Anthropic announced on August 25, 2026 that Claude's memory now works the same way everywhere: in chat, and in Claude Cowork sessions running in the cloud on web and mobile. Instead of summarizing a conversation once it ends, Claude now builds and updates memory "topics" while the conversation happens, showing every saved topic in one place that users can read, edit, or delete.
For most people, this shows up as a better product experience. For admins, it's a governance moment: a new organization-level sensitive-topics switch, automatic carryover of existing memories (sensitive topics included), and a hard deadline for exporting legacy memory. None of that requires a complicated response — but it does require a checklist, and this week is the right time to run it.
Quick Answer
Claude's memory update unifies how memory works across chat and cloud-based Cowork sessions, replaces end-of-conversation summaries with continuously updated topics, and adds a dedicated sensitive-topics control that is off by default at both the organization and individual level. Admins should review Organization settings → Capabilities, decide their organization's posture on sensitive topics, tell users how to view and manage their own memory, and export any legacy memory that is needed before September 9, 2026. Everyone else mainly needs to know that memory is now visible and editable as a list of topics, and that nothing sensitive is stored without an explicit, individual opt-in.
TL;DR
- One memory system, everywhere. Chat and Claude Cowork now share the same memory, but only when Cowork runs in the cloud — memory does not apply to local, on-computer Cowork sessions.
- Topics update continuously. Claude adds and edits individual memory topics as a conversation happens instead of waiting to summarize afterward.
- Full visibility and control. Every saved topic appears in Memory settings, where users can view, edit, or delete it individually.
- Conservative defaults. Memory is available for organizations to turn on but starts off for end users; sensitive topics stay off and hidden until an admin enables the capability and a user separately opts in.
- Nothing sensitive is stored by default, and certain categories — government ID numbers, financial account numbers, criminal history, and immigration status — are designed never to be saved regardless of settings.
- Existing memories carry over automatically, including any sensitive topics already saved under the prior system.
- Legacy memory export closes September 9, 2026. After that date, the option to export memory in the old format goes away.
- Vantage Point covers the related Opus 5, Microsoft 365 connector, and Cowork remote-session changes in our Claude Team admin readiness guide — read both together if your organization hasn't reviewed Claude admin settings recently.
What Changed?
Anthropic's announcement bundles several related changes into a single memory system. Here is what is actually different.
Memory now works the same everywhere
Previously, memory behaved somewhat differently between Claude's chat interface and Claude Cowork. Now the two share one memory system, with one boundary: it applies to Cowork sessions running in the cloud, not sessions running locally on a user's own computer. Confirm which mode your team relies on, since only cloud-based sessions build and use memory.
Topics are built and updated during the conversation
Rather than generating a single summary after a conversation ends, Claude now creates and updates discrete memory "topics" while the conversation is active — drawing on relevant context sooner and keeping it current, rather than reconstructing it from a static end-of-session snapshot.
Users see and control everything as a list of topics
Memory settings now present saved information as an itemized list of topics rather than an opaque blob of context. Users can open the list, read exactly what has been saved, edit an entry, or delete it — individually, without an all-or-nothing reset.
Sensitive topics get their own switch
Health information, religious beliefs, political views, and similar categories are handled separately from general memory. This sensitive-topics setting is off by default. In an organization, it stays hidden from users entirely until an admin turns on the capability; once enabled, each user still has to opt in individually before Claude will save anything sensitive for them. A user who has opted in can also remove sensitive topics from their memory at any time.
Existing memory carries over automatically
Anthropic says current user memories move to the new system automatically, including sensitive topics already saved. Nothing is deleted in the transition, but organizations that migrate should assume previously saved sensitive information now sits inside the new system's topic list.
Safety limits are built in
Claude is designed not to save certain categories: government-issued ID numbers such as Social Security numbers, financial account numbers, criminal history, immigration status, or anything else violating Anthropic's Acceptable Use Policy. When Claude can't save something for one of these reasons, it tells the user rather than saving it silently.
Who Is Affected?
The update reaches every Claude user, but the defaults and the controls that matter depend on plan type.
| Plan type | Memory default | Who can change it |
|---|---|---|
| Free, Pro, and Max (individual) | Available now; on by default on current app versions | Each user, directly in their own Settings → Memory |
| Team and Enterprise (organization) | Available for the organization; off until an owner turns it on | An organization Owner or Primary Owner, in Organization settings → Capabilities |
| Any plan — sensitive topics | Off by default everywhere | Org owner must enable the capability first (Team/Enterprise); then each individual user opts in separately |
Two groups should pay particular attention this week: organization owners deciding whether memory and sensitive topics suit their workforce, and anyone on any plan who relied on the legacy memory export option and needs to act before it disappears on September 9, 2026.
What Should You Do?
Admin checklist
- Review Organization settings → Capabilities. Confirm whether memory is currently on or off for your organization, and whether that matches your intended policy. This is the same admin area covered in our Claude Enterprise setup guide to SSO and provisioning if you haven't audited those controls recently.
- Decide your sensitive-topics posture before enabling it. Turning on the capability only exposes the choice to users — it does not turn memory of sensitive topics on for anyone until each person opts in individually. Decide as a policy matter whether that individual choice should be available at all in your organization.
- Tell users what changed. Let your team know memory now appears as an editable list of topics they can delete anytime, and — if you enable the capability — that sensitive topics require a separate, individual opt-in.
