Back to all posts
    108 Second Brains: What I Learned Setting Up Everyone Else's AI

    108 Second Brains: What I Learned Setting Up Everyone Else's AI

    After helping 108 people build their AI work system, I rebuilt everything.

    April 10, 2026
    Updated July 13, 2026
    7 min read
    230 views
    by Iwo Szapar

    After helping 108 people build their second brains to use AI smarter, I built a system that anyone using Claude Desktop can use.

    77% of the people who paid me to set up their AI work system never wrote a line of code.

    And they were scared of terminal.

    That forced a complete rebuild. Here is what 108 setups taught me about why V1 failed half the people who wanted it, and what I built to fix it.

    Who showed up

    I expected developers. Engineers comfortable in the terminal.

    That is not who bought.

    108 setups across 4 months. Silicon Valley, Dubai, Berlin, Toronto, Lisbon, Warsaw, Sydney.

    • Consultants made up 35% of buyers.
    • Content creators and marketers, 25%.
    • Senior professionals protecting proprietary methodology, 22%.
    • System builders, 10%.
    • And 8% were people so overwhelmed by their tool stack they just wanted someone to make it stop.

    Paul, a consultant in London, told me: "Nothing really integrates, so I'm just getting little flashes of brilliance here and there." That sentence describes 80% of my customer conversations. Smart people using five different AI tools that do not talk to each other.

    42% cited cognitive overload as their primary reason for purchasing. 28% had knowledge scattered across five or more systems. 18% wanted to scale their expertise without hiring.

    Nobody said "I want to learn Claude Code." They said "I want my AI to know who I am and do real work."

    57% chose premium tiers ($597 or $2,497). They did not want just the system. They wanted someone to set it up for them.

    What V1 got wrong

    V1 started in the terminal. That worked for more technical users, but it meant a steeper learning curve for everyone else. Memory lived in markdown files you updated yourself. The system worked, but keeping it current took effort. And effort is what most busy professionals run out of first.

    Vincent, one of the early buyers, described the experience: "For a non-engineer the technical setup has been and remains confusing and daunting."

    Lucas, who has ADHD like me, said: "I have ADHD, sometimes I start with something new, and before I know it, I try something else."

    Any system that requires you to maintain it will lose the people who need it most. V1 assumed discipline. V2 had to assume the absence of it.

    Full V1 vs V2 breakdown: iwoszapar.com/second-brain-ai/v1-vs-v2

    The five-month rebuild

    Every design decision came from a specific failure I watched happen in a customer's setup. Five months. Here is what changed.

    1. Memory that maintains itself

    Three things kill every AI memory system: bad data gets in, good data goes stale, everything drowns in noise.

    V1 used markdown files. You edited them by hand. The system worked, but it did not improve on its own.

    V2 moved memory to SQL. Nine memory tables: knowledge, decisions, sessions, contacts, goals, tasks, content drafts, configuration, retrieval patterns. Auto-decay detection flags stale knowledge before it produces wrong answers. A convergence layer prunes noise and synthesizes patterns. Memory gets sharper over time instead of just bigger.

    Vanja, one of the earliest buyers, said what everyone was thinking: "I don't want to keep saying the same to the LLM." You do not have to anymore.

    2. The terminal requirement disappeared

    This is the biggest change. Claude Desktop support means you never need to see a command line. Claude Cowork goes further: autonomous task execution for people who never want to touch code. Same CLAUDE.md rules, same MCP servers, same skills. The interface changed. The architecture stayed the same.

    Three platforms now. CLI for terminal power users. Desktop for the GUI-first majority. Cowork for fully autonomous, no-code operation. Same system underneath all three.

    Dominic, who bought during the V2 beta: "I'm not a coder. I don't write code. So it's incredible that this type of stuff is available."

    3. Onboarding became a conversation, 24/7

    V1: read documentation, figure it out.

    V2: a 6-step questionnaire detects seven behavioral patterns and matches you to one of five personas. An AI agent configures your entire system from your answers. You do not read documentation. You answer questions about how you work, and the system builds itself around your answers.

    4. Quality became visible, on autopilot

    45 dimensions scored. Specific fix suggestions. A PDF dashboard you can share. No other AI tool in the market tells you whether your setup is good or bad. They all let you guess. The health check does not.

