Claude Just Merged Its Memory — Here's What It Remembers Now

Open Claude right now — the app, not Claude Code — and ask it, “what do you remember about me?” A week ago, the answer would’ve depended on whether you asked in a regular chat or handed a task to Cowork, because those were two different memories that barely talked to each other. As of yesterday, that’s gone. There’s one memory now, and it follows you everywhere.

Anthropic shipped the change on August 25, 2026, and it’s already the kind of update that quietly rewires how the tool behaves without anyone sending you a memo. So here’s the plain version: what actually changed, what Claude keeps by default, what it won’t touch unless you say so, and the five minutes it takes to check.

What actually changed on August 25

Before this week, Claude chat and Claude Cowork — the version that works inside your files and apps on your computer — each built their own picture of you. Tell Claude in a regular chat that you’re a freelance copywriter working with three retainer clients, and Cowork wouldn’t know that. Hand Cowork a research task, and your next chat wouldn’t remember what it found.

Anthropic’s own release notes put it simply: memory now works as “a set of individual, categorized entries that Claude reads and updates during your conversations, replacing the previous daily memory summary.” That’s the real mechanical shift — not just chat-and-Cowork syncing up, but a move away from a once-a-day rollup toward something that updates while you’re mid-conversation.

Practically, it means this: mention in a chat that your manager hates long emails, and the next time you hand Cowork something addressed to that manager, it already knows to keep it short. Anthropic’s own framing is “the project you talked through, your manager’s preferences, or the client from last quarter” — and now Cowork starts from that instead of a blank slate.

One thing worth clearing up before you go further: if you searched “claude memory” today, you probably landed on pages about CLAUDE.md files, Claude Code’s memory tooling, or MCP memory servers. That’s a different, developer-facing thing. This post is about the memory feature inside the Claude app itself — the one under Settings, the one that’s on whether you’re a developer or not.

Source: Anthropic — “Claude’s memory works everywhere, and you decide what’s in it”

Who has it, and is it actually on

Yes — and this is the part that catches people off guard. Memory is on by default for Free, Pro, and Max plans, across web, desktop, and mobile. You didn’t opt in. Nobody asked. If you’ve sent Claude more than a handful of messages since yesterday, it’s already started building a profile.

Team and Enterprise workspaces are the exception — memory is off by default there, and an admin has to turn it on. If you use Claude through work, that’s worth confirming with whoever manages your workspace rather than assuming it matches what you see on your personal account.

This isn’t a US-only rollout, either. Search interest in “claude memory” spiked the same day in Germany as it did in the US, which tracks — the feature shipped globally, not region by region. If you’re using Claude anywhere outside the US, the defaults described here apply to you too; there’s no separate “international” version of the memory settings.

Here’s the part that actually matters for anyone using Claude for anything involving other people: memory doesn’t apply retroactively. Whatever Claude picked up from your chats before yesterday under the old system carries forward as-is. The new entry-based system governs everything from here on. So if you’ve never looked at what’s stored, today is a reasonable day to look — not because something changed for the worse, but because the feature just got a lot more capable, which means whatever’s already in there now reaches further than it used to.

What Claude keeps out — and what it never touches at all

Anthropic built two separate walls here, and they work differently. Worth knowing the difference.

Wall one: sensitive topics, off by default, opt-in. Health information, religious beliefs, political views, race or ethnicity, and gender identity don’t get saved to memory unless you specifically turn on “Include sensitive topics in memory” in Settings. Mention in passing that you’re managing a chronic condition, and by default, that doesn’t become a stored memory entry. If you flip the toggle on, it starts saving those going forward — it doesn’t retroactively scoop up sensitive details from chats before you opted in. And if you later turn it back off, Anthropic says it removes what was already saved under that category.

Wall two: some things are never saved, full stop, no toggle. Government ID numbers, criminal history, financial account numbers, and immigration status don’t go into memory regardless of your settings. That’s a hard product boundary, not a preference.

