Give Claude Code a memory that actually sticks.

Claude Code starts blind.

Claude Code starts every session from zero. No memory of yesterday's bug, no memory of last week's architecture call, no idea what the codebase looks like. You explain everything, every time.

The fix is two files and one plugin. A CLAUDE.md file Claude reads at startup, plus an Obsidian vault Claude can search across every session. Set it up once and Claude knows your project the next time you open the terminal.

Install and connect.

Open Obsidian → Settings → Community Plugins → search "Claude Code MCP" → install and enable the one by iansinnott. The plugin starts a WebSocket server on port 22360 the moment you enable it.

Now connect Claude Code to your vault. Open your terminal and run:

claude

Then inside Claude Code, run the IDE selector:

/ide

Pick Obsidian from the list. Claude finds the running server via a lock file in your Claude config directory. No manual JSON config needed. That's it. Claude can now read, write, and search your vault.

The file Claude reads at every session start.

CLAUDE.md is the file Claude Code reads at the start of every session. Drop one in any project root and it loads automatically. Treat it like a briefing for someone joining your team mid-sprint.

A starter you can paste and edit:

# Project Context
Stack: Next.js 14, TypeScript, Supabase, Vercel
Repo: ~/projects/my-app

# Standing Rules
- Always TypeScript, never plain JS
- Prefer server components over client components
- Check 02-Decisions/ before changing architecture
- Log every session summary to 03-Daily/

# Current Sprint (week of [date])
- [ ] Auth flow refactor (see 01-Projects/auth.md)
- [ ] Fix mobile nav regression
- [ ] Write API docs for /payments endpoint

Save this as CLAUDE.md at your project root. Claude reads it automatically every session.

Two things to know. First, Claude loads CLAUDE.md as context, not enforced rules. The more concrete the instruction, the more consistently Claude follows it. Second, you can drop one at ~/.claude/CLAUDE.md for instructions that apply across every project on your machine.

Treat the vault like departments.

How you organize your vault decides how useful it is. Give Claude a predictable place to look for context, decisions, and history.

00-Inbox/      Raw notes Claude dumps here. You triage later.
01-Projects/   Per-project folders. Each has its own CLAUDE.md.
02-Decisions/  Architecture decisions. One per file. Date, options, outcome.
03-Daily/      Auto-generated session logs. Your audit trail.
04-Resources/  Code patterns, API summaries, reusable snippets.
CLAUDE.md      The standing orders Claude reads every session.

The folder names don't matter. The discipline does. Once you commit to a structure, Claude can find anything you've written across every session.

How memory compounds.

The compound effect is the actual prize. It works like this:

Common traps.

1. Don't paste 2,000 lines of context into the chat to "remind" Claude. That's a workaround, not memory. Use CLAUDE.md.
2. Don't stuff every fact about your project into CLAUDE.md. Keep it to facts that matter every session. Move detail into 01-Projects/ or 04-Resources/ so Claude can fetch it on demand.
3. Don't skip git on your vault. If Claude corrupts a note, git checkout brings it back instantly.
4. Don't expect auto-discovery to fix a config problem. If /ide doesn't show Obsidian, the plugin isn't running. Restart Obsidian first.
5. Don't write CLAUDE.md once and forget it. Update it weekly with what you learned. The vault gets smarter only if you treat it that way.

You are the editor.

Treat the vault like documentation that future you will read. Claude writes the first draft. You decide what stays.

Try one today.

Four ways to put this to work before you close this tab:

1. Set up CLAUDE.md for your current project in 5 minutes.
2. Symlink your code repo's docs folder into 01-Projects/ . Claude sees both at once.
3. Add a single line to CLAUDE.md: "At session end, write a summary to 03-Daily/YYYY-MM-DD.md." Every session leaves a trail.
4. Try /memory inside Claude Code to view what it's already saved on its own.