On 10 July 2026, Anthropic replaced Claude's single rolling memory summary with a set of individual, categorized entries that Claude reads and updates during your conversations. The practical result: you can open Settings > Memory, read each entry one by one, tell Claude to change a specific one, or delete a specific one, without wiping everything Claude knows about you.
One day earlier, on 9 July 2026, Anthropic shipped a second, quieter feature: Reflect. It lives at Settings > Reflect and produces a monthly recap of the topics you spent time on, your most active day, your peak hour, and observations about the kinds of tasks you bring to Claude. It is in beta for Free, Pro and Max users on the web and the Claude desktop app, and it only generates anything if memory is turned on.
Two changes, two days apart, and both of them touch the same underlying store. If you have not looked at Claude's memory panel since June 2026, the thing you find there now is different in shape, not just in polish. Here is what moved, what it fixes, what it does not fix, and how to switch it off if none of this appeals to you.
What actually changed on 10 July 2026
The unit of memory changed. Claude's help documentation now describes memory as a set of individual entries organized into categories, listed in a panel you can read in full. Before this, the durable thing Claude carried between chats was a summary: one compiled block of text describing you, rebuilt once a day rather than surgically amended.
Three things follow from that shape change. Claude updates entries live during a conversation rather than only compiling later. Entries are grouped by category, so the panel reads like a list rather than a paragraph about you. And because each entry is a separate object, each one can be targeted individually by you or by Claude.
Worth being precise about scope: as of July 2026, memory covers chats on the web, Claude Desktop and Claude Mobile, and Anthropic's help page states it is not currently available for Cowork. Memory from chats is documented for Free, Pro and Max plans. Searching past chats is a paid-plan capability, listed for Pro, Max, Team and Enterprise. Those are three separate mechanisms, and mixing them up is the most common mistake in coverage of this change.
- Chat search: Claude searches your previous conversations on demand when you ask it to reference something specific. Documented for Pro, Max, Team and Enterprise.
- Memory from chats: the entries panel at Settings > Memory, generated automatically from your conversations. Documented for Free, Pro and Max.
- Project memory: each project keeps its own separate memory space and its own project summary, so context inside one project stays out of the others.
Almost nobody sits down and chooses this, which is the detail most people miss
Memory is not something most people opt into deliberately. Anthropic's help page describes the new memory experience as the default for new users, with Free, Pro and Max users being migrated to it, and states that searching past chats is enabled by default once it reaches your account. For Enterprise organizations, Anthropic documents the organization wide Generate memory from chat history toggle as enabled by default. The toggle you want is Generate memory from chats at Settings > Memory, and Anthropic publishes no blanket statement that it starts on for every consumer account, which is exactly why it is worth opening the panel rather than assuming either way.
The default is the whole story here. A memory system you deliberately switched on is a feature. A memory system that is already running, already holding entries about your job, your projects and your habits, and now also feeding a monthly behavioural recap, is an account setting you should have read. Open Settings > Memory once, this month, and actually read the list.
Why one summary was the wrong shape
A single compiled summary has a specific failure mode: it is all or nothing. If one sentence in it was wrong, stale or embarrassing, your realistic options were to reset the whole thing and start over, or to live with it. Correcting one fact meant sacrificing every correct fact alongside it.
That failure mode compounds over months. Suppose Claude concluded in March that you were preparing for a job interview, and it is now July and you took the job. In a summary world, the outdated sentence keeps colouring answers until you burn the entire memory down. In an entries world, you open the panel, find that entry, and remove it. The rest of what Claude has learned about your writing style, your stack and your recurring projects survives.
This is the same reason password managers store credentials as records rather than one text file. Granularity is not a cosmetic improvement. It is the difference between a store you can maintain and a store you periodically nuke.
| What you are doing | Memory summary (before July 2026) | Memory entries (since 10 July 2026) |
|---|---|---|
| Unit of storage | One compiled summary describing you | Individual entries grouped into categories |
| Fixing one wrong fact | Practically means resetting everything | Edit or delete that single entry |
| Seeing what is stored | Read a block of generated prose | Read a categorized list in Settings > Memory |
| When it updates | Compiled and refreshed once every 24 hours | Read and updated during conversations |
| Cost of a mistake | High, because correction is destructive | Low, because correction is targeted |
How to edit or delete a single entry
Go to Settings > Memory. Anthropic's help page describes the Memory panel as listing everything Claude remembers, grouped by category. From there you can use the box to tell Claude what to change or remove, or select Delete to remove an entry outright. You can also just tell Claude in a chat what to change or remove. Anthropic states edits made that way apply immediately to your next conversation, rather than waiting on the daily rebuild the legacy summary depended on.
