They are instructions that stay in effect across many conversations, so you do not have to type them again each time.

When you chat with Claude, the instructions you write in a message normally apply only to that conversation. If you say “keep your answers short” in one chat and start a new chat tomorrow, Claude begins fresh. A persistent instruction is different. You save it in a special place, and Claude reads it automatically at the start of your conversations. It might be a note about who you are, a rule such as “always use simple English”, or background about a project you work on every week.

Claude offers more than one way to do this, and each way has its own purpose and reach. Some apply to everything you do, some apply to one project only, some control only the style of writing, and some build up on their own from your chats. Choosing the right one for each instruction is the heart of this topic.

flowchart TD
    A["You, setting things up once"] --> B["Profile preferences: apply everywhere"]
    A --> C["Project instructions: apply inside one project"]
    A --> D["Styles: shape how replies are written"]
    A --> E["Memory: remembers things from your chats"]
    B --> F["Every new chat starts with the right context"]
    C --> F
    D --> F
    E --> F
        

A few simple words appear again and again. Here they are in plain English.

Without persistent instructions, you pay a small repeating tax in every new chat.

Picture how many times you might type the same sentences. “I am a teacher. Keep explanations simple. Use Indian rupees in examples. Do not use bullet points unless I ask.” Typing this once is easy. Typing it fifty times in a month is a waste, and sooner or later you will forget a line and wonder why the answer feels off.

Each layer has its own reach, and knowing the reach tells you where to put each instruction.

You can use several layers together. They are designed to complement each other. A sensible mental picture is a set of transparent sheets laid one on top of another: your general preferences at the bottom, your project guidance above that, and your chosen style shaping the final look of the reply.

flowchart LR
    A["Your message"] --> B["Claude reads your profile preferences"]
    B --> C["Adds project instructions if you are in a project"]
    C --> D["Uses memory where it helps"]
    D --> E["Writes the reply in your chosen style"]
        

This is the simplest and most widely useful place to start.

Profile preferences are account-wide settings that help Claude understand your general preferences. You write them once in a text box in your settings, and Claude considers them in your conversations. You will usually find the box in your Settings under the Profile area, where a prompt asks what preferences Claude should consider in responses. Menu names and positions can change, so check the Help Center if yours looks different.

The best preferences are short, specific, and useful in almost any conversation.

When instructions belong to one body of work, give them a home inside a project.

A project is a dedicated workspace for related chats and materials. Inside it you can add project instructions, which guide how Claude behaves in that project’s conversations, and project knowledge, which is a set of documents Claude can refer to. The two serve different purposes. Knowledge is what Claude knows. Instructions are how it behaves.

Because instructions live inside the project, they stay out of your other chats. This keeps your general account clean. A writing project can have rules about tone and structure, while a separate planning project can have rules about checklists and deadlines, and neither interferes with the other.

The two are partners that work at different levels.

When both apply, they work together. Your profile gives the general background and the project adds the specific guidance for the task at hand. If the two seem to point in different directions, it is wise to write the project instruction clearly enough to state the exception, so there is no doubt about which should win in that project.

A style keeps the feel of replies steady without changing what you ask.

Styles customize how Claude communicates with you. Unlike profile preferences and project instructions, which provide context and guidance, styles focus on how Claude formats and delivers its responses. You might choose a style that is concise, one that is more explanatory for learning, or one that is more formal for business writing. You can also create a custom style, for example by describing the voice you want.

Memory lets Claude carry useful details from one chat to the next, without you writing them down.

Claude can generate memory from your chats. When this is turned on, Claude saves useful details as you talk, for example your role, an ongoing project, or how you like things done, and can use them in later conversations. Unlike preferences, which you write yourself, memory grows from your conversations. Anthropic describes it as turning Claude from a stateless chat window into a collaborator that builds understanding over time.

You are in control. In Settings you can switch memory on or off, see what Claude remembers, and edit or delete individual entries. You can also simply tell Claude what you would like it to remember, and it updates its memory during the conversation. Projects have their own separate memory space, so context inside one project stays focused and does not leak into your other projects.

sequenceDiagram
    participant You
    participant Claude
    participant Memory
    You->>Claude: Chat about your work and preferences
    Claude->>Memory: Saves useful details
    You->>Claude: Start a new chat later
    Memory-->>Claude: Brings back relevant details
    Claude-->>You: Answers with that background in mind
        

Persistent does not mean permanent. You can always see, change, and limit what is kept.

