Why does Claude forget my writing style?

Claude forgets your writing style because tone lives in three weak places: short Project instructions, a global custom Styles dropdown that applies the same voice to every conversation, and chat-level corrections that vanish at the end of the session. None of them store a true brand voice, and none of them follow you to Gemini, ChatGPT, or Grok. You need a memory layer that holds the voice and re-injects it on every turn.

Three things conspire to erase the voice you train.

1. Custom Styles are account-wide, not project-aware. Claude's Styles dropdown lets you save a style (Formal, Concise, or a custom one you build from writing samples). The catch is that one Style applies to whatever conversation it is selected on. There is no way to bind a Style to a brand, a client, or a campaign and have Claude automatically pick the right one when you switch contexts.

2. Project instructions are short by design. You can paste tone guidance into a Project, but the field is meant for briefs, not full voice guidelines. A 2,000-word brand voice document gets truncated in the system prompt, and Claude leans on the first few sentences while ignoring the rest.

3. Voice corrections do not persist between chats. When you tell Claude "stop using em-dashes" mid-conversation, the rule lives only inside that chat. Open a new chat in the same Project and the em-dashes return. Auto memory captures some preferences in Claude Code, but it does not run in claude.ai where most writing actually happens.

The combined effect: Claude can do your voice for a few turns, then drifts back to its default cadence the moment a new conversation opens.

Voice drift looks like a small problem until you see how it stacks up across a week of work:

• Brand inconsistency across drafts. A landing page, an email, and a tweet all sound subtly different because each chat started from defaults.
• Edit time balloons. You stop trusting first drafts and start rewriting the opening of every reply to strip clichés Claude reintroduced.
• Team writers diverge. Two writers using the same Claude Project still produce off-voice copy because each of them taught Claude differently in their own chats.

Anthropic offers two features that brush against the problem.

Custom Styles are the most visible answer. You upload sample writing or paste instructions, save the style, and apply it from a dropdown in any conversation. This works well for one default voice. It breaks down when you write for multiple brands, because you have to remember to switch the dropdown manually, and there is no per-Project default. Anthropic's official styles guide walks through the setup.

Project instructions can carry tone rules into every chat inside a Project. Useful if your work is one-brand, one-Project. Limited because the instructions field is short, instructions do not version, and there is no way for a teammate to read or improve your voice rules without copy-paste.

These cover light personalization. They do not hold a brand voice the way an agency style guide holds one.

Writing style is rarely about one rule. It is dozens of micro-rules: sentence length, idiom preferences, banned words, banned phrasings, preferred metaphors, audience reading level, regional spelling, and a handful of voice-defining tics. That kind of nuance does not fit in a styles dropdown or a 500-character instructions field.

It also does not stay inside Claude. You draft in Claude, polish in ChatGPT, and headline-test in Gemini. Every hop resets the voice.

MemoryLake holds your brand voice in one place, then surfaces it to Claude every turn.

• Voice as a living document. Store your full style guide, banned-word list, and reference drafts in a MemoryLake Project. Claude pulls only the relevant rules per request, so the voice stays sharp without bloating the system prompt.
• Multi-brand voice switching. Keep one Project per brand. When you switch brands, MemoryLake swaps the voice, so the next chat opens in the right tone without you touching a dropdown.
• Same voice in Gemini, ChatGPT, Grok. The voice you trained for Claude works across every other AI in the workflow. Drafts stop sounding like a committee of bots.

MemoryLake scored 94.03% on the LoCoMo long-context benchmark — the top published result as of 2026 — with millisecond retrieval and AES-256 end-to-end encryption.

1. Create a Project for the voice. Sign in to MemoryLake, open Project Management, click Create Project, and call it something like "Acme brand voice". Upload your style guide, banned-words list, and three or four exemplar pieces through the Document Drive, then add the highest-value tone rules as short notes in the Memories tab.
2. Generate an MCP Server endpoint. Open the MCP Servers tab, click Add MCP Server, name it "Claude voice", and click Generate. MemoryLake returns an API key ID, secret, and endpoint URL. Copy the secret right away — it appears only once.
3. Connect Claude. In Claude Desktop, add the MCP entry to claude_desktop_config.json with your endpoint and Bearer token, then restart. For claude.ai in the browser, drop a one-line system prompt into the Project instructions field that points at your MemoryLake project ID, and the REST API pulls the voice in on every reply.