- Export legacy memory before September 9, 2026, if you need it. Anyone who wants a record of memory in the prior format should use Settings → Memory → Export legacy memory before the deadline. This is worth flagging to your broader team, not just handling for yourself.
- Check your data retention and industry agreements. Organizations operating under HIPAA, public-sector, or custom data retention terms should confirm with Anthropic how the updated memory system applies to their specific agreement before enabling it broadly.
What individual users should know
Most users don't need to do anything, but three things are worth knowing. First, memory now lives in Settings → Memory as a list of individual topics you can inspect anytime. Second, if your organization enables sensitive topics, opting in is your choice alone — your admin cannot see or edit what you've saved, and turning it back off later removes previously saved sensitive items. Third, export your memory in the legacy format before September 9, 2026 if you want a personal record; after that, the option is gone.
Why This Matters for Regulated Industries
For organizations in financial services, healthcare-adjacent, and other regulated sectors, the sensitive-topics switch is a data-governance decision wearing a product-settings costume. The categories Anthropic groups under "sensitive" — health, religious belief, political views, and similar — overlap with information your compliance function likely already treats as a distinct, higher-scrutiny data class. Two defaults work in your favor: the capability stays hidden until an admin deliberately enables it, and even then each person opts in individually rather than inheriting an organization-wide default. That gives compliance and IT time to set a documented policy — allow, restrict, or defer — before any employee sees the toggle. It's also worth confirming directly with Anthropic whether your organization's data retention agreement changes how memory behaves, particularly under HIPAA, a public-sector agreement, or a custom retention arrangement, since those commitments should govern the decision. If your organization hasn't yet built a complete data governance foundation for Claude beyond this one setting, our guide to building a secure Claude environment covers identity, access, and data governance as a full system.
Timeline
- August 25, 2026: Anthropic publishes the memory update, announcing unified chat/Cowork memory, continuously updated topics, and the sensitive-topics control.
- Week of August 24–28, 2026: The update rolls out; timing can vary by workspace, so confirm your own organization's settings rather than assume every account changed the same day.
- September 9, 2026: Deadline to export legacy memory in the prior format from Settings → Memory → Export legacy memory. After this date, that option is no longer available.
How Vantage Point Helps
Vantage Point is a Claude Partner Network Member that helps financial services and other regulated organizations turn Claude product changes into governed policy rather than ad hoc settings decisions. If your team needs help setting an organization-wide posture on memory and sensitive topics, documenting it for compliance, and communicating it clearly to users, our compliance and security solutions practice can assess the decision against your existing data governance framework, and our Claude services team can help configure and roll out the change itself. For teams that also need to train users, our advisory and change management practice can help build that internal communication. Contact Vantage Point to talk through your organization's posture on Claude's memory and sensitive-topics settings.
FAQ
Does this change apply to Claude Cowork?
Yes, with one boundary: memory applies to Cowork sessions running in the cloud on web and mobile. Sessions running locally on a user's own computer do not use this memory system, so a team that relies heavily on local Cowork sessions won't see the same continuity there.
Is memory on by default?
It depends on the plan. Individual Free, Pro, and Max users have memory available and on by default on current app versions. Team and Enterprise organizations start with memory available but off — an organization Owner or Primary Owner has to turn it on in Organization settings → Capabilities before anyone in that organization gets it.
What exactly counts as a "sensitive topic"?
Anthropic's own examples include health information, religious beliefs, and political views, describing the category broadly rather than exhaustively. If you're unsure whether something qualifies, treat it as sensitive for policy purposes until confirmed otherwise with Anthropic.
Who can turn on sensitive topics, and who decides for individual users?
Enabling the capability is an admin-level decision in Organization settings → Capabilities, and it does nothing on its own except make the option visible. After that, each individual user chooses whether to opt in personally. Admins cannot opt in on a user's behalf, and — per Anthropic's design — cannot view or edit what an individual user has saved.
What happens to sensitive topics a user already had saved if they opt out later?
If a user who previously opted in turns the setting back off, Claude removes the sensitive topics that were saved for that person. General, non-sensitive memory is not affected by that change.
Will Claude ever save Social Security numbers or similar IDs in memory?
No. Anthropic states that Claude is designed not to save sensitive identification numbers, financial account numbers, criminal history, or immigration status, regardless of memory settings. When Claude declines to save something for this reason, it tells the user rather than saving it silently.
What happens to memory a user already had saved before this update?
Anthropic says existing memories carry over automatically to the new system, including any sensitive topics that were already saved under the prior version. Review what carried over using the new topic list in Settings → Memory, since it may include more than you expect if sensitive topics were previously in use.
What happens if we miss the September 9, 2026 export deadline?
The specific option to export memory in the legacy format goes away after that date. Anthropic's standard data export options remain available, but a record in the prior legacy format must be exported before the deadline.
Sources
- Anthropic: Claude's memory works everywhere, and you decide what's in it
- Anthropic Help Center: Use Claude's chat search and memory to build on previous context
- Anthropic Help Center: Import and export your memory from Claude
- Anthropic Help Center: Claude release notes
This guide reflects Anthropic's public announcement and help center documentation as of August 26, 2026. Rollout timing and controls can vary by plan and workspace, so confirm current behavior in your own settings before finalizing policy.
