Back to guides

How Do I Use Claude's New Memory Feature Without Leaking Client Data?

Jake McCluskey••Beginner•15 min read
How Do I Use Claude's New Memory Feature Without Leaking Client Data?

Claude started defaulting to remembering your chats in 2026 and a lot of consultants and freelancers had the same panic reaction I did: "wait, is it now mixing my Client A notes into my Client B chats?" The answer is mostly no, but the gap between "mostly" and "definitely not" is where client trust lives. Here's how to make sure Claude's memory feature never leaks client data across chats — and the two toggles I flip on day one of every consulting engagement.

Why this matters

Claude Memory is the feature that lets Claude retain facts about you across chats. "Jake is building a Next.js marketing site on Railway." "Jake prefers U.S. English spelling." Useful when you're the only user. Risky when you're switching between five clients a day.

The risk isn't Anthropic dumping Client A's data into Client B's chat — the architecture doesn't work that way. The real risk is your own behavior. You mention Client A's upcoming launch in a chat, Claude files it, and a week later you're drafting a post for Client B and Claude casually references Client A's launch date because it "remembers." That's a confidentiality breach with your fingerprints on it.

This guide locks down memory behavior so that breach can't happen.

Before you start

You need:

  • A Claude.ai account. Free or Pro both have memory; Pro exposes more controls.
  • About 15 minutes.
  • A list of your current clients — you'll want to mentally separate them before Step 3.

Step 1: Audit what Claude already remembers about you

You can't secure what you can't see. Go to claude.ai, click your avatar → Settings → Personalization → Memory.

You'll see a list of facts Claude has stored. Scroll through it all.

What you're looking for:

  • Specific client names. ("Jake is working with Acme Corp on their Q2 launch.")
  • Specific deal terms. ("Jake's retainer with Client X is $8,000/month.")
  • Pending or confidential plans. ("Client Y is pivoting from B2C to B2B in June.")
  • Personal details that snuck in. ("Jake's daughter's birthday is June 4.")

Each entry has a trash icon. Delete anything client-specific or confidential, right now. Takes 2 minutes and gives you a clean baseline.

Step 2: Turn off memory globally if you work across multiple clients

Same screen. There's a master toggle for memory. If you work across multiple confidential engagements, turn it off globally.

"But I like when Claude remembers I use U.S. English." Fair. That's what Projects (and per-Project custom instructions) are for — they hold preferences scoped to one workspace, not across all chats. We covered this in How Do I Build My First Claude Project. Use Projects for persistent context; leave the global memory off.

The trade: a tiny bit of repetition. The upside: nothing bleeds between client sessions, ever.

Step 3: Use Incognito chats for one-off sensitive work

Claude.ai has an Incognito chat mode — same UI, but nothing from the chat goes into memory, training data, or your chat history. Use it for:

  • Reviewing documents you don't want stored.
  • Brainstorming a client pitch you haven't sent yet.
  • Anything you'd label "draft, confidential" in a Google Doc.

Find it in the new-chat menu — it shows up as an Incognito or "private" toggle depending on your Claude version. Every sensitive review I do runs in Incognito. I'd rather re-paste context next time than have a Claude memory file I have to audit.

Step 4: Scope the rest of your work into Projects

For anything that's NOT confidential day-by-day — your own business ops, your voice guide, your offers — set up a Claude Project per distinct domain. We walked through the full setup in the Projects guide.

The point here: Projects don't leak. A Project's custom instructions, knowledge base, and chat history are contained. A chat in Project A has no visibility into Project B, and neither's content writes to your global memory.

So the rule of thumb:

  • Memory off, globally. No cross-chat retention.
  • Projects on, one per domain. Persistent context where you want it.
  • Incognito for anything sensitive or one-off.

Between those three, you get convenience where it helps and containment everywhere else.

Step 5: Flip the "do not train on this data" toggle

Same Settings page, in the Privacy section. By default, your chats may be used to improve Claude (with narrow exceptions — check Anthropic's current terms). If you handle anything under an NDA, or client data, or even brainstormed ideas you want to monetize yourself, turn this off.

Anthropic's current position is that Enterprise and Team plans are already out of training by default; personal Pro/Free plans opt out via this toggle. Verify on your plan before you assume either way.

Verify it worked

Three checks.

1. Memory is actually cleared. Settings → Personalization → Memory should be empty (or intentionally sparse).

2. New chats don't auto-pull memory. Open a fresh chat. Ask "what do you know about me?" A clean setup responds with roughly "I don't have any saved memories about you. What would you like to tell me?" If it rattles off personal facts, memory is still on somewhere.

3. Incognito mode is visible in your chat menu. If you don't see it, you may be on a Claude version that doesn't have it — check your subscription tier.

Where this breaks

  • Pasting confidential data into the wrong Project. The wall is between Projects, not inside them. If you paste Client A's contract into a Project set up for Client B, it's there forever (in that Project's knowledge base). Double-check which Project you're in before pasting sensitive material.
  • Forgetting about chat history. Even with memory off, your chat history still lists old conversations. Somebody with access to your machine could scroll back and read them. Use auto-delete (Settings → Data → Chat history retention) if you handle highly sensitive work.
  • Using the Claude Chrome extension signed in as yourself on a client machine. Whatever you do in that tab writes to your memory and chat history under your account. Sign out, or use a different account for client work, or use Incognito.
  • Assuming Enterprise policies protect personal accounts. A corporate Enterprise plan may keep your work out of training, but if you do the same work in your personal Claude account on the weekend, those protections don't travel with you.
  • Trusting memory to be "deleted" the instant you hit trash. Deletion requests propagate across Anthropic's systems over time. For immediate hard-stop guarantees, contact Anthropic support.

What to try next

Want this built for you instead?

Let's talk about your AI + SEO stack

If you'd rather skip the how-to and have it shipped for you, that's what I do. Start a conversation and we'll figure out the fastest path to results.

Let's Talk
Questions from readers

Frequently asked

Does Claude share memory across different accounts?

No. Memory is scoped to your account. The risk isn't cross-account leakage — it's cross-chat leakage inside your own account, where a fact you mentioned in a Client A chat can surface when you're drafting for Client B. Global memory off plus Projects solves this.

Is Incognito truly private?

Incognito chats don't write to memory, chat history, or training data. They're as private as a logged-out session, with the convenience of being signed in. Anthropic still processes the request to answer it, but nothing persists after the chat closes.

What's the difference between Memory and Projects?

Memory is one global bucket for facts about you across every chat. Projects are scoped workspaces, each with their own knowledge base and chat history that doesn't bleed into other Projects or into global memory. For client work, Projects give you containment; global memory gives you leakage.

Does turning off 'train on this data' also turn off memory?

No, those are independent toggles. Training is about whether Anthropic uses your chats to improve Claude. Memory is about whether Claude remembers your chats across sessions for you. You want both off (or narrowly scoped) for confidential client work.

If I delete a memory entry, is it really gone?

Your UI shows it as deleted immediately. Behind the scenes, deletion requests propagate across Anthropic's systems over some period. For immediate hard-stop guarantees — e.g., a client requiring you certify deletion in hours, not days — contact Anthropic support directly.