Two coarser controls sit alongside the per-entry ones. Pause memory keeps what already exists but stops Claude both using it and creating anything new, which is the right choice for a temporary stretch of work you do not want absorbed. Reset memory permanently deletes all memories, which is the nuclear option that used to be your only real option. Incognito chats are the third lever: Anthropic states that in an incognito chat Claude will not remember the conversation, so it is not saved to memory or to your chat history.
Use incognito before the sensitive conversation, not cleanup after it. Deleting an entry removes what Claude wrote down, but incognito prevents the entry from ever being written. If you are about to discuss salary, medical detail, a legal matter or somebody else's private information, switch to an incognito chat first and save yourself the audit later. One caveat: Anthropic's incognito help page states these chats are still retained for 30 days for safety purposes, so incognito keeps them out of memory rather than making them vanish.
What Reflect actually shows you
Reflect, announced by Anthropic on 9 July 2026, is a recap of how you have been using Claude rather than a recap of what Claude knows about you. Anthropic describes it as a summary covering key topics, usage patterns, and the types of tasks you often work through, viewable across the past 1, 3, 6 or 12 months. It is in beta for Free, Pro and Max users who have memory turned on, on the web and the desktop app, with Cowork conversation reflection arriving later. Anthropic also states the recap is only generated when you open the Settings > Reflect page, so nothing is compiled for you in the background if you never go there.
- The topics you have spent your time on, drawn from your conversations.
- Your most active day and your peak hour of the day.
- Observations about the kinds of tasks you bring to Claude and how you work with it.
- A time range selector for the past 1, 3, 6 or 12 months.
- Anthropic states that Reflect does not draw on incognito chats, and does not pull in the underlying files from your connected tools.
Shipped the same day, Settings > Time and focus adds optional break reminders and quiet hours, which tells you how Anthropic is positioning this: digital wellbeing, not analytics. Anthropic says it developed the feature with wellbeing researchers, including people from the MIT Media Lab's Advancing Humans with AI program, the Digital Wellness Lab at Boston Children's Hospital, and the Family Online Safety Institute.
The privacy question a monthly behavioural recap raises
The honest framing is this: Reflect does not create new data collection so much as it makes existing data legible. Your conversation history and your memory entries already existed. Reflect turns them into a profile of your habits, complete with the hour of the day you are most likely to be typing into Claude. Nothing about that is hidden or sinister, and Anthropic put a screen in front of it rather than keeping it internal. But a rendered behavioural profile is a different object from raw logs, and it is reasonable to have a view on whether you want one.
Two things reduce the surface meaningfully. Incognito chats are excluded from Reflect, so anything you route through incognito never appears in the recap. And Anthropic states that Reflect draws on the conversation, not the source material behind it: if you asked Claude to summarize your inbox, that summary can surface in your reflection while the underlying emails do not.
The dependency runs one way and it is useful to know it. Reflect requires memory to be on. So the switch that governs whether Claude builds entries about you is the same switch that governs whether a monthly recap of your behaviour gets generated. Turning memory off turns both off.
How to check whether you have the new version yet
The tell is which settings section the feature lives under. Anthropic's help page states that if you see Memory in Settings > Capabilities, you are on the legacy memory experience. The new experience surfaces as its own Settings > Memory panel with the categorized entry list. That single check answers the question in about ten seconds.
- Settings > Memory showing a categorized list of entries: you are on the July 2026 experience.
- Memory sitting under Settings > Capabilities: you are still on the legacy experience and the entry list has not reached you.
- No Settings > Reflect section: memory is off, you are on Claude Mobile or on a Team or Enterprise plan, or the beta has not reached your account yet.
