How Forward Future built an AI "second brain" that works on its own: from a single email stream to structured memory

Disclaimer: the downloaded content is only part of the article. The text cuts off at the paywall (“Keep Reading – Available to paid subscribers”), so what follows is based solely on the public portion. It does not include the later sections, such as what went wrong or what the author would do differently, although the subtitle announces them.

The author, Nick Wentz of Forward Future, recounts that a few months ago his AI spotted a broken link in an email to a partner. Nobody had asked it to. It found the correct link, updated the draft and flagged the change for him to review. For him, it was the first sign that his system worked: an AI that understands his work well enough to know when something needs attention without being told.

His thesis is that most people use AI like a search engine: you ask, get an answer and move on. He wanted a system that continuously assesses what is happening in the company, connects the various apps and starts useful work on its own initiative. He calls it his “second brain.” It is built around Grok Bot, which serves as both interface and orchestrator. The system remembers relationships and decisions, prepares meetings, tracks commitments, monitors business performance and coordinates a team of specialized agents. Some routines run on a schedule and others are triggered when new information arrives. He can also launch tasks simply by talking to it.

The main message for non-power users is that it all started with a single email routine. According to the author, if you can explain to a new employee how you triage your inbox, you can build version one.

Start with a single workflow. Wentz began experimenting with Grok Bot in August 2026 and gave it a single task: review his email every morning. The first version connected Gmail, HubSpot and Google Drive. It read incoming messages, retrieved relationship and business context and helped him decide what deserved his attention. One example: a sponsorship email mentions a company with which a deal was negotiated six months earlier. Previously he would have had to dig through old threads, check the CRM and track down the latest proposal. Now he wanted the agent to do that work.

He then added sources one at a time. Fathom, via its API, to pull transcripts of new meetings. Then Asana, Calendar and Slack, which make it possible to link conversations to the work they generate. Later he added the audience and revenue platforms: Beehiiv, YouTube and Stripe. Today the system reviews important emails and new transcripts roughly every 30 minutes during the workday. Other workflows react to changes in the connected systems, such as new transcripts, HubSpot deals changing stage or overdue Asana tasks, combining scheduled checks with event-based triggers. You don't need every integration to get started.

Recipe for a first morning briefing. The article proposes five steps with Grok Bot:

1. Connect email. Open Grok Bot, choose “Connect apps” and connect Gmail, starting with read-only access. HubSpot or other sources can be added later.

2. Give instructions. Paste a prompt in which the assistant acts as an executive assistant. Every morning it reviews new emails since the last briefing and uses earlier correspondence as context. It prioritizes messages that require decisions or replies, pending commitments, upcoming deadlines and significant risks or opportunities. It ignores newsletters, promotions and routine notifications. It delivers the three most important items, and for each one explains what happened, why it matters, what the user should do, and links to the original email. The briefing should take no more than two minutes to read. It must distinguish facts from assumptions, flag anything unverified and not send emails or modify records without approval.

3. Test and refine. Ask Grok Bot to run the briefing immediately and check whether it caught the important messages and filtered out the noise. The author's first briefings summarized too much, and he improved them by telling it not to summarize the inbox and to surface only what requires a decision, affects an existing commitment or presents a significant risk or opportunity. A few rounds of refinement, he says, make a big difference.

4. Automate. Once the output is satisfactory, ask it to save the process as a reusable skill and schedule it for weekdays at 7:30 Pacific time. Delivery happens in the same conversation, and it flags any sources it could not access. Then confirm the schedule and use “Test” to verify the routine.

5. Add context. After a few days, connect HubSpot or another relevant source. That way the assistant can recognize that a seemingly routine email is tied to an active negotiation or a pending commitment. According to the author, that is where it becomes truly useful: instead of reporting what is new, it starts identifying what matters. The goal is a two-minute daily briefing with priorities, context and recommended next steps. The article's visual example is illustrative, using fictitious data instead of sensitive information.

A usable memory. Chat history is useful for remembering preferences, such as favoring bullet points over long sentences. It is less reliable for precise business commitments, changing terms and information that must stay accurate over time. That is why the author built a more structured memory system, with Convex as the primary database. It organizes Forward Future's information into records of people, companies, projects, conversations, deals, decisions and commitments, and tracks the relationships among them.

Those relationships get complicated fast. A single company can simultaneously be a sponsor, an interview guest, a technology partner and the maker of a product being considered for coverage. The system must recognize those connections and keep business relationships firmly separate from editorial decisions. Each important record includes its history, associated documents, outstanding responsibilities and links to the original sources. The system updates them as new information arrives, distinguishing confirmed decisions from proposals, assumptions and outdated information. Thus, a tentative sponsorship date does not carry the same authority as a signed agreement.

Example: six videos, several months, one deal. They recently negotiated a six-video sponsorship package with a new partner. The deal required spacing the integrations over several months, included stipulations regarding competitors and prohibited back-to-back placements. The problem was that none of this lived in a single place. The negotiation took place over email and in meetings, the deal was in HubSpot, the commercial terms in the contract, and the production schedules and responsibilities somewhere else. The second brain ties it all together. If a new email proposes dates that clash with an existing commitment, the system is designed to detect the discrepancy and flag it. The key, according to the author, is not remembering everything but knowing which information reflects the current commitment: a proposed date, a revised offer and a signed contract do not carry the same authority.

Assessment. The accessible portion offers two practical takeaways. Start with a small, well-scoped workflow, with read-only permissions and human approval before sending or modifying anything. And treat memory as structured data with a hierarchy of authority, not as a conversation history. The details of the failures and what the author would change remain behind the paywall and cannot be summarized here.