AI drafts have a familiar problem. The first version is competent and completely generic, so you fix it: you cut the buzzwords, swap a word the brand never uses, flatten an exclamation point, tighten a sentence. Then the next draft arrives and you make the exact same edits again. The model never learned, because nothing was watching.
Voice memory is what watches. It notices the edits you make to AI drafts and stops making you repeat them.
It learns from the edits you already make
There is no separate setup, no style questionnaire, no prompt to maintain. When SurfacedBy drafts something and you edit it before it ships, that edit is the lesson. Across every channel where it drafts for you, content and community replies alike, it captures the difference between what it wrote and what you approved.
Cut “empower” and “leverage” enough times and it learns you do not use them. Keep swapping “users” for “teams” and it stops writing “users.” None of that requires you to articulate a style guide; you just keep editing the way you always do, and the drafts drift toward your voice.
A voice you can see and edit
This is not a black box that quietly drifts. Once a week, the captured edits are distilled into a short list of plain-language patterns you can read, and each one traces back to edits you actually made.
Because it is visible, it is correctable. If a pattern is wrong or too aggressive, you edit it or remove it, and the drafts adjust. And it is capped on purpose: the system keeps a tight number of patterns rather than accumulating every quirk forever, so your voice stays a clear set of rules instead of a sprawling, contradictory mess.
Per platform, because you do not write a reply like a landing page
The voice that works on a landing page is the wrong voice for a Reddit thread. So the patterns carry a channel: some are global, and some apply only where you write that way. The marketing register you keep on a content page can be exactly the thing you drop in a community reply, and voice memory keeps those straight instead of flattening everything into one tone. The same learned voice then feeds both the content drafts and the community replies, each in the right key.
One boundary worth being clear about: voice memory is about how you sound, not what is true. It learns tone, word choice, and rhythm; the facts about your brand live in your brand knowledge, separately. So teaching it your voice never risks teaching it a wrong claim, and you stay the editor of both.
A note on timing
SurfacedBy changes often. We build from a mix of customer requests, our own research, and our own daily use of these drafting tools, so the inspector and the channels here reflect how voice memory looked at launch. The distillation and the per-platform handling have gained depth since, and they will keep gaining it. It works quietly in the background as you write; the more you edit, the more it sounds like you.