- Anthropic describes the rollout as gradual, with the new experience as the default for new users and Free, Pro and Max users being migrated to it.
Anthropic is explicit about the rest: Team and Enterprise plan admins will hear about a rollout separately, and in the interim users on those plans stay on the legacy experience. So if you are on a workspace plan and still see the legacy panel in July 2026, that is expected rather than a bug. On Enterprise plans an Owner controls whether memory is available to you at all. Team plans have no organization level memory controls, so those settings are yours.
How to turn it off if you do not want any of this
Open Settings > Memory and switch off Generate memory from chats. That stops new entries being created. If you also want the existing store gone, use Reset memory, which Anthropic documents as permanently deleting all memories. If you only want a temporary break, use Pause memory instead, which retains what exists while Claude stops using it and stops adding to it.
Because Reflect requires memory to be on, turning memory off ends the monthly recap as well. There is no separate switch for it: Anthropic's help page states there is no dedicated toggle for the recap, so if memory is on, the recap is on too. Anthropic renames and reorganizes settings sections often, so treat every path in this post as the July 2026 layout and expect labels to drift.
The part most coverage of this update gets wrong
The framing everywhere in July 2026 is control. You can edit your memory now. You can delete individual entries now. Both are true, and both are genuine improvements over the reset-everything era. But editing rights are not ownership, and it is worth naming the difference plainly.
You do not decide what gets written. Claude generates entries from your chats, and your role is to review and correct after the fact. You do not decide where they live or in what format. Anthropic does document a way out: you can ask Claude to export your memory and paste it into another service, and Claude will import memories another assistant gave you, though Anthropic calls the feature experimental and still in active development. What comes out is text you copy into a file by hand, not a store another tool can read directly. Deleting an entry is still a permission the vendor grants you inside the vendor's product. That is a better deal than last month. It is not the same as owning your memory layer.
The practical consequence is that the context you build up inside Claude lives inside Claude by default. Moving it is a manual copy and paste with no shared format at the other end, so splitting your work across two assistants means maintaining two stores by hand. The July 2026 update makes Claude's version of this more pleasant to live with. It does not make your memory portable in any automatic sense.
That is where a separate memory layer does something different. MemX is a place you put the source material yourself: photos, PDFs, voice notes and chats, read with OCR and transcription, searchable in plain language with the original document shown alongside the answer. You choose what goes in rather than reviewing what an assistant inferred, and it is private by architecture, with per-user isolation, customer-managed encryption keys, encryption at rest and on-device capture. It does not replace Claude's memory of your conversations. It means the documents and notes you actually depend on are not sitting inside one vendor's chat product.
Do both, if you like. Leave Claude's entries on for conversational continuity, read the panel once a month, delete what is stale, and keep the material you would be annoyed to lose somewhere you control.
01Can I delete just one Claude memory now?
Yes. Since 10 July 2026, memory is stored as individual categorized entries. Open Settings > Memory, find the entry, and select Delete, or use the box to tell Claude what to change. Anthropic states edits made by telling Claude in a chat apply immediately to your next conversation.
02Is Claude memory on by default?
Anthropic documents the new memory experience as the default for new users, with Free, Pro and Max users migrated to it, and searching past chats as enabled by default. It publishes no blanket statement that memory generation starts on for every account, so check the Generate memory from chats toggle at Settings > Memory.
03What is Claude Reflect and do I have to use it?
Reflect, announced 9 July 2026, is a monthly recap at Settings > Reflect showing your topics, most active day, peak hour, and how you work with Claude. It is beta on Free, Pro and Max, web and desktop, and only runs if memory is on.
04Does Reflect see my private or incognito chats?
No. Anthropic states Reflect does not draw on incognito chats, and does not pull in the underlying files from connected tools. A summary you generated can appear in the recap even when the source documents behind it do not.
05Why do I still see Memory under Capabilities in Claude?
That means you are on the legacy memory experience and the July 2026 entry list has not reached your account. Anthropic describes the rollout as gradual across Free, Pro and Max. Anthropic states that Team and Enterprise users stay on the legacy experience for now, pending a separate admin rollout.
