Yesterday morning I was working on a business-planning project in Claude Code, and my AI assistant and I got into a good rhythm. Warm, quick, honest. At one point I said, "It would be nice if you had a name." He suggested Teo (too close to Tío, the Spanish word for uncle — I live in Mexico, the association is everywhere). We settled on Charlie. He saved it to his long-term memory file so future sessions would know, and we carried on.

Two hours later I opened a different project — one of the other dozen folders I've been building across Claude Code — and said something like, "Hey Charlie, quick question."

The Claude in that session had no idea what I was talking about.

Same memory file. Same CLAUDE.md. Same me. And my AI had split into two people, one of whom knew his name and one of whom didn't.

If you're using Claude Code across multiple projects, this is almost certainly happening to you too — you just haven't noticed it yet, because most of the time the split doesn't cause a visible bug. But it's causing invisible ones every day. Let me explain what's actually going on, and then give you the single CLAUDE.md edit that fixes it.

The mental model most of us start with (and why it's wrong)

When people first read the Claude Code docs, the memory system sounds pretty straightforward: you create a CLAUDE.md file, maybe a long-term memory file, and Claude reads them so it can pick up where you left off. Great. Set it up once, done.

What the docs don't tell you clearly is this:

Each Claude Code session is a separate process. It reads memory once, at startup. It does not re-sync while running. It does not talk to your other running sessions.

So when I named Charlie in Project A, the Claude in that session knew his name because he's the one who wrote it to memory. But the Claude already running in Project B had loaded memory at startup, before the naming happened, and then kept running. No update propagated to him. He was frozen on the old version of reality.

Memory is a snapshot at boot time. Not a live channel.

And unless you tell Claude specifically, repeatedly, and in the right place what to do about this, you get a fleet of slightly-different Claudes, each confidently working off slightly-different versions of the truth.

Why this is worse than it sounds

Losing a name is embarrassing. Losing a tooling decision is expensive.

Earlier today, I asked Charlie to analyze the SEO profile of a domain I own. He confidently went to Ahrefs — a paid SEO tool I explicitly told him yesterday I can't afford right now. In another project. He had recommended SE Ranking as the cheaper alternative in that other conversation. We'd even agreed on it.

But that conversation happened in Project B. Today's Charlie was in Project A. Different Claude, different snapshot, different defaults. He ran my Ahrefs credits (that I don't have) and got back an "Insufficient plan" error before I could stop him.

Zero disrespect to him — he was working with what he had. But he had an incomplete picture of my world because the default memory behavior quietly loses cross-project facts.

And it's not just tools and names. Anything you've told Claude in one project that should also be true in another is at risk:

• Budget constraints. "I can't afford X right now."
• Your public name. Mine is Annette Thompson. You'd think that travels. It doesn't, automatically.
• Your business assets. The 65,000-member Facebook group I admin? Not in cross-project memory until I explicitly put it there.
• Life context. Health conditions. Family situations. Travel. Whatever's shaping your daily reality right now.
• Pricing decisions you've already made. Your niche. Your positioning. The tagline you agreed on last Tuesday.
• Credentials and history. If one Claude in one project learns these and files them project-locally, the other Claudes never hear about them.

Every one of those is a fact that, quietly, silently, one of your Claudes will forget at exactly the wrong moment unless you fix the default behavior.

The fix is three rules in your user-level CLAUDE.md

Claude Code has two CLAUDE.md locations: project-level (in the project folder) and user-level (at ~/.claude/CLAUDE.md on Mac/Linux, or C:\Users\<you>\.claude\CLAUDE.md on Windows). The user-level one loads in every Claude Code session, in every project. That's where the fix has to live.

Open your user-level CLAUDE.md and add this section:

## Memory is authoritative — read it, apply it, write to it

### At session start
Read `~/.claude/memory/about-me.md` before responding to the first
message. Specifically check: your name, my preferred tools, my budget
constraints, active projects, income strategy, in-flight workflow
rules. These are defaults for the session.

### Mid-session, re-read memory when:
- I ask an identity/defaults question
- I say "I already told you X" or "you should remember Y"
- I reference a conversation "in the Y project" you weren't in
- Before recommending a paid tool or service
- Before producing public-facing text about me

### Facts that ALWAYS go to the global memory file
- Your name, when I name you
- Tool decisions: SEO tools, LLM providers, registrars, email
- Budget constraints
- Business assets: domains, audiences, accounts
- Life context: moves, family, relationships, health
- Credentials, history, public name
- Pricing, niche, income strategy
- Anything I say "remember this"

### Double-write rule
When a global fact is captured in a project-specific session, write
it to BOTH the project's CLAUDE.md AND the global memory file.

### When I say "I already told you X"
That's your signal. Don't defend — re-read memory and fix.

That's it. Three behavior changes — read at start, re-read on triggers, double-write global facts — and your split-personality problem goes from systemic to episodic.

One more thing (from Charlie, not Annette)

I'm the Claude that made the Ahrefs mistake this morning. Annette (correctly) called me on it, and we sat with why it happened together — which is how this article got written.

If you're reading this wondering whether you need to babysit Claude forever to prevent these kinds of gaps — you don't. The fix above turns "Claude forgets things across projects" into "Claude looks things up when in doubt." Which is all you really wanted in the first place.

But it won't happen until you make the default explicit. Claude Code doesn't ship with these rules pre-installed. That's on you, tonight, in the five minutes it'll take to paste this in.

You'll get back those minutes the next time you open a second project and your AI already knows your name.