How Almond works: capture, keep current, compound, prioritize

The problem this solves before we explain how it works

Most AI tools for performance marketing have solved the same 30 percent of the job: platform mechanics, bid logic, creative scoring. The other 70 percent is business context. Why the account is structured the way it is. What you tested and killed in March. What leadership decided on Monday. That context lives in meeting recordings, Slack threads, email and three people's heads. No AI sees more than a fraction of it, and what does get captured goes stale the moment it is saved.

Almond is built to close that gap. The product is a Business Context Engine, also called an agentic memory layer, and it works in four steps: capture, keep current, compound, prioritize.

Step 1: Capture, automatically and continuously

The first thing Almond does is connect to the places where your business context actually lives. That list includes Google Ads, Meta, Slack, Notion, Google Sheets, email, meeting notes and your CRM. All connections are read-only. The setup takes about four minutes.

Read-only matters for two reasons. It removes the security concern that comes with giving a system write access to live ad accounts, and it means the connection can be made without an engineering ticket.

Once connected, Almond reads continuously. It is not waiting for someone to upload a brief or paste a summary into a prompt. When your Monday strategy call ends, the notes are ingested. When a Slack thread decides to pause a campaign, that decision is captured. When a Google Ads experiment concludes, the result is recorded alongside the reasoning that preceded it.

The before state most performance marketers recognize: "I spent twenty minutes typing background before I could ask the actual question." The after state Almond is designed to create: you ask the question, and the context is already there.

Step 2: Keep current, without human re-entry

Capturing context once is not enough. Performance marketing context changes every week. Budgets shift. KPIs get redefined at a quarterly business review. A competitor launches and reframes the whole creative strategy. Brand guidelines get updated.

Most AI tools in this space hold a static brief. You load it once and it ages. The tool keeps optimizing for a KPI you dropped two months ago because nobody re-uploaded the document.

Almond's context layer is updated as events happen. Every meeting, experiment and decision that flows through the connected sources updates the context store in real time. Nothing requires a human to re-enter, re-upload or remember to flag it.

A concrete example from practice: a budget reallocation decided on Monday is reflected in the agent's context on Monday. The agent does not keep recommending spend toward a campaign the team decided to sunset, because the decision to sunset it was captured when it was made, not when someone eventually got around to updating a shared doc.

Step 3: Compound, across quarters and years

A prompt window holds weeks of context at best. Human memory is inconsistent and non-transferable. Neither compounds.

Almond versions every piece of context it captures. Every test result, every creative decision, every account restructure is retained and linked to the reasoning that produced it. Over time, this builds what the brand notes describe as a set of Principles: patterns learned from hundreds of decisions across your account history.

The practical difference shows up in two scenarios.

First, the repeated-mistake scenario. Without compounding context, an AI agent proposes the exact test you ran and killed in March. With it, the response is: "You tested this in March. It failed on creative, not targeting. Here is the variant worth running now."

Second, the staff-turnover scenario. When your best account lead resigns, they take their relationships and their opinions. Without a context layer, they also take the reasoning behind every structural decision in the account. With Almond, that reasoning is versioned and retained at the system level. The account does not reset when a person leaves.

For agencies specifically, this is the difference between expertise that lives in people and expertise that lives in the agency. One of those is an asset on your balance sheet. The other walks out the door.

Step 4: Prioritize, by relevance, recency and business impact

Connecting more sources does not automatically produce better answers. Several practitioners in the brand notes describe the opposite experience: "We connected all our sources and the answers got worse." Fourteen recommendations with no sense of which one matters this week.

The reason is that not all context is equal. A throwaway Slack message and a CEO strategy call both contain text. They do not carry the same weight.

Almond classifies every piece of ingested context by three dimensions: relevance to the current question, recency relative to when decisions were made, and business impact based on what the decision affected. A strategy call outranks a casual thread. A decision made last week outranks one made eighteen months ago, unless the older one established a structural principle that still governs the account.

The output of this classification is prioritized context, not a data dump. Instead of fourteen recommendations, the agent surfaces the three that matter this week, with the reason each one made the list.

What is built on top of the engine

The Business Context Engine is the foundation. Three products are built on it.

Canvas is an AI visual dashboard. It surfaces live boards of charts, tables and AI answers, all drawing on the context layer. Prebuilt templates cover everything from account audits to ad hoc analysis. Reports are shareable and client-ready without reformatting.

Workflow Builder is a visual automation tool with 30 prebuilt Google and Meta automations. The key distinction from trigger-based automation is that every workflow runs on the context layer. It acts on judgment, not just conditions. A workflow will not scale a campaign that the team decided to sunset on Monday, because the decision to sunset it is already in the context store.

Integrations are built on open protocols, including MCP. This means the context layer works inside Almond's own tools or inside any AI client your team already uses. The context is not locked to a single interface.

Security, because context is sensitive

The more context a system holds, the more the security model matters. Almond's approach on the points that matter most to performance marketing teams and agencies:

• Your context is never used to train Almond's models or anyone else's.
• Every context store is client-scoped and isolated. One client's strategy cannot surface in another client's recommendation, enforced at the data layer.
• You can bring your own LLM, so business context stays within the model boundary your security team has already approved.
• Every piece of context is versioned and auditable. You can always answer the question that matters after something goes wrong: what did the agent know, and when did it know it?
• All data is encrypted in transit and at rest, with role-based access controls.

Where Almond fits

Almond is in early access. Every team that joins gets a detailed 1:1 Meta and Google ad account audit with the CEO, recorded. The waitlist requires a work email.

The honest framing for where it fits: if your team is already using AI for performance marketing and finding that the outputs require constant context-setting, go stale quickly, or repeat mistakes the account already made, the gap is almost certainly the memory layer, not the model. That is the specific problem Almond is built to solve.

World's First Agentic Brain for Autonomous Performance Marketing. Learn more at almond.ai.