With memory on, you can manage things with plain conversation.

You do not always need to open settings. If memory is available, you can say, in a normal sentence, what you would like Claude to remember, and it will update its memory. You can also correct something it got wrong, or ask it to forget something, and the change applies to your next conversations.

Match each instruction to the layer that fits its reach.

flowchart TD
    A["I have an instruction"] --> B{"Should it apply to every chat?"}
    B -- "Yes" --> C["Profile preferences"]
    B -- "No" --> D{"Does it belong to one project?"}
    D -- "Yes" --> E["Project instructions"]
    D -- "No" --> F{"Is it only about writing style?"}
    F -- "Yes" --> G["A style"]
    F -- "No" --> H["Say it in this chat, or let memory pick it up"]
        

The same idea appears in other Claude products, under different names.

Persistent instructions are read again and again, so it pays to write them carefully once.

Let us follow an imagined writer named Kavya as she sets up all the layers.

Kavya now opens a new chat and gets answers that already fit her work, her client, and her voice, with no repeated explaining. She also reviews her memory list now and then, and tidies her instructions when a client changes.

Layers can sometimes disagree, and some limits always remain.

Because several layers can apply at once, they may occasionally seem to contradict each other. Perhaps your preferences say “use bullet points” while a project says “write in flowing paragraphs”. In general, more specific guidance is meant to refine the general guidance, and what you say directly in a chat about the task at hand usually takes priority for that chat. The best fix, however, is to avoid the conflict by writing clear instructions and stating exceptions directly.

It is also important to know that persistent instructions work within Claude’s built-in safety behavior. They can shape tone, format, and focus. They cannot be used to make Claude ignore its guidelines or act against its principles, and instructions that try to do so are not followed. Likewise, if an organization has set controls, those may take precedence over your personal settings.

A few common habits cause most of the trouble.

Persistent instructions save effort, with a small cost of upkeep.

Quick answers to the questions beginners ask most often.

It means the instruction stays in effect across many conversations instead of ending when one chat closes, so you do not have to repeat it.

In your Settings, in the Profile area, where a text box asks what preferences Claude should consider in responses. Menu names can change, so check the Help Center if yours looks different.

Preferences apply broadly across your account. Project instructions apply only inside one project and are best for guidance tied to a specific body of work.

Styles shape how Claude writes and formats its replies, such as concise, explanatory, or formal, rather than what Claude knows or what task it does.

You write preferences yourself. Memory is built from your chats, with Claude saving useful details over time. Both are visible and controllable by you.

Yes. The memory area in Settings lists what Claude remembers, and you can edit or delete entries, pause memory, or turn it off.

Yes. You can use an incognito chat, or turn memory off for a single chat before you send the first message.

No. Some kinds of information, such as government ID numbers and financial account numbers, are never saved to memory, even if you ask.

One-time needs, such as a special tone for today’s task. Saving them permanently would affect later chats where you do not want them.

No. Persistent instructions work within Claude’s built-in guidelines. They can shape tone, format, and focus, but they cannot override safety behavior.

Several layers may apply, instructions may be unclear or outdated, or a chat may be very long. Check each layer, simplify the wording, and remind Claude in the chat if needed.

In the Claude Help Center, in the articles on personalization features and on how memory works, which are kept up to date as the product changes.

Persistent instructions let Claude keep your wishes across many conversations, so you stop repeating yourself. Claude offers several layers for this. Profile preferences hold general notes about you that apply across your account. Project instructions hold guidance for one body of work, alongside project knowledge. Styles shape how replies are written. Memory lets Claude carry useful details forward from your chats, and you can view, edit, pause, or turn it off, while certain sensitive information is never saved. The skill is matching each instruction to the right layer, writing it clearly and briefly, and reviewing it as your life and work change. Saved instructions guide Claude within its safety behavior, and they do not replace your own checking of important answers.