Stop re-explaining yourself every session. Decide what should carry forward, what should be looked up, and what should stay apart.
Chat search looks up past conversations when you ask. Memory saves topics as you chat and applies them to new chats, including cloud Cowork tasks. Use memory for what stays true and search for what happened, draw boundaries with projects or memory-off chats before context blends, and review Settings > Memory monthly. Pause, reset, delete a topic or go incognito, depending on how much you want gone.
The first few minutes of a lot of Claude conversations go on re-explaining: your role, the project, who the reader is, how you like things written. Claude now has two separate features aimed at that, and they do different jobs. Chat search looks something up when you ask. Memory carries standing context into new conversations without being asked. People tend to switch one on and expect the other, and that is where the confusion usually starts.
The skill that pays off is deciding what should carry forward, what should be looked up, and what should stay apart. Everything below about how the features behave comes from Anthropic’s own help articles, mainly the one on chat search and memory, which I read on October 15, 2026.[1]
Search past chats is available on the paid plans (Pro, Max, Team and Enterprise) on the web, Claude Desktop and Claude Mobile. You ask Claude to find what you discussed before, and it pulls the relevant material into the current chat. Anthropic says these searches use retrieval-augmented generation, which in plain terms means Claude looks things up in your history rather than holding all of it in mind, and each search shows up in your chat as a tool call.
Memory works differently. It’s on by default for Free, Pro and Max plans. On Team and Enterprise plans, the organisation’s owner decides whether memory is available, and it stays off for each member until they turn it on.[1] Claude saves it as a set of individual topics while you chat, not as a summary written once the conversation ends. Anthropic’s own example is a deadline that moves: mention it once and your next conversation already knows.
| Search past chats | Memory | |
|---|---|---|
| What it’s for | Finding and reusing specific past work | Standing context carried into new chats |
| How it starts | You ask, and a search tool call appears in the chat | Claude saves topics as you chat, and you can say “remember this” |
| Plans | Pro, Max, Team, Enterprise | On by default for Free, Pro, Max. Team and Enterprise: owner enables it, then each member turns it on |
| Where to switch it off | Settings > Memory, “Search and reference chats” | Settings > Memory, “Generate memory from chats” |
| What it leaves out | Incognito chats. Project chats are searched only inside their own project | Incognito chats, sensitive topics by default, and a short list Claude won’t save at all |
Summarised from Anthropic’s help article, read October 15, 2026. Menu names are for the current Settings > Memory layout.
Here is the rule I’d work from.
Let memory hold what stays true, and use search to find what happened.
Your role, your reader and your preferred format belong to memory. “What did we decide about the vendor shortlist” belongs to search. Memory holds topics, not transcripts, so it’s the wrong place to look for the exact wording of last month’s decision. Search runs when you prompt it, so it’s the wrong place to keep your standing preferences, and you end up retyping them each time you open a chat. Our earlier piece on how ChatGPT and Claude remember you covers the first rollout; this guide covers the current Settings > Memory experience.
You trigger a search by asking for one, so the wording matters. Anthropic’s examples are plain: “What did we discuss about [topic]?”, “Can you find our conversation about [subject]?” and “Let’s continue where we left off with [project].”[1] Claude can search all your chats outside projects. Inside a project, it searches only that project’s conversations.
A finance lead closing the month asks, in a fresh chat, what was agreed last quarter about how to treat a supplier accrual. The reply opens with a search tool call and comes back with a short summary and a link to the original conversation. That link matters more than the summary. Anthropic says references to previous conversations come with citations back to the original chats, so you can open the source instead of trusting a paraphrase.[1]
Three habits make the request work better, and they come from my own judgment, not from Anthropic’s documentation:
| 1. Recover a decision. Search my past chats for our discussion of [TOPIC]. Tell me what we decided, what was still open, and which chat each point came from. If you can’t find it, say so instead of guessing. 2. Pick up where you stopped. Let’s continue where we left off with [PROJECT]. Find the last conversation about it, tell me in five lines where we stopped, then ask me what I want to do next. 3. Reuse something that worked. Find the [DOCUMENT, e.g. client update] we wrote last time. Use its structure and tone for this week’s draft, and quote the lines you’re reusing so I can check them. 4. Check before you rely on it. Before you answer, look through our past chats for anything on [TOPIC] that contradicts what I just said, and list it with the chat it came from. |
Replace the bracketed parts. Each prompt asks for the source chat, so you can open it and confirm before the answer goes anywhere.