Read that carefully, though — “not saved to memory” and “not part of the conversation” are two different sentences. If you paste your passport number into a chat to get help drafting a form, Claude won’t turn that into a standing memory entry about you. But it’s still in that conversation, subject to Anthropic’s normal chat-retention rules. The memory feature is a control on what gets remembered long-term, not a redaction tool for what you type in the moment. More on that distinction below, because it’s the one people get wrong most often.

The five-minute setup: see it, edit it, turn it off

You don’t need to trust a blog post’s description of what’s in your memory. You can just look. Here’s the walkthrough.

Step 1 — Find it. Open Claude, click your profile icon, go to Settings → Capabilities, and look for Memory. You’ll land on a page with everything Claude has stored, sorted into categories — work context, preferences, ongoing projects, people you’ve mentioned.

Step 2 — Read what’s actually there. Click into a category. Expect to see short, factual-sounding notes: “Works as a freelance copywriter, primarily B2B SaaS clients.” “Prefers responses under 200 words unless asked for detail.” “Currently planning a product launch for early October.” It reads less like a diary and more like notes a sharp assistant would jot down.

Step 3 — Edit or delete what you don’t want. Every entry has an edit and a delete option. You can also just tell Claude directly in a chat — “forget that I mentioned my old job” works, and Anthropic says it updates the memory accordingly. No need to hunt through settings for a one-off correction.

Step 4 — Decide on sensitive topics. If you want Claude remembering health or personal-belief context (useful if you’re using it for something like meal planning around a condition, or a values-driven writing project), flip Include sensitive topics in memory on. Otherwise leave it — the default is designed to be the cautious choice.

Step 5 — Know the difference between pausing and resetting. This is where most people get confused, so slow down here:

• Pause memory — Claude stops making new memories and stops using the ones it has, but keeps them stored. Anything you discuss while paused doesn’t get added even if you turn memory back on later. Good for a single sensitive conversation you don’t want influencing anything downstream.
• Reset memory — permanently deletes everything. Anthropic is explicit that this can’t be undone. Good for a genuine clean slate — new job, new client roster, or just wanting to start over.

If you re-enable memory after pausing, both regular and sensitive-topic memory (if you’d opted in) come back on. It’s not a one-way ratchet.

Real example: what this looks like end to end

Take a specific, ordinary case. Say you’re a solo consultant. On Monday, in a regular Claude chat, you mention you’re working with a client called Meridian Logistics on a supply-chain audit, that the engagement runs through November, and that your contact there — call her the project lead — prefers bullet points over prose in status updates.

Under the old system, that context lived in chat. If you opened Cowork on Wednesday and asked it to draft a status update for Meridian, it wouldn’t know any of that — you’d paste in the context again, or write it from scratch.

Under the new system: you open Cowork on Wednesday, say “draft this week’s Meridian status update,” and Claude already has the engagement name, the timeline, and the bullet-points preference loaded in. It writes a first draft that doesn’t need the preamble. That’s the entire pitch of the unification — less re-briefing, more picking up where you left off.

The trade-off runs the other direction too, and it’s worth sitting with for a second. If you’d rather Cowork not know about Meridian for some reason — maybe you’re testing a template and don’t want client specifics bleeding in — you now have to actively manage that, either by pausing memory for that session or by not mentioning Meridian in chat in the first place. The convenience and the leakage risk are the same mechanism.

A second example, further from client work: say you’re planning a kitchen renovation and you’ve been using Claude chat to compare contractor quotes, track a budget ceiling, and note that you’re leaning toward one bid because the contractor could start in three weeks instead of six. Weeks later, you open Cowork to draft a polite decline email to the contractor you’re not hiring. Under the new system, Cowork already has the budget context and the timeline reasoning — it can draft something specific (“we ultimately went with a contractor who could start sooner”) instead of a generic template you’d have to hand-edit. Low-stakes example, but it’s the same mechanism as the consulting one: say it once, it’s available everywhere, for better or worse.

How people are actually reacting

Anthropic’s own announcement post picked up thousands of likes within hours, and the tone of the reaction online skews toward relief more than alarm — a lot of “finally” and not much “wait, what.” Anthropic’s product team framed it as a direct response to user feedback, and the top replies mostly matched that: fewer re-briefings, less copy-pasting context between chat and Cowork.

