Phase 3 · Claude · Level 2 · Practitioner
Styles and memory: shaping tone and what Claude retains
By the end, you'll be able to…
- Apply a built-in or custom style so Claude writes in a voice you choose
- Use memory so Claude carries your role and preferences across chats, and control what it keeps
- Decide what belongs in a style, in memory, in project instructions, or in the prompt
Why it matters
Two features quietly personalise Claude: styles shape how it writes, and memory shapes what it knows about you across conversations. Used well, you stop repeating 'I'm in operations, keep it plain and short' every time. Used carelessly, Claude drifts in a voice you didn't intend or remembers something you'd rather it forgot. This lesson shows how to steer both. Just as important, it shows how to check and switch them off.
Styles: how Claude writes
A style controls Claude's voice (its tone, structure and level of formality) separately from what you ask. You set the style once and it shapes every reply until you change it, so you're not re-typing "keep it concise and plain" at the top of each prompt.
Claude ships with a handful of built-in presets you reach from the plus menu in a chat, under Use style. The usual set is Normal, Concise, Explanatory, Formal and Learning. The names describe them fairly well: Concise trims the padding, Explanatory adds more teaching and detail, Formal tightens the register, Learning walks you through things. For a lot of work, simply switching to Concise or Formal is enough.
Where it gets useful is a custom style. From the same menu, Create & edit styles lets you build your own in one of two ways: paste a writing sample and let Claude infer the voice from it, or describe the style you want in words. A sample is the stronger option when you have one. Claude copies patterns from real examples better than it follows adjectives.
Create a style called 'Fernway internal'. Voice: plain UK English, warm but efficient, for busy non-technical colleagues. Short sentences and scannable structure with headings where it helps. Avoid corporate filler ('leverage', 'circle back', 'reach out'). Never over-promise; if something's uncertain, say so plainly.Why this works: A style is a standing instruction about voice, so naming the audience, the register and concrete do/don't rules gives Claude a target it can hit consistently across every chat, far more reliable than saying 'be professional' once in a prompt.
Once saved, a custom style is tied to your account, so it follows you across browsers and the mobile app. One honest caveat worth knowing: styles are gradually being folded into a broader personalisation feature Anthropic calls Skills, so the exact menu may shift. The underlying idea (a saved, reusable voice) is what's durable. Anthropic's personalization-features help article has the current path.
A style is per-account and about voice. That's different from project instructions, which apply only inside one Project and can set rules and voice for that strand of work. If a tone is for one project, put it in the project instructions; if it's how you always like Claude to write, make it a style.
Memory: what Claude retains about you
Memory lets Claude carry context between conversations. Rather than meeting you fresh every chat, it builds a running summary of the professional things worth remembering (your role, the projects you work on, how you like responses formatted, your working style) and draws on it next time. Ask for a report and it can already know you want plain English and a manager-friendly summary, because you told it that last week.
Memory fills in two ways. Automatically, as Claude notices recurring things you mention; and intentionally, when you just tell it: "Remember that I'm in operations at a mid-size UK firm and I always want plain English." The intentional route is the one to lean on, because you control exactly what goes in.
Add to your memory: I'm an operations manager at a mid-size UK company. I mostly summarise documents, draft internal emails and analyse spreadsheets. Default to plain UK English, short and scannable, and always name your source when you're working from a document I've given you.
Why this works: Explicit memory instructions put you in charge of your own profile: you decide what persists rather than leaving Claude to infer it. Being specific about role and preferences means future chats start already tuned, without re-explaining.
Crucially, memory is transparent and controllable. You can see and edit everything it holds under Settings → Capabilities, where there's a View and edit memory option, each remembered item shown as a short line you can change or delete. You can pause memory (it keeps what it has but stops adding) or reset it (delete the lot, which is permanent). And for a one-off conversation you don't want kept at all, there's an incognito chat (a ghost icon) that saves nothing to history or memory, ideal for a sensitive or throwaway task.
Two subtleties matter. First, memory is project-aware: a Project keeps its own separate memory space and summary, so your onboarding Project doesn't bleed into your general chats. Second, memory availability and the related chat-search feature can differ by plan. Memory itself is broadly available, but check the memory and chat-search help article for today's detail.
Where should a preference live?
Four homes, and choosing the right one saves a lot of friction:
- In the prompt: a one-off ("make this one formal"). Nothing to save.
- In a style: how you want Claude to write, across everything. Voice, tone, register.
- In memory: durable facts about you and your work: role, recurring tasks, standing preferences.
- In project instructions: rules and voice for one strand of work, kept out of everything else.
Roughly: style is voice-everywhere, memory is you-everywhere, project instructions are one-project, and the prompt is just-this-once.
Show me everything you currently have in memory about me and my work, as a plain list. Flag anything that looks out of date or that you inferred rather than being told directly, so I can correct it.
Why this works: Periodically asking Claude to recite its own memory surfaces anything stale or wrong before it quietly shapes answers. A simple habit that keeps personalisation from drifting somewhere you didn't intend.
Try it now
Common mistakes
- Letting memory fill itself and forgetting to look. Automatic memory is convenient but can pick up a stray detail or an outdated fact. Check the memory list occasionally and prune it, because it's shaping every answer silently.
- Putting project-specific rules in memory. A rule that only applies to one client or project doesn't belong in your global memory, where it will leak into unrelated chats. Use project instructions for that.
- Forgetting incognito exists. For a sensitive or one-off task you don't want retained, use an incognito chat rather than trusting yourself to delete the memory afterwards.
- Over-trusting a personalised answer because it sounds like 'your' Claude. A confident reply in your own house voice, from a Claude that 'knows you', feels more trustworthy than a stranger's, and the voice and the memory change the tone and framing, never the underlying accuracy. A wrong figure in your preferred style is still wrong. Personalisation makes Claude fit you; it does not make it more correct, so keep checking the facts exactly as before.
Keeping current
Styles and memory are both actively changing: styles are migrating toward the broader Skills feature, and memory's controls and reach keep expanding. The durable ideas (a saved voice; a controllable, transparent profile of you) will remain even as menus move. Check the personalization-features help article, the memory help article, and the Claude release notes. Accurate as of 13 July 2026.