Claude's memory update: what changed, and the September 9 deadline

Per Anthropic's own announcement, memory used to live separately in chat and in Cowork, Anthropic's file-and-document agent. Now it's one system. Coverage from TechCrunch and Engadget confirms the same details. For anyone using both, or anyone who used Claude's memory before this week, here's exactly what moved.

What actually changed

• Topics, not a black box. Everything Claude remembers is listed as an individual, editable entry under Settings > Memory > Topics. You can open, correct, or delete any single one. Anthropic's own example is fixing a company name once and having every future conversation pick up the correction, rather than re-explaining it each time.
• Written live, not summarized afterward. Claude now adds to memory during a conversation as things come up, instead of generating a summary once the chat ends. That's part of what makes the chat-to-Cowork handoff close to instant.
• Unified across chat and Cowork, but only in the cloud. Per Anthropic's support documentation, memory now flows both directions between Claude.ai chat and Claude Cowork: something you established in a chat is available when a Cowork task runs, and the reverse. This applies to cloud-run Cowork sessions, now the default execution mode. Anthropic states plainly that Cowork sessions run locally on your own computer, the older, non-default mode, don't use memory at all.
• Projects stay siloed. Memory tied to a Claude Project remains separate from your general memory pool and from other Projects. Anthropic frames this as a deliberate boundary, not a gap: a "household chores" project's memory doesn't leak into a "research" project.

The sensitive-topics setting, and what's always excluded

Off by default, Claude does not save health information, race or ethnicity, religious beliefs, political views, gender identity, or similar categories to memory. A new toggle, "Include sensitive topics in memory" in Settings > Memory, lets you turn that on, useful if, say, you want Claude to remember a dietary restriction without re-explaining it every time you ask for recipes. Turn it on and it only applies going forward, not retroactively, and Claude notifies you in-app each time it actually saves something in that category.

A separate, harder line sits underneath the toggle: government ID numbers (including Social Security numbers), criminal history, immigration status, and financial account numbers are excluded from memory regardless of the sensitive-topics setting, along with anything that would violate Anthropic's Acceptable Use Policy. That's not a preference you can override.

Who gets it, and when

Plan                What changed today
--------------------------------------------------------------------------------
Free, Pro, Max       On by default, web + desktop + latest mobile app version.
Team, Enterprise      Part of this update too, but OFF by default. An admin must
                      enable it org-wide; individual users then opt in.
HIPAA / public-sector /
custom-retention orgs  Not available at all under those agreements.

Source: Anthropic's own support documentation, verified 26 Aug 2026.

This update is scoped to the consumer product: Claude.ai chat and Claude Cowork. It does not touch the Claude API or Claude Code, which have their own, separate memory-adjacent features (an API "memory tool" for long-running agents, and Claude Code's CLAUDE.md-based project notes) that this announcement doesn't change. If you only use Claude through the API or the CLI, none of the above applies to you.

If you used Claude's memory before this week: the September 9 deadline

Anthropic's support page, checked directly today, carries an active notice: users are being migrated off the previous memory system, and if Claude appears to have forgotten something from before the migration, the fix is to go to Settings > Memory and export the legacy memory before 9 September 2026. After that date, the export option disappears. Anthropic's guidance for anything the migration dropped is to paste the exported legacy memory back into a conversation, calling out the specific part that seems to have been forgotten. If you've been a Claude memory user for a while and haven't looked at this screen recently, this is worth five minutes before the deadline, not after.

What to actually do with this

• Audit what's already there. Settings > Memory > Topics is a real list, not a settings toggle. It's worth a look even if you've never touched Claude's memory before, since it may already contain more than you expect.
• Decide on the sensitive-topics toggle deliberately. Leaving it off is the safer default on a shared or work device; turning it on makes sense mainly for a genuinely personal, single-user account where the convenience (not re-explaining a health or dietary detail every session) outweighs the tradeoff.
• Know that Incognito isn't instant deletion. Claude's Incognito mode skips memory and chat history. On Team and Enterprise plans specifically, Incognito chats still appear in org-level data exports under the organization's retention policy, so admins can still see them even though memory and history don't.
• If you use both chat and Cowork, expect context to follow you. A research thread you had in chat can now surface inside a Cowork task automatically. Useful for continuity, but worth knowing if you were relying on the two being separate.
• Team and Enterprise users: an admin has to enable this first before anyone in the organization gets it, and it's still gated per-agreement (HIPAA, public-sector, and custom-retention orgs don't get it at all).

How this compares to ChatGPT and Gemini

OpenAI took a different direction with ChatGPT's memory earlier this year: since a June 2026 overhaul several outlets called "Dreaming," it increasingly synthesizes what it knows about you in the background, automatically, rather than relying only on a manually curated list you build with "remember this." Google's Gemini sits closer to Claude's model, with an explicit "Saved Info" list you review directly. Claude's update keeps it at the more transparent end of that range: nothing gets inferred silently, and everything lands as a discrete, named topic you can inspect and correct. For the full mechanics of all three, including where Gemini Notebook fits in, see our ChatGPT vs Claude memory comparison.

What Anthropic hasn't said

• How topics are actually generated. Anthropic hasn't published the mechanism for how conversation content gets clustered into discrete topic entries. Treat "topic" as a product term, not a documented technical process.
• Team/Enterprise rollout timing. Anthropic's announcement doesn't state a rollout date for Team and Enterprise at all, only that admins control availability once it's enabled for their organization.
• Independent user experience so far. This update is barely a day old at publication time; there isn't yet a meaningful body of independent user reports on how the migration or the new sensitive-topics behavior holds up in practice. Worth revisiting in a few weeks.

People also ask

What exactly changed in Claude's memory on 25 August 2026?

Memory became a list of individually editable topics under Settings > Memory, shared between Claude.ai chat and Claude Cowork (previously separate), written live during conversations, plus a new toggle to include sensitive topics like health or political views, which stay off by default.

How do I see everything Claude remembers about me?

Go to Settings > Memory > Topics. Every saved memory is listed individually there, and you can open, edit, or delete each one.

Does turning off memory delete what's already saved?

No. "Pause memory" stops Claude from using or adding to it but keeps existing entries; "Reset memory" is the separate, permanent option that deletes everything, and Anthropic states it can't be undone.

Will Claude remember my health or political views?

Not by default. Those categories, along with race, ethnicity, religion, and gender identity, are excluded from memory unless you turn on "Include sensitive topics in memory" yourself. Even then, Claude notifies you each time one is saved.

Does this update apply to Claude Code or the API?

No. It's scoped to the consumer product: Claude.ai chat and Claude Cowork. The API has a separate, unrelated memory tool for long-running agents, and Claude Code has its own CLAUDE.md-based project notes; neither is affected by this announcement.

I'm on a Team or Enterprise plan, do I have this automatically?

No. It's part of the same update but off by default; an organization admin has to enable it, after which individual members can opt in. It's also unavailable to orgs under HIPAA, public-sector, or custom-retention agreements.

I used Claude's memory before this update, do I need to do anything?

Check Settings > Memory. Anthropic's support page states legacy memory can be exported only until 9 September 2026; after that, whatever you haven't exported is gone. If Claude seems to have forgotten something from before the migration, Anthropic's suggested fix is to paste the exported legacy memory back in and flag the missing part directly.