The more interesting reaction, if you dig past the announcement thread, is a kind of shrug about loyalty. One widely-shared take put it bluntly: people switch between ChatGPT and Claude without a second thought regardless of which one “remembers” them better — the idea that memory would lock you into one assistant the way, say, your photo library locks you into a phone brand, just isn’t playing out that way. If you’ve been assuming a growing memory profile is a reason to stick with one tool, the early evidence says most people don’t actually behave that way.

There’s a smaller, genuinely useful signal buried in the reaction too: a handful of users flagged that memory felt like it was updating too aggressively right after the rollout — noticeably more often mid-conversation than they expected. If that happens to you in the first few days, it’s worth checking Settings → Memory to see what’s actually landing in there; a fast-updating memory is still an editable one.

What’s notably thin, at least in the first 24 hours: complaints from professionals specifically worried about client confidentiality now that memory spans both chat and Cowork. That’s either a sign the sensitive-topics wall is doing its job, or a sign most people haven’t thought it through yet. Given how new the change is, it’s worth assuming the latter until proven otherwise — which is exactly why the “what to keep out” section below exists.

Claude vs. ChatGPT vs. Gemini: how the memory models actually differ

All three major assistants now do some version of “remembers you across sessions,” but the mechanics — and what counts as deleted — are genuinely different. This matters more than the marketing copy suggests, especially the last column.

That last row is the one to remember if you only remember one thing from this table. Claude’s model — delete the topic, it’s gone — is the most direct of the three. ChatGPT and Gemini both require you to think about deletion at the conversation level, not the fact level, which means a single offhand detail can be harder to surgically remove than you’d expect.

What this means for you

If you’re a casual user who mostly asks Claude for everyday help — recipe ideas, travel planning, the occasional email draft — this is a pure convenience upgrade. Do nothing. Memory will quietly make Claude feel less like it forgot you overnight. Your first action: skim Settings → Memory once, just so you’re not surprised later.

If you’re a freelancer or consultant working with named clients — the Meridian example above is you. Your first action: before your next Cowork session, glance at what’s currently stored under work-context entries and delete anything tied to a client you don’t want auto-surfacing in other work.

If you’re in HR, legal, healthcare, or handle other people’s private information — treat this the way you’d treat any tool that just got a longer memory: the sensitive-topics wall is off by default, which is good, but it doesn’t cover names, case details, or workplace situations you’ve described in plain language. Your first action: read the “what to keep out” list two sections down, and pause memory before any session involving a specific person’s private situation.

If you use both Claude chat and Cowork regularly for the same work — you’re the exact audience this update was built for. Your first action: stop re-explaining project context to Cowork. Say it once in chat, confirm it shows up as a memory entry, and let it carry over.

If you manage a Team or Enterprise Claude workspace — memory is off by default for you, unlike consumer plans. Your first action: decide deliberately whether to turn it on, and if you do, tell your team explicitly rather than letting them discover it mid-task.

If you’re switching between Claude and ChatGPT or Gemini regularly — don’t assume “I deleted the chat” means the same thing everywhere. Your first action: if you need something gone from ChatGPT or Gemini, go find the standing memory entry (or every relevant past chat) — don’t just delete the conversation and assume the fact went with it.

If you’re a developer who also uses Claude Code — this update doesn’t touch Claude Code’s CLAUDE.md files or its separate memory tooling. That’s still its own system. Your first action: don’t conflate the two when you’re troubleshooting — “memory” in the app and “memory” in Claude Code are unrelated systems that happen to share a name.

If you share a device or a Claude account with someone else — family plan, shared work login, whatever the arrangement — memory is tied to the account, not the device. Your first action: if you’re sharing credentials with someone (not recommended, but common), assume they can see everything memory has stored, including anything you’ve mentioned about them. Worth a conversation before it becomes an awkward discovery.

Edge cases and troubleshooting

“I turned off sensitive topics, but Claude still seems to know about my health stuff from a month ago.” That’s expected if you’d previously opted in and are now opting back out — Anthropic says turning the setting off removes existing sensitive-topic memories, but there can be a short lag, and anything discussed before you ever opted in wasn’t captured as a memory entry in the first place (it just existed in that chat transcript).