Treat what comes back the way you’d treat a colleague’s recollection of a meeting: useful, and worth checking against the minutes before you reuse a number or a commitment. If a long thread has started to wander, a new chat that pulls in the old one by search is often cleaner than carrying on in the same window. Why long threads degrade is covered in our explainer on context rot.
Anthropic lists the everyday context Claude remembers: your role, projects and professional context, the people and places in your work and life, your communication preferences and working style, technical preferences and coding style, and project details and ongoing work.[1] That’s a working profile of you, not a transcript. You can also tell Claude to “remember this” to save something directly.
Two limits are worth knowing before you put anything personal in. By default, Claude doesn’t store topics on personal or sensitive subjects, such as health, race, ethnicity, religious beliefs, politics and gender identity. You can opt in with Include sensitive topics in memory, under Settings > Memory. Once it’s on, Claude saves those topics going forward, not retroactively, and a notice appears above the message box each time it saves one. If you decline the notice or turn the setting off later, Claude removes the sensitive items it already saved.[1]
Some information isn’t saved even if you ask: government ID numbers, criminal history, financial account numbers and immigration status. Claude tells you when it can’t save something for this reason. My own view is to leave the sensitive-topics setting off unless there’s a specific reason to turn it on, because a working profile rarely needs it.
A marketing manager who writes the same kind of update each week can use that. She tells Claude once that her reader is a finance director who wants the answer in the first line and the figures in a table, and later chats start from there. Memory is built for exactly this kind of standing preference. Paste something like the briefing below, then open Settings > Memory and check what was saved.
| Please remember this about how I work. My role: [e.g. head of finance operations at a mid-sized distribution business]. Who I write for: [e.g. a finance director who wants the answer in the first line and figures in a table]. How I like drafts: [e.g. plain language, short paragraphs, UK spelling, assumptions listed separately]. What I’m working on this quarter: [e.g. the supplier consolidation and the year-end close]. Update these if I tell you something has changed. |
Keep it to working preferences. Leave out client names, figures and personal details you wouldn’t want stored.
If you work in Claude Code and want the project-file version of this idea, that’s a separate mechanism with its own rules. Our explainer on CLAUDE.md and MEMORY.md files covers it.
Memory gets risky when two contexts that shouldn’t meet end up in the same pile. The tool to reach for is the project. Each project has its own memory. Inside a project, Claude uses that project’s memory alongside your memory from chats outside projects, but not memory from other projects. New memory is saved to the project.[1]
There is one detail worth knowing. If you have memory turned on, Claude can reference a project’s memory in chats outside projects too. To keep a project’s memory fully separate, turn off Use account memory under Memory on the project’s page. A consultant who runs each client as a project and leaves that setting on has separated the conversations, but not necessarily the memory.
Draw the boundary before Claude learns anything: one client, one case, one project.
A boundary set in week one costs a minute. Set after three months of mixed chats, it means deleting topics by hand. Anthropic describes four options:
The memory-off setting has to be chosen before the first message, because Anthropic says you can’t change it once the chat has started. Incognito chats are retained for 30 days by default, or longer under an organisation’s own retention setting, and they aren’t used for training.[3] Incognito isn’t available inside a project, so in a project you use the memory-off switch.
Who needs which boundary differs by role. A consultant or agency lead needs a project per client. An HR partner handling individual cases needs the memory-off chat or incognito, and should keep named individuals out of anything saved. A marketer needs a project per brand voice, and a finance lead may need one per entity so periods and subsidiaries don’t blur.
Memory is shared between chat and Claude Cowork when Cowork runs in the cloud. What Claude remembers from your chats is available in cloud Cowork tasks, and what comes up in a task carries back to chat. Anthropic’s example is asking Cowork to draft an update for your manager: it already knows who that is and how they like updates written.[1] Whatever is in your memory is now shaping tasks that run without you watching.
