Every brand I talk to is trying to build a second brain.

The motivation is the same across all of them. They want autonomous AI agents that can actually do work on the brand’s behalf. Reply to a creator. Draft a campaign. Diagnose a CRO issue. Spot a recurring support pattern.

For any of that to work, the agent needs to know what the brand knows. Things like the decisions, conversations, history. The rates you’ve paid creators. The campaigns that worked. The reasons you walked away from things.

So brands are building a brain. Good instinct.

But most of them will end up doing it wrong.

We built Gentic Brain for the brands that want to do it right. gentic.co/brain

You’ve probably seen the posts.

Andrej Karpathy wrote a great gist about an LLM-maintained, compounding markdown artifact. Tobi Lütke posted on X about knowledge compounding in the open. Lots of people online are running personal versions of this. A markdown wiki that an LLM curates over time.

Karpathy’s pattern works for a person. The LLM-maintained markdown is structurally fine when the surface area is one human’s thinking. Maybe a small team.

A brand is a different problem. Even with an LLM doing the curating, a single artifact strains under brand-scale volume – hundreds of decisions, thousands of customer conversations, every campaign, every creator, every support pattern worth remembering.

Synthesis at that scale either collapses (everything compressed into vague generalizations) or fragments (the LLM gives up on a single artifact and you end up with a directory of related markdown nobody reads).

But the actual problem is shape. The wiki is the wrong shape for an agent doing operational work.

A retention agent asking “what rate did we pay the last creator like this one?” doesn’t want a wiki paragraph saying “our creator rates typically land between $3-5k.” It wants Kathy’s actual email from May 14 with the $4,800 rate quoted in her own words, cited. A support triage agent asking “is this complaint a pattern or a one-off?” doesn’t want a summary. It wants the five raw tickets, timestamps included, so it can say “the pattern started six weeks ago and here’s each instance.”

Wikis flatten evidence into decisions. Agents need both – the decisions AND the un-flattened evidence underneath.

For a brand, the brain needs a different shape. Two layers, working together.

Layer one: vectorize everything, chunk by chunk. Every Slack thread worth keeping. Every Doc someone drops in. Every creator email. All chunked, embedded, and stored where any agent can retrieve it semantically. This is the raw memory.

Layer two: wikify the important stuff, deliberately. When a capture isn’t just a record but the decision – the brand positioning, the architectural call, the competitive finding – you tell Mother (our AI coordinator) to promote it. Stable identifier. Cross-links. The original capture stays attached as evidence.

The wiki gives agents the decisions. The vector layer gives them the evidence those decisions were based on. Agents need both to act proactively.

Five surfaces. Each one a single gesture – a Slack reaction, a pasted link, a forwarded email.

1. Slack threads – capture a conversation by reacting to it.

You’re in a mapped Slack channel and a thread runs across a decision.

The whole back-and-forth, every reply, every message, flows into Brain within seconds. Or if you’d rather type:

2. Pasted URLs – capture a document or page by dropping the link.

Anywhere Mother is mapped:

Mother fetches it, parses it (Google Docs, Drive PDFs, generic PDFs, blog posts, marketing pages, product pages), and stores the rendered text with the original URL kept as the citation.

3. Raw text — capture a paste when there’s no URL.

Creator emails, inbound form submissions, hand-written notes, meeting snippets — anything the content lives in your message rather than at a link:

The text goes in with whatever metadata Mother can infer (sender, date, thread context).

4. Forwarded emails — capture an email that has no link.

Marketing emails are the obvious case. A Klaviyo flow, a competitor’s newsletter, a launch announcement - half of them have no “view in browser” link, so there’s no URL to paste anywhere.

So forward the email to Mother. Mother pulls out the email’s actual contents, hosts it at a stable URL, and replies with a one-line command to drop in Slack:

From there it’s the same pipeline as any saved page: the text, the images, and the email’s subject line as the title.

5. Images – capture the visuals, not just the text.

When you save a Klaviyo email, an ad creative, or a product page, the visuals carry as much meaning as the copy. Brain pulls the alt-text on every image automatically. Brands that want more can opt into two further tiers: vision-model descriptions (Brain captions each image so it surfaces in normal text search) and image embeddings (images become searchable by visual similarity – you can even query Brain with an image and get the closest matching captures back).

Capturing puts something into Brain's raw memory. Promoting it to the wiki says this one matters more than the rest. Most of what you capture should stay raw – the creator emails, the Slack back-and-forth, the competitor pages.

They're evidence, and evidence is most useful in its original, un-flattened form. But every so often a capture isn't just a record, it's a decision the whole brand will keep referring back to: your positioning, the reasoning behind how you priced, the definition of your core customer, the competitive finding that shifted the roadmap, the architectural call you don't want to relitigate six weeks from now. That's what belongs in the wiki.

You tell Mother – @Mother promote this to the wiki – and she gives it a stable identifier, cross-links it to related entities, and keeps the original capture attached underneath as the evidence it was built on.

The test is simple: if an agent or a teammate would otherwise have to re-derive the answer from scattered raw captures every time they need it, promote it to the wiki. If they just need to find the source, leave it raw and let retrieval do the work.

Once you’ve put a few things in, asking is open-form natural language. No syntax to memorize:

Each answer comes back with clickable citations – Slack thread permalinks, document URLs, the original creator email’s metadata.

If you specifically want the authoritative wiki view rather than the raw evidence underneath, ask the wiki directly:

When Brain doesn’t know, it’ll tell you. It won’t fall back on public model knowledge to fabricate an answer.

Once Brain is in place, your agents stop guessing.

A creator outreach agent checks Brain before drafting and references the actual rate you negotiated with similar creators. A retention agent checks Brain before proposing a campaign and pulls the reasoning from the last one that worked. A support triage agent checks Brain and flags that the same complaint has come in five times this month with citations to each ticket.

Today’s Brain captures what people write down. Next is capturing the brand’s lived activity — campaigns sent, ads served, tickets resolved, reviews left. Starting with Meta and Google ad performance (already flowing through Gentic), then Klaviyo, then support and reviews.

Go to gentic.co/brain. Authorize Slack. Pick the channels Mother should live in. Channels you don’t pick stay private. Retrieval is free. Capture meters at a few cents per item.

As with any Gentic product you can also use the Brain MCP and connect to it from Claude, Claude Code/Cowork, ChatGPT and any AI platform with MCP support.

The proper way to build a brand brain is a vector layer underneath, a curated wiki on top, and a capture loop that doesn’t ask your team to change how they already work. That’s the brain your agents need to do real work.

Feel free to reply to this email or DM Bora on LinkedIn with questions.