Claude Just Started Remembering Your Client Conversations. Here's What to Do.

On August 25, 2026, Anthropic enabled Claude Memory across Chat and Cowork — unified, persistent, and on by default for every personal account. Starting today, Claude retains what you tell it across every future session, automatically, without you asking it to.

For most personal users, this is a useful feature. For attorneys and CPAs using personal Claude accounts for client work, it is a client confidentiality problem that professional services firms need to fix today.

What Changed

Before August 25, Claude's memory in Cowork and memory in Chat operated separately. Now they are unified into a single persistent profile that Claude maintains across every interaction.

The defaults are:

• Free, Pro, and Max plans (personal/individual accounts): Memory is ON by default.
• Team and Enterprise plans: Memory is OFF by default.

When memory is active, Claude is not just recalling what you said in the current conversation. It is building and updating a persistent profile — facts, names, context, preferences — that carries into every future session on that account.

That means the attorney who mentioned a client's name and the nature of their dispute last week? Claude may carry that into their next session. The CPA who walked Claude through a client's tax situation in June? That context may still be there.

The Client Confidentiality Exposure

Attorney-client privilege protects confidential communications between attorneys and clients. What it does not clearly cover is a third-party AI system that has been given access to those communications and is actively retaining them in a persistent memory profile.

IRS Circular 230 §10.20 requires tax practitioners to maintain client confidentiality. The AICPA Code §1.700 (Confidential Client Information Rule) prohibits disclosing or using confidential client information without client consent. Neither rule was written with persistent AI memory in mind — and neither provides a clear safe harbor for information that a practitioner voluntarily fed into an AI tool.

The practical exposure is not that Anthropic will read your client files. It is narrower and more immediate:

• Memory items can surface in future sessions in unexpected ways
• A practitioner using one account for multiple clients could see memory from Client A appear in a session focused on Client B
• Memory stored on a personal account is not covered by a business data processing agreement
• If a matter goes into discovery, the existence of client-specific AI memory items is a question that did not exist six months ago

None of this is theoretical. It is the natural consequence of using a personal productivity tool — one designed for individual use — for work that carries professional confidentiality obligations.

The 4-Step Fix

This takes about 10 minutes. Do it before your next Claude session.

Step 1: Check your plan. Log into your Claude account at claude.ai and go to Settings > Plan. If you are on Free, Pro, or Max, memory is on by default. If you are on Team or Enterprise, memory is off by default and you can stop here — but verify that none of your team members are using individual Pro or Max accounts for client work.

Step 2: Audit what Claude has stored. Go to Settings > Memory > Topics. You will see a list of memory items Claude has accumulated. Read through all of them.

Step 3: Delete any client-specific items. Delete any memory item that references:

• A client name or identifying detail
• A matter, case, or engagement description
• Financial figures, deal terms, or settlement amounts
• Legal strategy, tax positions, or advisory guidance you provided
• Any engagement-specific context whatsoever

If you are not certain whether an item could identify a client, delete it. The cost of over-deleting is minimal. The cost of leaving client-identifying memory in place is not.

Step 4: Decide on your plan structure. If anyone at your firm is using personal Claude accounts for client-facing work, you have a governance gap that predates today's memory update. The right structure for firm-level client work is a Claude Team or Enterprise account — which provides business terms of service, a data processing agreement, and memory off by default.

If upgrading the whole team is not immediate, the interim fix is to manually clear memory after any session involving client data and to set a firm policy that personal AI accounts are not used for client work.

What Is Safe to Leave in Memory

Not all memory needs to be deleted. Claude Memory is genuinely useful when it contains non-client-identifying professional context.

Safe to keep:

• Your preferred output format ("I prefer numbered lists over bullet points")
• General professional context ("I work in tax advisory for small businesses")
• Firm location and size ("12-person firm in Chicago")
• Communication preferences ("keep explanations concise")

Delete immediately:

• Any client name, company, or identifier
• Any matter or engagement description
• Any financial, legal, or advisory detail tied to a specific client
• Any document content you submitted from a client file

The test: if this memory item appeared in a session with a different client, would it be a problem? If yes, delete it.

The Broader Pattern

This is not the first time a useful AI feature has created professional services compliance exposure on arrival. Claude Fable 5's 30-day data retention window created a similar gap in June. The pattern repeats because AI tools are built for consumer and developer use cases first, and professional services obligations are a specialized compliance layer that the default settings do not account for.

The firms that stay ahead of this are the ones who treat AI tool updates the same way they treat regulatory changes: as something to review and respond to promptly, not something to circle back to when it becomes a problem.

Today that means auditing your Claude Memory settings. Next time it will be something else.

This week's action: Log into Settings > Memory > Topics. Spend 10 minutes reviewing and clearing any client-specific memory items. Done before your next session.