Style Memory

The reason output sounds like your agency, not like a generic LLM.

The 30-second version

Style Memory is a 7-field JSON profile stored on each custom agent: tone descriptors, banned phrases, required sections, voice examples, structural pattern, reading level, and formality. You upload 1–5 samples of your past work, AtelyaOS extracts the profile, and every Composer run respects it. Unlike a ChatGPT custom GPT, the profile is structured, editable, and persistent.

Why this matters

LLMs default to a flat, helpful, slightly sycophantic voice. If your agency has a distinctive tone — minimalist B2B prose, punchy DTC voice, formal RFP-grade English — that default voice is the wrong starting point. Re-prompting "in the style of past work" every time wastes tokens and produces inconsistent output.

Style Memory turns "the way we write" into a small, inspectable artefact. You can read it, edit it, copy it between agents, and audit what the Composer is actually being told.

How it works

The 7-field profile

Every Style Memory profile fits this shape:

{
  tone_descriptors: string[]      // up to 20 short adjectives
  banned_phrases:   string[]      // up to 50 do-not-write items
  required_sections: string[]     // labels that must appear
  voice_examples:   string[]      // rule-with-evidence lines
  structural_pattern: string      // one-line flow summary
  reading_level: 'executive' | 'general' | '8th_grade' | 'postgrad'
  formality: 1 | 2 | 3 | 4 | 5    // 1 = very casual, 5 = highly formal
}

That schema is the public contract — every Composer prompt receives those fields verbatim.

Where it lives

Style Memory is per agent, stored on bots.style_rules. A single workspace can have many agents, each with its own profile. There is also an optional per-workroom style override for one-off cases (e.g. "this proposal needs to feel more formal than the agent's default").

[SCREENSHOT: agent builder Style tab with extracted rules]

How an agent learns

From the agent's Style tab, upload 1–5 samples of past deliverables — proposals, recaps, kits — that represent the voice you want.
Click Extract. AtelyaOS sends the samples through an extraction LLM call and produces a draft profile.
Review and edit. Add tone descriptors, prune banned phrases, set the formality dial.
Save. The next workroom that uses this agent will respect the profile.

Style Memory does not auto-update from new deliverables. It changes only when you re-extract or hand-edit. That's intentional — implicit drift is the failure mode you came here to avoid.

How it differs from ChatGPT

Persistent across sessions. ChatGPT custom GPTs forget when you start a new chat. Style Memory is stored in your workspace.
Inspectable. You can read every line the Composer gets. ChatGPT hides its internals.
Editable. Bad output? Open the profile, add a banned phrase or pin a structural pattern, and the next run respects it.
Per-agent. One workspace, many voices. ChatGPT can't keep five distinct voices straight.

Style Memory and tiers

Custom agents (and therefore Style Memory profiles) are gated by the per-workspace agent cap of your plan — see custom agents. Pro and above let you keep an unlimited number of agents, each with its own profile.

Common pitfalls

Uploading mixed-voice samples. If three samples are formal proposals and two are casual social posts, the extraction will average them and produce a muddy profile. Upload samples of one voice at a time.
Treating tone descriptors as marketing taglines. "Bold" and "innovative" are useless. "Direct, second-person, no rhetorical questions, sentence-medium length" is useful.
Skipping banned phrases. This is the highest-leverage field. List the words that scream "AI wrote this" — delve, unleash, seamless, etc. — and the output gets dramatically better.
Editing then forgetting to save. Extracted rules are a draft until you hit save.

What's next

Custom agents — the agent that holds the profile.
Custom agents per client — the recommended pattern.
Briefs and plans — what else feeds the Composer.

Style Memory

On this page