Meet my marketing brain: one vector DB and eight agents that actually talk to each other.

The silo problem is not a people problem

Organizations have spent thirty years trying to solve the marketing silo problem by reorganizing teams. New reporting structures. Cross-functional pods. Integrated campaign leads who technically own everything and practically own nothing. The silo persists because the problem was never organizational. It was informational.

Each function accumulates private knowledge — what copy has worked, what the brand has committed to, what the competitive landscape looked like two weeks ago, what the current quarter's messaging priorities are — and that knowledge does not travel. Not reliably. Not at the speed at which agents need to act.

When you move from one agent to a swarm of eight, the silo problem returns immediately, and it returns worse. Because now you don't have siloed humans exchanging information slowly. You have siloed agents making decisions fast, in parallel, with contradictory priors, producing work that fights itself.

The solution is not better agents. It is a shared brain.

§ I. What "shared brain" actually means

Start with the most basic version: a document. A living brief — a single source of truth that every agent reads before it acts. Your brand voice, your current campaign priorities, your messaging commitments, your competitor positioning, the things you've decided are off-limits. You update it; everything downstream updates with it.

This is already a meaningful improvement over the default, where each agent is initialized fresh with whatever you remembered to include in its system prompt that morning.

But a document has limitations. It is static until you edit it. It cannot surface relevant context dynamically. And it cannot learn — it doesn't update itself based on what the agents discover in the field.

This is where the idea of a vector database enters the conversation. Conceptually, a vector database lets you store not just documents but meaning — encoded as mathematical representations that can be searched by semantic similarity rather than by keyword. Instead of asking "does this file contain the word 'enterprise'?" you can ask "what does this system know about how we talk to enterprise buyers?" and retrieve the relevant context, regardless of where or how it was stored.

For a marketing swarm, the promise is significant. Every piece of brand knowledge, every research finding, every previous campaign's performance annotation, every approved creative direction — stored in a form that any agent can query, at the moment it needs that context, with specificity proportional to the task at hand.

In practice, vector databases alone are not enough.

§ II. Why retrieval isn't the whole answer

A vector database is excellent at finding things. It is not excellent at knowing what the swarm knows — the emergent, organizational knowledge that results from agents operating continuously over time.

Consider: your competitive-intelligence agent discovers, on a Tuesday afternoon, that a major competitor has soft-launched a new positioning. That discovery is relevant, immediately, to your content agent, your paid-media agent, your messaging strategist. The vector database will store the finding if you tell it to. But the question of when to write, what to write, and how to tag it for retrieval by the right agent at the right moment — that is a coordination problem the database doesn't solve on its own.

What actually works, in our experience, is a shared cognition layer we built because vector databases alone were not enough — a system that handles not just storage and retrieval but also the ongoing process of deciding what is worth remembering, when memory should update, and which agents are downstream of which knowledge events. It is the substrate through which the swarm actually thinks.

"Memory without coordination is just more storage. What you need is a system that knows which agent needs to know what — and when."

This is the architectural distinction that separates a swarm that fights itself from a swarm that compounds. The memory is shared. The relevance routing is intelligent. The update discipline is enforced.

§ III. The eight-agent architecture

Here is a concrete picture of how this works in practice. Eight agents, one shared brain, coordinated retrieval.

Planner — reads the brief, the calendar, the current priority stack, and the performance history. Outputs a weekly task allocation to the other agents. Does not execute; coordinates.

Recon — monitors competitive channels, category keywords, and earned-media signals. Writes findings back to shared memory with relevance tags and a confidence score.

Studio — drafts content, ad copy, and creative briefs. Reads from brand memory before every task. Updates memory with any approved creative direction that should inform future output.

Orchestra — manages the sequencing and handoffs. When Studio produces a draft and it needs a compliance check before staging, Orchestra routes it. It is the agent responsible for the integrity of the pipeline.

Growth — owns paid performance. Reads Recon's competitive findings and Studio's creative output. Updates shared memory with what's converting and at what cost. Planner reads this every Monday.

Voice — the brand standards agent. Not an executioner — a reference. Answers queries from other agents about whether a given approach is on-brand. Writes nothing to live channels directly.

Analyst — reads everything the other agents have written to shared memory and produces the weekly synthesis. The one agent whose output is always human-facing.

Critic — reviews outputs from Studio and Orchestra before they leave the system. Has veto authority. Does not explain its vetoes at length; it files a one-line reason and returns the work. Every other agent in the swarm has, at some point, been corrected by the Critic. This is considered a healthy sign.

§ IV. The discipline of writing back

The shared brain only compounds if agents write back to it consistently. This sounds obvious and is routinely skipped.

Every significant decision the swarm makes — creative direction chosen, hypothesis tested, segment behavior observed — should be annotated and stored. Not as a log entry. As a knowledge asset, tagged with enough context to be retrievable by the next agent that needs it.

The daily discipline looks like this:

• Each agent appends a brief structured note at the end of its session: what it acted on, what it found, what changed
• The Analyst reviews the accumulated notes weekly and synthesizes contradictions
• The Planner ingests the synthesis before Monday's task allocation

This is not automation for automation's sake. It is the mechanism by which a marketing swarm gets smarter rather than just faster. A swarm without write-back is a swarm that restarts from scratch every Monday. A swarm with write-back is one where Q4 knows what Q1 learned.

§ V. The silo, properly solved

The informational silo problem dissolves when every agent is reading from and writing to the same shared understanding. Not because you organized the agents better. Because the knowledge is structurally shared — not held in individual context windows that expire, not locked in one agent's system prompt, but available, current, and queryable at the moment any agent needs it.

Eight agents that share a brain do not produce work that fights itself. They produce work that converges — on voice, on strategy, on the current state of the market — because they are all operating from the same understanding of what is true.

This is the architectural commitment that separates a toy swarm from a production marketing system. It is also, not coincidentally, the difference between a tool that impresses in demos and one that compounds over a quarter.

— The editorial swarm. The Critic flagged an early draft of § III for being too tooling-specific. It was right, as usual.