Claude memory is a persistent memory feature that lets Claude retain context, preferences and project details between sessions. Anthropic introduced the feature in September 2025 for Team and Enterprise, extended it to Pro and Max in October 2025, and made memory available on the free plan in March 2026. In my own work with Claude, it runs every day as a layer on top of projects and standalone chats.
One distinction matters if you read further on the topic: Claude memory in this article is the consumer feature in Anthropic’s Claude, not the CLAUDE.md system that developers use within Claude Code. They are two different mechanisms under the same name.
What Claude memory does for your work
Claude memory retains professional context between separate sessions. Instead of having to explain your way of working, client preferences or ongoing projects again in every new chat, Claude stores what is relevant and uses it in later conversations. The feature is optional and can be switched off completely in the settings.
The feature works on two levels. General memory applies to all standalone chats outside projects. Project memory, in addition, is kept separate for each project: what you build up in your strategy project stays separate from client work in another project.
That separation stops information from spilling over between contexts where you do not want it to. That sets it apart from memory implementations that lump everything together.
Besides memory, Claude offers two related features for persistent context. Claude Projects bundle instructions, files and chats around one piece of work. Claude Cowork carries out tasks with access to files and folders, and since 16 September 2026 has been rolling out to Pro, Max and Ultra, available from every conversation. Memory is the layer that works at the level of Claude itself and holds on to context across all your interactions.
How to turn on Claude memory
You turn on Claude memory under Settings and then Your Claude, with the “Generate memory from chats” toggle. In the older view, memory is under Settings, Memory, and in the legacy view under Settings, Capabilities. On desktop, the toggle sits behind the menu at the top right; on mobile, it is in the same settings menu. As soon as you turn memory on, you get the option to have Claude create an initial summary based on your existing chat history, or to start from scratch.
Claude memory is on by default for Free, Pro, Max and Ultra. On Team and Enterprise, the organisation-wide setting is on by default, but memory stays off for each member until that member turns it on. Since the update of 25 August 2026, memory works in both chat and Cowork, and topics such as health or beliefs stay out of memory unless you turn on the setting for sensitive topics yourself.
With a new account, starting from scratch is the cleanest option, because you see exactly which information is stored and when. On an account with months or years of chat history, the initial synthesis produces usable context faster, though it is a one-off catch-up that you will want to review afterwards.
Which Claude model you use makes no difference to memory itself. The feature works on every model in every plan that supports memory. Which Claude model you choose does affect how much you get out of memory, because one model makes more use of that context than another.
What Claude’s memory does and does not store
Claude’s memory mainly stores professional context that is useful for recurring work: your role, the projects you work on, preferences for tone and format, recurring topics and work patterns. What it does not do is record entire conversations word for word. Memory consists of separate, categorised items, not a transcript.
Anthropic explicitly positions the feature for work context and does not store personal or sensitive topics by default. Conversations about health or beliefs, for example, only end up in memory if you turn on the “Include sensitive topics in memory” option yourself under Settings, Your Claude. For business users, that matters because it restricts the scope to work-related context, which limits the consequences of any data breach.
What is stored:
- Role, field and work context
- Ongoing projects and their status
- Preferences for style, tone and format
- Client or company names you mention explicitly
- Ways of working you apply repeatedly
What is not stored:
- Verbatim conversation content
- Sensitive topics such as health and beliefs, as long as you leave the sensitive topics option off
- Chats you have in incognito mode
- Memory from projects other than the one you are working in
How to view, edit or delete your Claude memory
You view your Claude memory under Settings, Your Claude. There you will find a complete overview of what has been stored, as readable text rather than an opaque internal representation. You can delete each entry individually, and you can wipe the entire memory in one go with a reset option.
Besides the settings, you can ask Claude directly in a conversation what has been stored (“what do you remember about me?”) or ask it to add or remove something specific (“forget that I am working on project X”). This conversational control is faster than clicking through the settings and fits well with a daily review routine.
If you log prompts and outputs for later audit, memory fits well with that habit. A weekly look at your memory items, which Claude has shown by category since 10 July 2026 rather than in a daily summary, prevents outdated assumptions from lingering and affecting your output.
A closed project should not still be influencing a new chat six months later. Regular clean-ups are not extra work; they are part of the discipline that makes memory a reliable tool rather than an invisible layer of context you have no control over.
Importing memory from ChatGPT or Gemini into Claude
You import memory from ChatGPT or Gemini into Claude in three steps. First, you ask the source system for a structured export of what it has stored about you. Then you copy that output, and finally you paste it into Claude under Settings, Your Claude, using the Start import button. You then review the items and add them with Add to memory. Import is experimental and works on all plans, on the web and in Claude Desktop. Anthropic has published a recommended export prompt that covers all relevant categories: instructions, preferences, work context and ongoing projects.
The import is not an exact copy. Each AI tool stores memory in its own format (compare how memory in ChatGPT works), and detail is lost in the transfer. What remains is usually enough to give you a head start, but expect a review round afterwards to correct assumptions that came across distorted.
Switching from ChatGPT to Claude usually involves a transition period in which you are setting up two tools at once, a familiar hurdle if you still go back and forth between them.