“I deleted a memory entry, but Claude still referenced it in a new chat right after.” Give it a beat — entries update live, but a chat already in progress when you make the edit may not immediately re-read the updated memory set. Start a fresh conversation and check again.

“Does this affect Claude Code or the API?” No. This unification is specifically the consumer app’s chat-and-Cowork memory. Claude Code’s memory files and the API’s handling of context are separate systems that Anthropic hasn’t merged into this.

“I use Claude on my phone and my laptop — is memory the same on both?” Yes. It’s account-level, not device-level, so what Claude remembers follows your login across web, desktop, and mobile.

“Can I have memory on for personal use but off for anything client-related?” Not as a permanent split — there’s no per-project memory toggle yet. Your workaround is the pause/resume flow: pause before anything client-specific, resume after, or keep a separate account for client work if the split matters enough.

“What happens to memory if I cancel my subscription and go back to Free?” Anthropic hasn’t published specific guidance on this scenario as of this writing; the safest assumption is that your account-level memory persists across plan changes unless you explicitly reset it, since memory is tied to the account rather than the tier.

“Someone told me Claude’s memory used to be ’three separate pools that sync once a day’ — is that true?” That specific description isn’t backed by anything in Anthropic’s own documentation, current or past. What we can verify: the prior system used a daily memory summary, and chat and Cowork didn’t share memory before this update. If you’ve seen that “three pools” framing elsewhere, treat it as an oversimplification that’s now doubly outdated.

“I’m worried this data trains the model.” Memory and model training are two separate settings — don’t conflate them. Whether Anthropic can use your chats to improve Claude depends on a distinct opt-in setting (or whether a conversation gets flagged for safety review), not on whether memory is on. Check both settings independently if this matters to you; see the next section for specifics.

What this can’t do

It won’t tell you exactly how it’s weighing conflicting memories. If you told Claude in March you hate bullet points and in July that you love them, there’s no visible “which one wins” logic exposed to you — you’ll just have to check the current entry and correct it if it’s stale.

It doesn’t give you project-level or client-level separation. As covered above, memory is one pool per account. If you need hard walls between different clients’ contexts, you’re managing that manually or using separate accounts — there’s no built-in workspace-per-client structure on consumer plans.

It’s not a substitute for your own records. Memory entries are Claude’s summarized notes, not a verified, auditable log. Don’t rely on it as your system of record for anything you’d need to produce later — for that, keep your own notes.

It doesn’t protect what you paste into a conversation, only what gets promoted into long-term memory. Repeating from earlier because it’s the mistake people actually make: pasting sensitive information into a chat still puts it in that chat’s transcript, governed by Anthropic’s regular retention rules — it just doesn’t automatically become a standing memory entry unless it fits a category memory tracks.

It doesn’t change your data-training settings. Turning memory on, off, or resetting it entirely is unrelated to whether your conversations are used to improve Claude. Those live under a separate privacy setting, and Anthropic’s own retention documentation notes that if you’ve allowed your chats to be used for model improvement, that data can be retained in de-identified form for up to five years in training pipelines — a completely different clock than the memory feature runs on.

What professionals should keep out of memory entirely

This is worth its own section, because “memory is convenient” and “memory is safe for client work” are different claims, and the gap between them is where people get burned.

Treat Claude’s memory as a convenience layer, not a client-record system. Even where Anthropic says a category is never saved, the underlying conversation may still exist elsewhere in its systems, subject to normal retention and safety-review rules. A few categories to keep out of prompts and memory as a matter of habit, not paranoia:

• Login credentials, API keys, passwords, or access tokens — for anything, ever
• Full contracts, unreleased pricing, or negotiation positions
• Government ID numbers, financial account numbers, or immigration details (Claude won’t memory-store these, but don’t paste them into chats either)
• Protected health information, diagnoses, or therapy notes
• Any employee, customer, or client data that identifies a real person alongside sensitive details — compensation, performance issues, health status
• Exact proprietary source code or unreleased product plans, if your employer would object to that leaving your machine

The practical habit that solves most of this: use minimal, reversible context. Write “an enterprise client in a regulated industry” instead of the client’s actual name. Swap exact figures for ranges. Use a placeholder like [CLIENT_A] in your prompts. And for anything genuinely sensitive, pause memory first, then check afterward that no new entry got created from that session.

