What Claude Remembers Between Conversations
7 min read
In March, a marketing lead tells Claude that her team writes for finance directors, avoids exclamation marks, and never uses the word "solution." In May she opens a brand-new chat, asks for a webinar invitation, and the draft comes back in exactly that register without a single reminder. A week later she asks for something for a friend's birthday party and gets a paragraph written for finance directors, because the same background followed her into a chat where it did not belong.
Both results come from the same feature. If you want to use it deliberately instead of being surprised by it, you need to know what Claude saves, where it keeps it, and which switch controls what. The Complete Beginner's Guide to Claude covers the basics of working with Claude. This article is only about what carries over between conversations.
Two different things: chat search and memory
Anthropic's help pages describe two separate mechanisms, and people mix them up.
Chat search lets Claude look through your earlier chats when a question calls for it. It is available on the paid plans (Pro, Max, Team and Enterprise) on the web, in Claude Desktop and in the mobile apps. When Claude uses it, the search shows up in the current chat as a tool call, so you can see that it happened. It covers chats outside projects and the conversations inside each individual project, and a search never crosses a project boundary. You can switch it off under Settings, then "Your Claude," with the "Search and reference chats" toggle.
Memory is different. Instead of searching old conversations, Claude builds up a set of short topics from your chats, saved while the conversation is happening rather than summarized afterward. You can also ask directly: "remember that I prefer metric units" saves it on the spot. According to Anthropic's help page, memory is on by default for Free, Pro and Max users on the web, in Desktop and on mobile. On Team and Enterprise plans the organization's owners decide whether memory is available, and it stays off for each member until that member turns it on.
Chat search
- Looks back through old conversations when a question needs them
- Shows up in the chat as a visible tool call
- Paid plans only
- Never searches incognito chats or crosses into another project
Memory
- Keeps short topics about you and your work
- Saved during chats, or when you say 'remember this'
- On by default for Free, Pro and Max
- Can be viewed, edited, paused and deleted
What memory tends to hold, and what it never keeps
Anthropic says memory typically covers your role and professional context, the people and places that come up in your work, your communication and technical preferences, and details of ongoing projects. That is useful background: it is the difference between explaining your job in every chat and not having to.
Two limits are worth knowing before you share anything personal.
- Sensitive topics are left out by default. Health, religion, politics and gender identity are not saved unless you turn on "Include sensitive topics in memory."
- Some information is never saved. Anthropic lists government ID numbers, criminal history, financial account numbers and immigration status.
Not a vault
Those exclusions describe what memory does, not a promise about everything you type. What you paste into a chat is still part of that chat. Read the privacy article in this series before you paste anything you would not email to a colleague.
Projects keep their own memory
If you use Claude for more than one body of work, this is the part that matters most. Each project has its own memory. Inside a project, Claude draws on that project's memory plus the account-level memory from your chats outside projects. It does not use the memory of your other projects.
Take a freelance designer with two clients, a dental practice and a craft brewery. Notes about the practice's cautious tone stay inside the practice's project, and the brewery's playful voice stays in the brewery's. Neither leaks into the other, which is the point of having two projects.
There is one wrinkle. New memories created in a project are saved to that project, but with memory on, Claude can still reference them in chats outside projects. If a client's details should stay strictly inside the project, open the project's page and turn off "Use account memory." That separates the project from your account-level memory so nothing flows in from outside.
If you have not set up a project yet, Setting Up a Claude Project You'll Actually Keep Using walks through it.
Three controls, and what each really does
Claude gives you more than an on and off switch. The differences matter.
- 1
Pause: keep what exists, stop adding
Under Settings, then "Your Claude," the "Memory from chats" toggle pauses memory. Existing memories stay but Claude stops using them and stops creating new ones. Conversations that happen during the pause are not added later, so a pause is a real gap, not a delay.
- 2
Reset: delete everything, permanently
Resetting removes all memories, project memories included, and it cannot be undone. Use it when you want a true clean slate, not when you only want to fix one wrong detail.
- 3
Off for one chat: decide before you start
Open a new chat, click the "+" button and turn off "Memory" before you send the first message. The setting cannot be changed after that first message, and the chat still stays in your history and remains searchable from other chats.
For something you want to leave no trace at all, there is incognito. You start one with the ghost icon in the upper right of a new chat outside a project. Incognito chats are not saved to memory or to your chat history, and Claude does not search them later. They are available on every plan, including free accounts.
Reading, editing and deleting what Claude kept
Memory is not a black box. Go to Settings, then "Your Claude," and select "Manage" next to "Your memory." The memories appear as topics. Open one to read it, use the edit icon to change it, or delete it. You can also ask in a chat: "forget that I work in healthcare," or "update my memory: I changed teams in September."
Deleting a chat does not delete its memories
Memories created from a conversation stay after you delete the conversation. If something sensitive came up in a chat, delete the chat and then delete the matching memory topic separately.
A ten-minute memory audit
Do this once, then again every few months.
- Open "Manage" next to "Your memory" and read every topic. Delete the ones that are out of date: an old job title, a project that ended.
- Look for anything that mixes two parts of your life. Work notes sitting next to personal ones are a sign you want a project, or incognito, for one of them.
- Open a new chat and ask: "What do you remember about how I like my writing edited?" The answer shows you what Claude would apply without being told.
- If an answer surprises you, edit the topic, or say plainly in the chat what is wrong. Both work.
- For any project that holds a client's or an employer's information, check the "Use account memory" toggle on the project's page and decide whether you want it on.
Tip
A memory is only as good as the last time you read it. A wrong detail that sits there for months shapes every draft quietly, and it is the hardest kind of mistake to notice, because the output still looks fluent.
When memory helps and when it gets in the way
Memory pays off for stable facts: your role, your audience, the format you always want. It gets in the way when the context is temporary or sensitive: a one-off task for someone else, a draft that needs a different voice, a conversation you would rather keep separate. In those cases the cleanest answer is not a clever prompt but one of the three controls above, chosen before you start.
If you work across several assistants, remember that each one handles this differently. What ChatGPT's Memory Actually Remembers and What Grok's Memory Actually Remembers cover the same question for those tools.
Official sources
Checked on October 24, 2026. Features, plans and names change often, so the vendor's own pages are the final word.