Back to the log

Note

#57 - Design a Company Brain / History, Memory & Knowledge / The Modern Enterprise Canvas

November 18, 2026KnowledgeAgentsAutomation
Part 1 of Build a Company Brain, a 6-part series. This is Design — what a company brain is and how to lay it out. Architecture and Operations follow in officehours 58 and 59, with hands-on study halls in between. Seats for the next session are limited — check the event page: lu.ma/intelcraft.

Link to Folder with Everything

2026.11.18.officehours.57

Deck

Recording

officehours 57 — full session recording (Nov 18, 2026)
Watch original recording

Watch on Google Drive

Summary

This kicked off a six-part build: by the end of the series, everyone has a private company brain their team and their AI can both read. Part 1 was Design — not tools, but how to lay the thing out so it survives contact with real work.

What a company brain actually is

Rahul's definition: a company brain helps people convert right knowledge into right action. It keeps the work and the knowledge in one structure that people and AI can both read. No single tool is the answer — he scanned roughly a hundred second-brain / company-brain tools and reverse-engineered them into a standard shape, and the lesson was that the design matters more than any product. The tools (LM Wiki, YC's gBrain, Tiago Forte's Second Brain, PARA, Obsidian, Every.to's compound engineering, OB1) mostly assume you know Git and code. The whole point of this series is to get a working brain with none of that.

The one distinction to hold onto: history, memory, knowledge

These are three different things, and conflating them is why people get lost. History is everything that happens in a system — events, logs, every chat — just a record. Memory is what gets selected out of history as important for the work at hand, saved for a specific worker or shared with the team; memories come and go as things change, and they aren't necessarily true. Knowledge is the subset that's been decided and validated — facts two people would both agree are right, central and durable. Most chat assistants keep their own private memory; what a real brain adds is shared, validated knowledge on top of history and memory.

The Modern Enterprise Canvas

The piece that's unique to IntelCraft — Rahul has been building it since 2011. It's the inventory: a single view of the whole company. Responsibility areas across the top (GTM, R&D, Operations/Infra, Finance, Products/Client, Leadership/KB), seen through four contexts — people, processes, information, systems. Put that inventory in one place and both people and AI know who's who, what the processes are, and what information needs tracking. It answers the question “What is the organization?”

The IntelCraft vocabulary: chat, workflows, routes, spaces

A deliberately plain way to talk about any AI system. Chat is where a person talks to something. Workflows are where things happen (Zapier, n8n, Make, code). Routes are how you connect to tools and data. Spaces are where the files and data live. None of these is “the brain” — they're how you connect a user to it: the user works in chat, the knowledge lives in a space, workflows do the work, routes wire workflows to spaces. Designing the message and how it moves through that system is the most important part.

Design before tools, and naming before everything

The frameworks are patterns, not products: GTD for capture/clarify/organize, PARA for where knowledge goes, OKR for where you're moving, VTO for where the company is headed, and an AI policy template for the rules (who can use which tools, which models are allowed). The non-negotiable starting point is a naming convention — consistent names for projects, areas, and threads across every tool. Rahul showed his own GTD before/after: a 2013-era tangle of inboxes and Zapier automations collapsed into a simpler Todoist + Notion setup. The payoff is that if everyone names things the same way, you can tell an agent “look at everything for GTM” and it just works — and the day you stand up a shared brain, everyone's existing work imports into the same place. Even with no shared brain yet, a consistent naming convention is the thing that makes one possible later.

The live demo: one commander, many workers

The part that drew the most questions. Rahul runs GrokBot as a commander and ChatGPT, Claude, Devin, and Lovable as workers. Tasks land in Todoist; a rubric (literally “if it's content or copy, send it to Claude; if it's this, send it there”) and a README he refined over a few runs tell GrokBot how to triage, label, and route. His agent queue showed the spread live — 2 tasks for Claude Code, 2 for GrokBot, 16 for Devin, 28 for ChatGPT. Then the trick: GrokBot opens a browser inside its own cloud computer, goes to claude.ai or chatgpt.com, and says “pick up this task from Notion and do it.” No API wiring — it just drives the apps a person would use. “It's stupid, it's simple, but it works.” A human-review step (“Rahul Reviews”) sits in the middle of the generated workflow diagram.

Connecting it, cheaply and privately

Minimum viable is native one-to-one connectors (Todoist, Notion, ChatGPT, Lovable) plus a cloud-synced OneDrive desktop every agent can reach — desktop, projects, and inbox folders shared across all of them. The advanced path, for a whole company, is an MCP gateway like Composio to consolidate and secure connectors with granular, atomic access (one server that only reads Notion and Outlook, another that can read and write). For people who'd rather own their stack, Rahul is a long-time open-source advocate: PikaPods for dead-cheap hosting (~$2/app/month — Airtable and Notion clones, Open To Do, Docmost, ArchiveBox), and Coolify to run your own cloud on AWS or a home machine, deploy straight from GitHub (CI/CD without paying Vercel per deploy), and self-host LibreChat, n8n, Hermes Agent, and memory services. Obsidian (Markdown knowledge, heavy PARA following), Linear (a Notion-like take on GitHub issues/Jira), and Maddy's Humble AI were all offered as valid swaps — anything with an MCP connector works, because these tools are all just a place to put information, move it through stages, and get it done.

Next steps

  • [Rahul Singh] Publish the research catalog of ~100 knowledge-management tools, the naming conventions, and a sample AI policy template on IntelCraft.
  • [Rahul Singh] Publish the session handouts and system schematics (what goes into the brain).
  • [Rahul Singh] Develop an offline path to Notion data that syncs to external storage (GitHub/S3 backup).
  • [The group] Draft an AI policy for personal or team use from the provided template, and share feedback on the company-brain concept.
  • [Everyone] This is a six-part build — the next session fills up fast. Check lu.ma/intelcraft and intelcraft.ai for the next office hours and the hands-on study hall.

Details

  • Session: officehours 57, November 18, 2026 — Part 1 (Design) of the Build a Company Brain series. Hosted by Rahul Singh with the IntelCraft team.
  • The hands-on companion: a study hall follows each office hours — smaller, in person, collaborative, not recorded. Handouts get published so you can follow along, but the room is where the building happens. Seats are limited.
  • Backup practice: because Granola keeps only the transcript and summary (no audio), the sessions are now also captured on Gemini for the video, and anything recorded can be repurposed into written posts or a podcast later.
  • This summary was written by Claude from the Gemini session notes and the session recording. The internal pre-session discussion that happened to be on the same recording is not included.