How to build a Brand Memory Layer: a setup guide
Most brand guides don't transfer well to AI tools. They're written for humans — full of judgment calls, subjective descriptors, and examples that assume context the reader brings. An AI model starting a new session doesn't bring that context. It starts with everything it was trained on and nothing your brand specifically locked.
A Brand Memory Layer fixes that gap. It's the encoded set of inputs — specific enough for a model to act on, structured enough to survive across sessions — that your AI draws from before it writes a word.
Here's how to build one.
What a Brand Memory Layer actually contains
The common mistake is treating this like a compressed version of the brand guide. That's still written for human readers, which means it's still full of descriptive language ("bold but approachable") that a model can't operationalize.
A Brand Memory Layer contains four types of content, each serving a different purpose:
| Type | What it locks | Why it's there |
|---|---|---|
| Brand facts | What the product actually does, who it's for, what the pricing model is | Prevents the model from hallucinating features or overpromising |
| Voice rules | Banned words, required patterns, tone directives with examples | Prevents drift back toward generic AI defaults |
| Positioning decisions | The one claim you own, how you frame against alternatives, what you don't say | Keeps every output aligned to the strategic angle |
| Past-work examples | 3–5 pieces of approved output with brief notes on why they worked | Shows the model what "good" looks like in practice |
All four are necessary. Brand facts alone produce accurate but tonally flat output. Voice rules alone produce on-voice output that misrepresents the product. Positioning without examples produces strategic-sounding content that doesn't connect.
The five things to lock first
Start here before trying to build the full layer. These five inputs produce the most immediate improvement in output quality:
1. Your positioning sentence. One sentence that states what you do, who it's for, and what makes it different — written the way you'd say it on a sales call, not the way it sounds in a brand deck. This is the anchor every draft pulls against.
2. Your three banned words or phrases. Every brand has language it's instinctively wrong about. List the three most common offenders from recent AI output — the words that signal the model has lost your voice.
3. Your tone descriptors with examples. Not "warm but professional." Something like: "Write like a smart peer explaining something true — no filler affirmations, no manufactured enthusiasm. Here's a paragraph that hits it: [example]."
4. Your audience's actual job. Not a persona. The specific work your reader is trying to get done when they come to your content. The model performs better when it has a concrete reader doing a concrete task.
5. One thing your product doesn't do. This is the guardrail most teams skip. State one thing the product doesn't claim to do — the overreach that would damage trust if the model invented it. That constraint is worth more than a dozen positive attributes.
Once these five are in place, run a test session and compare the output against a session without them. The difference is usually visible in the first paragraph.
How to format it for AI use
Brand memory that lives in a PDF or a Notion doc doesn't help. It has to be structured for the tool reading it.
The most reliable format: short declarative statements organized by type, not by section. Avoid prose explanations — the model doesn't need to understand why a rule exists, just what to do.
A working example:
VOICE RULES
- Never use: leverage, seamless, transformative, best-in-class
- Active voice only. No passive constructions like "content is generated."
- Short paragraphs. End most paragraphs with a single short sentence.
POSITIONING
- We are an AI marketing platform that keeps brand context across sessions — not a writing tool that resets every chat.
- We do not claim to replace strategists or guarantee results.
- When comparing to alternatives, frame on context persistence and process discipline, not feature counts.
AUDIENCE
- Marketing operator or agency lead who already uses AI, is getting mediocre output, and suspects the problem is their workflow, not the model.
Three sections, each with clear directives. This is the format that translates to consistent output — not because it's comprehensive, but because it's specific enough for the model to act on without interpretation.
Building from existing brand materials
Most teams already have something — a style guide, a messaging framework, approved copy. The build process is extraction, not creation.
Work through each document and pull the statements that are specific enough to act on. Filter aggressively. "Our tone is human and relatable" is too vague. "We never start a sentence with 'We believe' or 'We're passionate about'" is specific. Only pull what passes that test.
For past-work examples: select pieces where someone on the team said "this is what we're going for." One or two per content type (blog, email, social) is enough. Include a one-sentence note on why it worked — "this gets the voice right because it doesn't oversell the feature."
The extraction pass takes a few hours the first time. Most of that time is deciding what to leave out.
How to know when it's working
Output quality isn't the only signal. The sharper test: does a new AI session produce an on-brand first paragraph without you re-explaining the brand in the prompt?
If the answer is yes, the layer is doing its job. If you're still leading every session with "remember, we're a B2B platform and our tone is…" — something isn't loaded, or it's structured in a way the model isn't drawing from.
Three signs the layer needs a revision:
- The banned words keep appearing. The list is incomplete or too abstract.
- Tone is right but facts are wrong. Brand facts aren't specific enough.
- Output sounds on-brand for the category, not the company. Positioning isn't locked tightly enough — the model is falling back on industry defaults.
Copper Sun locks these inputs at the session level so they surface automatically in every project — without pasting context into each prompt. That's the difference between a Brand Memory Layer that works once and one that holds across a team's work over time.
Frequently Asked Questions
How often should I update the Brand Memory Layer?
When something locked has changed: a positioning shift, a new product feature, a banned phrase that the team adopted. Review at least quarterly and after any major brand or product update. The layer becomes stale when the output starts diverging from what the team expects — that divergence is usually the first signal.
How much context is too much?
When the additions are descriptive rather than directive. "Our tone is inspired by the best B2B content shops" is descriptive — it tells the model what to aspire to without telling it what to do. Every item in the layer should generate a specific behavior, not an aspiration.
Can I use the same Brand Memory Layer across different content types?
The core facts, voice rules, and positioning should travel everywhere. Past-work examples benefit from one or two per content type — a blog post example doesn't tell the model what a good email looks like. Keep the shared base tight and add format-specific examples where the work requires them.
What's the difference between a Brand Memory Layer and a system prompt?
A system prompt sets the model's behavior for a session. A Brand Memory Layer is the content that goes into that prompt — or into a persistent memory system — so the model has brand-specific knowledge to draw from. The layer is the substance; the system prompt is the delivery mechanism.