    How the whole system fits together: iwoszapar.com/second-brain-ai/how-it-works

    What the system actually does

    I run my entire business through the Second Brain. Not as a demo. As production infrastructure.

    A few examples from dozens of use cases:

    • Morning briefing. One command opens Gmail, WhatsApp, and the CRM simultaneously. Reads the calendar. Scans for overdue follow-ups. Pulls active tasks. Writes a plain-language summary of my day before I open anything. A disciplined person checks their calendar, reviews their CRM, scans their inbox in order. I built a system that does all three in parallel and hands me a summary.
    • Content creation. Spawns a creator agent and an evaluator agent. They do not share context. The creator writes. The evaluator scores against a 100-point rubric without seeing the creative process. No confirmation bias. Below 70, the creator revises.
    • Sales pipeline. The system reads a prospect's full CRM record, scrapes their LinkedIn for recent activity, pulls the previous email thread. Drafts outreach that references specific context from our conversation. A Gmail hook catches the send and writes a CRM entry automatically. I never log outreach manually. 200+ prospects with active records requires better infrastructure, not better memory.
    • Accounting. Imports CSVs from my Polish bank and Revolut. Parses transactions. Categorizes. Reconciles against Stripe records. Financial tracking is painful for someone with ADHD. Sustained attention to detail with no immediate reward. One command.

    80+ skills. 30+ agents. 20+ MCP servers. The AI talks to Gmail, Stripe, LinkedIn, WhatsApp, GitHub, Google Calendar, Google Analytics, your CRM. It controls your browser through Computer Use. Runs scheduled tasks, fires hooks, coordinates agent teams.

    Full feature list: iwoszapar.com/second-brain-ai/features

    Muhammed's description was the most concise: "Empty head; effective work; I'm elevated by AI that understands me and preps all real work for me."

    You own everything

    Celine, a privacy-conscious executive in Zurich, told me: "I don't want to give them my brain."

    That concern is the other half of the adoption problem. People will not invest in a system that stores their knowledge on someone else's servers.

    ChatGPT stores your memory on OpenAI's servers. Copilot lets your employer's IT admin see every prompt. Cancel your subscription, lose your context.

    V2 runs on your machine or your own cloud (Supabase). Cancel the subscription and keep every file, every memory, every configuration. The CLAUDE.md file that powers the system is readable by any LLM. No vendor lock-in. Your data stays yours.

    The economics

    A typical knowledge worker spends $200+ per month on disconnected AI subscriptions. ChatGPT Plus. Perplexity. Notion AI. Grammarly. Various browser agents. Each one starts from zero every session. None of them talk to each other.

    • Second Brain DIY — $197. One time. Not per month. One AI that reads your email, checks your calendar, updates your CRM, drafts your content, and remembers what happened last Tuesday. Not six tools that each do one thing badly.
    • Kickstart — $597. Same system, plus an AI agent configures everything from your questionnaire answers, 48 hours of priority async support, and the first year of MemoryOS Pro included ($349 value).
    • Done-With-You — $2,497. Two-hour onboarding call, full infrastructure deployment, monthly updates, 12 months of MemoryOS.
    • MemoryOS runs free for basic health checks. Standard ($199/year) adds full monitoring and all MCP tools. Pro ($349/year) adds monthly repo updates with new agents, skills, and tools pushed directly to your setup.

    All tiers: iwoszapar.com/second-brain-ai#pricing  ·  MemoryOS: iwoszapar.com/memory-os

    1,227 days

    I have used AI every day for 1,227 consecutive days. Not casually. In production. Building products, running a business, writing content, managing a pipeline, doing accounting.

    What I know after 1,227 days: the AI gets better every quarter. The models improve. The tooling improves. But the people using AI do not automatically improve with it. They plateau at whatever level of infrastructure they built in the first week. The model advances. The context stays stuck.

    That is what the Second Brain fixes.

    • Context that compounds.
    • Memory that prunes itself.
    • Personalization matched to how you actually work.
    • Quality scoring that tells you whether your setup is good or slowly degrading.

    I set up 108 of these. Each one taught me something about the gap between what AI can do and what people actually do with it. V2 is what I built to close that gap.