The limit is on the local side. Cowork sessions that run on your own computer don’t use memory. Whether you have a choice depends on your version: with the new Claude experience, Anthropic says tasks run in the cloud and there’s no option to run them locally, while on older setups you can switch under Settings > Cowork with Run new tasks in the cloud.[2] Our decision guide on where Cowork tasks should live goes through that choice, and our plain-English guide explains what Claude Cowork is.
Two more details. You can run a single Cowork task without memory by clicking “+” and turning off Memory, but the setting is fixed when the task starts and can’t be changed later. And Anthropic’s cloud page says search of past cloud sessions is title-only for now, so you can find a session by its title but not by what’s inside it.
| Draft this week’s update to my manager on [PROJECT]. Use what you already know about who they are and how they like updates written. Before the draft, list the remembered details you’re relying on, one per line, so I can correct the ones that are out of date. Then write the update in under 200 words, with decisions needed from them at the top. |
The list in the middle is the point. It turns a hidden assumption into something you can check in ten seconds.
Memory is only as good as what’s in it, and you can see all of it. Open Settings > Memory, where everything Claude remembers is listed under Topics. Select a topic to read it, use the edit icon to change it, or select Delete to remove it. A fix applies to every conversation from then on.[1] You can also change memory from a chat by telling Claude what to remember, change or forget, and the update applies to your next conversation.
One mechanic catches people out. Deleting or letting a conversation expire doesn’t remove the memory entries generated from it. Those stay until you delete them as individual memories. If you delete a chat because it held something sensitive, check Topics for what it left behind.
Read what Claude has saved about you before you rely on it, the way you’d read a new colleague’s notes.
A short monthly check is enough to keep that honest. It’s my suggested routine, not an Anthropic feature.
| Step | What to do | What you’re looking for |
|---|---|---|
| 1 | Open Settings > Memory and read the Topics | Anything out of date, such as an old role, a finished project or a changed reporting line |
| 2 | Edit what’s slightly wrong, delete what’s wrong | Client or personal detail that shouldn’t be stored |
| 3 | Check what’s missing | Standing preferences you keep retyping. Tell Claude to remember them |
| 4 | Open each client project and check Use account memory | The setting matches the boundary you meant to draw |
Do it on the first Monday of each month, or when a project ends.
It also helps to be clear about who owns what. Claude saves and retrieves. You decide what’s worth carrying and whether to rely on what comes back.
| What Claude does | What you still own | How it gets checked |
|---|---|---|
| Saves topics from your chats and applies them to new ones | Deciding what belongs in memory and what stays out | The monthly review of Topics in Settings > Memory |
| Searches past chats and summarises what it finds | Deciding whether to rely on the summary | Open the cited chat before reusing a number or a commitment |
| Uses memory to shape Cowork tasks in the cloud | The final update and anything sent in your name | Ask Claude to list the remembered details it used, then correct them |
A Future Factors working view. Claude retrieves, you decide.
Each control does something different, and the names sound closer than they are. Pause memory keeps what Claude already knows, including sensitive topics if you have them on, but Claude won’t use memory or make new memories. Conversations while memory is paused aren’t added if you turn it back on, and unpausing turns both your memory and sensitive-topics memory on again. Reset memory permanently deletes everything, project memories included, and can’t be undone. If you re-enable memory afterwards, you start from scratch.[1]
Search has its own switch. In Settings > Memory, switching off Search and reference chats stops Claude searching your past chats while leaving memory as it was. Pick the lightest control that does the job.
| Control | Existing memory | New memory | Past chats |
|---|---|---|---|
| Pause memory | Kept, but not used | Not made | Search has its own switch |
| Reset memory | Permanently deleted, including project memory | Starts from scratch if you re-enable | Search has its own switch |
| Delete one topic | Only that entry removed | Continues | Unchanged |
| Search and reference chats off | Unchanged | Unchanged | Claude stops searching them |
| Memory off for one chat | Not used in that chat, and no past-chat search from it | Not saved to memory | Stays in history; findable from other chats |
| Incognito chat | Not used | Not saved to memory | Not in history, not searchable |
Based on Anthropic’s help articles on memory and on incognito chats, read October 15, 2026.
