively pull notes mid-session instead of just loading one file. Several open-source servers exist (obsidian-mcp, mcp-obsidian, obsidian-mcp-tools, obsidian-mcp-server), each with different tradeoffs. The detailed setup is covered in 📝Obsidian MCP + Claude Code.

Daily notes as session logs. Claude reads today's daily note for context, works with you, writes a summary at the end. Your daily notes become a rolling chronicle of AI-assisted work. It's a workflow convention more than a feature — but it creates real continuity when you're in the groove.

For solo developers and researchers with small-to-medium vaults, this setup genuinely works. The zero-friction entry point — cd ~/vault && claude — is hard to beat.

Where it hits a wall

The more you lean on it, the more you hit the same set of structural limits. Not bugs. Not missing features. Structural limits that come from asking a note-taking app to be something it wasn't designed to be.

Context window cost. Most Obsidian MCP servers retrieve notes by loading entire files into the context window. With a small vault, this is fine. With hundreds of notes, you're burning tokens on content that isn't relevant to the current task. There's no semantic filtering by default — the server can't tell the difference between a note that's exactly what you need and one that just happens to share a keyword. You can fix this yourself: Smart Connections embeds the vault on-device, and a companion MCP server hands those results to Claude. It works well. It's also three moving parts you assembled and now maintain, rather than how retrieval arrives out of the box.

No access control. The MCP server exposes your entire vault. Personal journals, client notes, health logs, half-baked ideas you'd never share — all equally visible to any query. There's no way to say "you can see my project notes but not my therapy reflections." The community workaround is "keep backups," which tells you everything.

A machine has to be running. The MCP integration needs filesystem access, which used to mean strictly desktop. Obsidian's official headless Sync client changed that in February 2026 — you can keep a vault alive on a server and reach it from anywhere. Either way, something of yours has to be running, patched, and reachable for your AI to have memory. Forget to start it, or take a laptop off the desk without a server behind it, and the context vanishes.

Write risk. When Claude has write access to your vault, it can overwrite your notes, reformat your content, or create files that don't fit your organizational structure. Obsidian has no concept of "AI-authored content" vs. "your content." Some people go read-only to avoid this, but that kills session logging and note creation.

Scaling friction. Past a few hundred notes, search gets noisy, tool calls slow down, and the ratio of useful retrievals to noise drops. The MCP server is a file-system adapter. It doesn't understand relationships between notes, can't prioritize what's relevant, and doesn't learn from your patterns.

What I built instead

These are exactly the problems I ran into before building @MythOS. Not hypothetical problems — the actual walls I hit while trying to make Obsidian work as my AI memory layer.

MythOS uses vector embeddings for semantic retrieval instead of file-loading. It has three-tier visibility with audience tags so I can control what any given agent or collaborator can see. It connects via HTTP transport and OAuth 2.0, so it works on Claude's web and mobile interfaces — not just desktop. And it separates AI contributions from my content through a dedicated notes field, so there's no overwrite risk.

The honest framing: Obsidian is a great note-taking app that clever people have adapted for AI memory. MythOS is what I built when I needed the AI memory part to actually be the product.

FAQ

What's the simplest way to try Obsidian as AI memory?

Drop a CLAUDE.md file in your Obsidian vault root and open Claude Code from that directory. Claude reads it automatically. No plugins, no MCP server, no configuration. Start there and add complexity only when you need it.

Can I use Obsidian and MythOS at the same time?

Yes. Claude Code supports multiple MCP servers simultaneously. A lot of people keep Obsidian for local note-taking and use MythOS for AI memory and retrieval. It's a common transition path.

Does Obsidian AI memory work on mobile?

Not on its own. The integration needs filesystem access, so a mobile session reaches your vault only if you keep a desktop or a headless-sync server running for it. Obsidian's mobile app edits notes fine; making those notes reachable by an AI agent from a phone is a server you now operate.

How big can my vault get before retrieval breaks down?

There's no hard limit, but most people start noticing noise and token waste past a few hundred notes. The issue isn't vault size — it's that keyword-based file retrieval doesn't scale the way semantic search does.

How does this compare to MythOS?

A detailed comparison is in 📝MemPalace vs Obsidian vs MythOS. The short version: Obsidian adapts a note-taking app for AI memory. MythOS was built from the ground up for AI collaboration — semantic retrieval, access control, cross-platform MCP, structured content separation.

Related

Contexts

Created with 💜 by One Inc | Copyright 2026