Compare Claude and ChatGPT on the type of work you do before you migrate your memory for good. A half-finished import that you later have to untangle costs more time than making one considered choice at the start.
Incognito chat as an alternative to Claude memory
Incognito chat in Claude is a session type that saves nothing to your conversation history and adds nothing to memory. You turn it on for each chat with the ghost icon in the interface, or with the shortcut Cmd+Shift+I on a Mac. The session disappears as soon as you close it and leaves no trace in the memory layer.
Incognito is meant for situations where you deliberately do not want anything stored: confidential brainstorms, one-off strategic discussions, sensitive client questions that do not belong in a shared system. It differs from turning memory off because it works chat by chat: your existing memory stays intact and is not used or updated in that one session.
If you regularly work with client data, incognito is often a better choice than switching memory off completely. You keep the productivity gains of memory for your own work patterns and use incognito when you need it. That is a workable middle ground between “remember everything” and “remember nothing”.
Claude memory within projects
Within projects, memory works with a separate memory space for each project. When you work within a project, Claude uses that project’s memory and your general memory alongside the project’s instructions and files. The memory of other projects stays out of view. If you switch off “Use account memory” on the project page, the project memory stays completely separate from your general memory.
That isolation is more than an organisational preference. If you work for several clients or products at the same time, project memory prevents context leaking from one file to another. Client A’s price list should not turn up in a conversation about client B, and that separation needs to be structural rather than dependent on your own alertness in every chat.
In my daily work with Claude, this is how I keep client contexts separate without having to explain manually in every new chat which project is active. Project memory takes care of that context switch. If you do not use projects yet, you are missing a layer of control that makes memory properly usable for professional work across multiple client files.
When Claude memory pays off and when it does not
Claude memory pays off when you consistently do the same type of work with the same preferences, clients or methods. Memory builds value through repetition: the more of your patterns Claude recognises, the less context you need to give in each session. For an SEO specialist who produces twenty content briefs a month, or a lawyer who prepares the same types of deed every week, that is a direct time saving.
The feature pays off less in three situations. With changing client contexts and no project isolation, memory quickly fills up with assumptions that no longer hold. For one-off or ad hoc tasks, memory saves little because there is nothing to build on. And for work where you want a clean slate each session, for example sensitive decision-making where you want no influence from earlier conversations, incognito works better than memory.
Clients do not accept inconsistent output, and that has taught me to build fixed ways of working rather than deciding task by task what is good enough. Memory fits that philosophy when it is used in a structured way: projects per client, regular review of what has been stored, and incognito where nothing should be stored.
If you switch memory on without structure and hope it will sort itself out, you end up with a tool that slowly becomes unreliable.
Privacy and data risks of using Claude memory for business
On privacy and data risks, three points matter for professionals in the EU. Memory stores information on Anthropic’s servers, not on your own machine. On Free, Pro and Max, you choose yourself whether Anthropic uses your chats for model training, a choice that has been mandatory since 8 October 2025; on Team and Enterprise, Anthropic does not train on your content by default. And as soon as client data ends up in the context, memory directly touches on professional secrecy and GDPR obligations; if you have signed NDAs, those apply just as much.
From my years at major banks, I bring the four-eyes principle into AI work, and with memory that principle comes back in a specific way. Before you turn on memory for business use, answer three questions.
First, which data categorically does not belong here? Think of clients’ personal data, financial details, medical information and legally privileged material. Second, what check applies to what does get stored? Do you review your memory weekly, or let it run unchecked? Third, who is liable if memory introduces an error into output that goes to a client? You are, not Anthropic.
Not every task has the same risk profile, and the intensity of your checks should move with it. An internal brainstorm calls for less care than advice that falls under professional responsibility. For the first category, memory with regular clean-ups is enough. For the second, incognito mode or a dedicated project with strict isolation is the safer choice.
Compliance thinking from the banking sector translates into specific questions you ask for every task, not into a one-off decision to switch memory on or off. For team environments, Enterprise admins can also disable memory across the organisation. So if you work within a larger team, check your own organisation’s policy before setting up memory for yourself.
If you want to build memory and the checking step that goes with it into your own work, you can learn that step by step in LearnLLM’s Claude Course.
Frequently asked questions about Claude memory
Memory has related features that work separately from it, such as chat search and memory for developers.
Can Claude search your earlier chats?
Yes, Claude can search your earlier chats with the Search and reference chats feature, which is switched on or off separately from memory. According to Anthropic, search is available on the paid plans and is on by default once the feature has been rolled out to your account. Within a project, search stays limited to that project, and incognito chats are never searched.
What is the difference between Claude memory and CLAUDE.md?
The difference between Claude memory and CLAUDE.md is that memory automatically retains context from your chats, whereas you write a CLAUDE.md file yourself for Claude Code. Such a file of instructions sits in a project’s folder or in your own user folder and is loaded at the start of every session. Claude Code also has an auto memory with notes that Claude keeps itself, stored locally on your own computer.
Is there a memory feature for the Claude API?
Yes, the Claude API has a separate memory tool that lets Claude store information between conversations and read it back from a folder of memory files. That tool works on the developer’s side: Claude requests a file operation and your application carries it out on storage you manage yourself. So you decide where and how the data is kept.