On Team and Enterprise plans there is an organisation layer. Owners turn memory on in Organization settings > Capabilities, and members then manage their own settings. Owners can’t view or edit a member’s individual memories. If an owner turns memory off for the organisation, all existing memory entries for everyone are deleted immediately. Anthropic also says memory isn’t available to organisations with HIPAA, public-sector or custom data retention agreements.[1]
So the settings are personal but the habits are shared. Teams can agree the conventions themselves, and a facilitated session shortens the learning curve, because the hard part is agreeing the habit, not finding the toggle. Training builds the capability, a project per client gives it somewhere to be used, and a monthly review keeps the habit going.
| Decision | Our default | Who checks, and when |
|---|---|---|
| What gets its own project | Each client or case, with Use account memory off | Team lead, when a new client starts |
| What goes in memory | Role, reader and format preferences. No client names or figures | Each person, in the monthly review |
| What needs incognito or memory off | Anything about an individual employee or candidate | Team lead, spot-checked quarterly |
A Future Factors working view. Adjust the defaults to your own data policy and ask your owner what your organisation allows.
The reference below sums up the rule, and the first step in the list is worth doing today.
| When you catch yourself… | Reach for | Where it lives |
|---|---|---|
| Retyping your role, reader or format | Memory | Tell Claude to remember it, then check Settings > Memory |
| Asking what was decided last month | Search past chats | Ask in plain words and open the cited chat |
| Mixing two clients or cases | A project each | Turn off Use account memory on the project’s page |
| Discussing something you want out of memory | Memory off for this chat | Click “+” before your first message |
| Discussing something you want out of history too | Incognito chat | Ghost icon, outside projects |
| Wanting a clean slate | Pause or reset | Settings > Memory |
Based on Anthropic’s help articles, read October 15, 2026.
| # | Do this | Done when |
|---|---|---|
| 1 | Read your Topics in Settings > Memory | You’ve edited or deleted anything wrong |
| 2 | Decide the sensitive-topics setting | You chose on purpose. Leave it off unless you have a reason |
| 3 | Paste the memory briefing | Settings > Memory shows your role, reader and format |
| 4 | Create a project for each client, case or brand | Use account memory is set deliberately on each |
| 5 | Use a search prompt the next time you re-explain a past decision | You opened the cited chat and checked it |
| 6 | Put the five-minute review in your calendar | The first Monday of the month is booked |
Steps 1 to 3 can be done in one sitting.
Start with one piece of recurring work, such as your weekly update, and see how much less you retype over two weeks.
No. Inside a project, search is limited to that project’s conversations, and Claude uses that project’s memory alongside your memory from chats outside projects, not memory from other projects. A chat outside a project can search your other non-project chats.
Not automatically. Anthropic says memory entries generated from a conversation aren’t removed when the conversation is deleted or expires. Open Settings > Memory, find the topic and delete it there. You can also tell Claude in a chat to forget something.
On Team and Enterprise plans, owners can’t view or edit your individual memories, and you control your own memory setting once an owner enables the feature. Anthropic’s article also says memory entries are included in standard conversation history exports and retained under the organisation’s policies, so don’t treat memory as invisible to your organisation.
Anthropic says you can import memories from other AI assistants or export your Claude memory for backup or migration, and describes the feature as experimental and still in active development. Check its separate help article for current steps.
The Memory switch in the plus menu only appears if memory is already on for your account, your organisation allows it and your plan includes it, and it doesn’t appear in incognito chats. If Memory isn’t under Settings, you may be on the legacy experience that some Team and Enterprise organisations still have, where it sits under Settings > Capabilities.
Everything about how the features work comes from three Anthropic Help Center articles, read on October 15, 2026: the main article on chat search and memory, the article on Claude Cowork in the cloud, and the article on incognito chats. We did not test the features in a live account, and Anthropic updates these pages often, so check your own Settings screen. The Stable Facts Rule, the Boundary Rule, the Read-It-Back Rule, the five-minute memory review and the suggested prompts are Future Factors working methods based on our own judgment, not Anthropic features. The scenes in this guide are illustrations, not client stories. We used no statistics.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.