Frequently asked questions

Does Claude have memory like ChatGPT? Yes, and as of August 25, 2026, Claude’s version is arguably more unified than ChatGPT’s — one memory pool spans chat and Cowork, versus ChatGPT’s two-layer system of saved memories plus chat-history reference.

Is Claude’s memory on by default? Yes, for Free, Pro, and Max plans. It’s off by default for Team and Enterprise workspaces until an admin turns it on.

How do I see what Claude remembers about me? Settings → Capabilities → Memory. Everything is listed by category, and you can open, edit, or delete individual entries.

Does turning off memory delete what’s already stored? No — that’s what “pause” does versus “reset.” Pausing stops new memories and stops using existing ones, but keeps them. Reset permanently deletes everything, and Anthropic says that action can’t be undone.

Does Claude remember sensitive topics like my health or beliefs? Not unless you turn that on specifically. It’s a separate opt-in toggle from general memory, and it’s off by default.

Will Claude Code or the API pick up my chat memory? No. This update covers the Claude app’s chat-and-Cowork memory only. Claude Code has its own separate memory system built around files like CLAUDE.md.

Can I stop Claude from remembering one specific conversation without turning memory off entirely? Yes — pause memory before that conversation and resume after, or use an incognito/temporary chat where available, which Anthropic says can help keep a conversation from being remembered.

Does memory mean my data is being used to train Claude? Not automatically. Memory and model training are governed by separate settings. Check both if that distinction matters to you — see the “What this can’t do” section above.

What’s the difference between deleting a chat and deleting a memory? Deleting a chat removes that conversation from your history. It does not automatically delete any memory entry Claude derived from it — those live separately and need to be deleted on their own.

Is this the same as the “Claude Dreaming” feature I’ve read about? No — Claude Dreaming was a separate, earlier concept some third-party writeups used to describe persistent memory behavior. This is the current, official mechanism: Settings → Memory, entry-based, spanning chat and Cowork. If you’re reading an older comparison piece that predates August 25, treat its description of “how Claude remembers” as outdated.

Does pausing memory affect Cowork tasks that are already running? Anthropic’s guidance covers new conversations going forward once you pause — a Cowork task already mid-run when you pause isn’t specifically addressed in the current documentation, so the safe assumption is to pause before starting anything you don’t want remembered, not mid-task.

The bottom line

The headline here isn’t really “Claude got smarter.” It’s that the boundary between “things I told Claude in chat” and “things Claude knows when I hand it a task” just disappeared — and that’s a genuinely useful change for anyone using both surfaces, with a genuinely real cost for anyone who wasn’t thinking about what they’d been casually mentioning. The good news: the controls are real, they’re specific, and checking them takes less time than reading this sentence took you to get here. Go look at Settings → Memory once. It’s worth the five minutes.

If you’re building AI into how you actually work — not just chatting with it, but handing it real tasks — our Claude for Small Business course walks through exactly this kind of workflow-and-privacy tradeoff, with the guardrails built in from the start.

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 — Let Claude use your computer in Cowork
• Anthropic Help Center — Release notes
• Anthropic Privacy Center — How long do you store my data?
• Anthropic Privacy Center — Is my data used for model training?
• 9to5Mac — Anthropic update unifies memory feature across Claude Cowork and chat
• TechCrunch — Claude Cowork finally remembers what you told the app in chat
• The Register — Claude and Cowork now share what they know about you
• SiliconANGLE — Anthropic updates Claude’s memory to enhance customization and protect sensitive topics
• OpenAI — Memory and new controls for ChatGPT
• OpenAI Help Center — How does “Reference saved memories” work?
• Google — Get personalization in Gemini Apps
• Google — Get personalization with memory of your past Gemini chats