Every Ai Writes Like You
Most writing an Ai hands back sounds like the model, not like the person who asked for it. CogmemAi Voice fixes that where it holds for every tool at once: the memory layer. One command measures writing you wrote yourself, turns the habits into rules, and stores them with your memory. From the next session on, Claude Code, Cursor, claude.ai, your companion and your support assistant all write in your voice. A second command checks any draft against it before you send.
npx cogmemai-mcp voice learn notes.md emails.txt paper.docx
What One Command Produces
The measurements are computed locally. The voice document is written from them and from a few short excerpts, in checkable terms, so a reviewer can hold a draft against it.
A voice profile, as a rule
One voice_profile rule that names the voice and describes it the way an editor would: rhythm and how much it varies, register, point of view, how sentences open and close, the marks you use and the marks you never use, vocabulary level, signature words, formatting habits. It loads first in every session and tells the Ai to write in that voice.
Eight to twelve voice rules
Each one concrete enough to check in a draft and tied to a measurement: "NEVER use a semicolon; the writer used none in 4,200 words." "ALWAYS use contractions; the writer runs about 60 per 1,000 words." "ALWAYS open about one sentence in ten with a short transition." Edit, delete or add rules afterward like any memory.
The numbers, kept
A voice_metrics memory holds every measurement, so the check can compare a draft with your writing on another machine or inside the hosted endpoint. Learning again replaces the earlier set, moved to the trash rather than deleted. Global by default; --project keeps a brand voice apart from your own.
Two Commands, Then Nothing to Remember
1. Learn
npx cogmemai-mcp voice learn notes.md emails.txt paper.docx --name "Scott"
# 4,210 words in 3 samples; 212 sentences
# "plain direct engineer voice": 11 rules
# memories: 13 storedWriting you wrote yourself, 300 words at least and 1,500 or more for a good profile. --dry-run prints the measurements and the excerpts and sends nothing. --project learns a brand voice for this project only.
2. Write, anywhere
# In Claude Code, Cursor, claude.ai or any connected Ai:
"Draft the reply to the vendor."
# The voice rules are already loaded. It writes like you.Rules load in every session of every Ai connected to your memory. The profile rule asks the Ai to run voice_check on any draft longer than a paragraph before handing it back.
3. Check before you send
npx cogmemai-mcp voice check draft.md
# Drift check: 35/100
# high: uses 5 semicolons; the writer uses none
# medium: sentences average 26 words; the writer averages 10
# Review: coverage 0, 5 violations
# FAILEDDrift findings in plain words, then the review's coverage and violations against the voice document. Passes at 80 with no violations; exit code 1 otherwise, so it gates a script or a send.
No account yet? Get a free API key, run npx cogmemai-mcp setup, then the commands above.
What It Measures
Everything below is computed on your machine from your own writing, and again from each draft at check time. The same numbers drive the rules and the drift findings.
- Rhythm Mean, spread and median sentence length; the share of short sentences (8 words or fewer) and long ones (30 or more); sentences per paragraph.
- Openers and closers The first word of every sentence and how often a transition (However, Also, So, For example) opens one; how often a sentence ends as a question or an exclamation.
- Punctuation per 1,000 words Semicolons, colons, em and en dashes, parentheses, exclamation and question marks, quotation marks, ellipses. A mark at zero becomes a NEVER rule.
- Person and register Contractions, first person, second person, numbers, adverbs ending in -ly, words in capitals.
- Vocabulary Average word length, the share of long words, type-token ratio, and signature words: the frequent words that are not function words.
- Formatting Bullets and numbered items as a share of lines, headings, bold spans.
The drift check flags, in order of weight: a mark you never use appearing at all; sentence length far from yours; contractions or first person flipped; vocabulary heavier or lighter; transitions, paragraphs and bullets out of habit; none of your frequent words present.
What It Is Not
Not a detector bypass
Voice is never measured against an Ai detector and is not built to pass one. It exists so the email sounds like you, the post sounds like your company, and the paper reads the way you write. The output is still a draft you read before you send it.
Not a style setting in one tool
Custom instructions live in one product and are lost when you switch. Voice lives in your memory with your rules and your intent, so the same habits hold on Claude, GPT and Qwen, in an editor, a chat window or a hosted endpoint, and they stay when you change any of them.
Questions
What does cogmemai-mcp voice learn actually do?
It reads writing you wrote yourself (.txt, .md, .html, .docx or .pdf) and measures it on your machine: sentence length and how much it varies, paragraph length, the first word of every sentence and how often a transition opens one, every punctuation mark per thousand words, contractions, first and second person, numbers, adverbs, average word length, signature words that are not function words, and bullets, headings and bold. The measurements and a few short excerpts go to the MSGR gateway once, and the reply is stored as rule memories: a voice_profile rule that describes the voice in checkable terms, eight to twelve voice rules such as "NEVER use a semicolon; the writer used none in 4,200 words", and a voice_metrics memory with the numbers. About a minute.
How does the voice reach my other tools?
Rules load first in every session of every Ai connected to your CogmemAi memory: Claude Code, Cursor, claude.ai, a companion, a support assistant on the MSGR gateway. Nothing else to configure. The profile rule tells the Ai to write in that voice and to check longer drafts with voice_check before handing them back.
What does voice check compare?
It measures the draft the same way and reports drift in plain words: "uses 5 semicolons; the writer uses none", "sentences average 26 words; the writer averages 10", "no contractions; the writer uses about 60 per 1,000 words", "none of the writer's frequent words appear". Then the review judges the draft against the voice document and returns coverage and violations. A draft passes at the threshold (80 by default) on both with no violations, and the exit code is 1 when it fails, so it works in a script or as a gate before anything is sent.
Can I keep a brand voice separate from my own?
Yes. The default is global, because a person's voice applies everywhere. Add --project and the profile is learned for the current project only, so a client site, a product blog or a company account gets its own voice and never mixes with yours. The check takes the same flag.
Does my writing leave my machine?
The measurements are computed locally. A few short excerpts (about 7,000 characters in total) are sent with them to the MSGR gateway so the voice document can quote real habits; --dry-run prints exactly what would be sent and sends nothing. The resulting rules are stored in your CogmemAi memory, encrypted at rest like every other memory.
Is this a way to beat Ai detectors?
No, and it is never measured against one. It is for writing in your own voice, or your brand's, on your own work: the email that should sound like you, the post that should sound like your company, the paper that should read the way you write. The output is still a draft you read before you send it.
What does it cost?
The learn command and the local drift check run on every CogmemAi plan. The judged review inside voice check runs on the paid tiers. Included in Pro and Teams.
Start With Your Own Writing
A minute from your own words to rules every Ai follows. Included in CogmemAi Pro and Teams.