Skip to content

The Team Second Brain

A small team already has all the knowledge it needs — scattered across everyone's individual AI chat history, one account at a time. A two-person shop and a fifteen-person team share the same problem: each person's AI tool learns from that person alone. Nothing compounds across the team, and nobody has an enterprise knowledge-platform budget to fix it.

The pattern already exists — it just stops at one person

This pattern comes from Andrej Karpathy, who describes it in a gist he calls the LLM Wiki: a large language model (LLM) "incrementally build[s] and maintain[s] a persistent wiki — a structured, interlinked collection of markdown files that sits between you and the raw sources" (Karpathy, 2026). Instead of an AI tool re-deriving an answer from scratch every time (the way most retrieval-augmented search works), it reads new material once, folds it into the existing wiki, and keeps that wiki current going forward. Three layers make it work: the raw sources (immutable), the wiki itself (the LLM-maintained markdown), and a short schema file that tells the LLM how to structure and update it.

Karpathy names the target use explicitly: "Personal... Research... Business/team... anything where you're accumulating knowledge over time" (Karpathy, 2026). Team use is not an edge case of this pattern. It is one of the three cases Karpathy named it for.

The pattern also answers the objection a small team raises first: who has time to keep a wiki current. Karpathy's answer is that an LLM does the keeping: "Humans abandon wikis because the maintenance burden grows faster than the value... The wiki stays maintained because the cost of maintenance is near zero" (Karpathy, 2026). A wiki a person has to update by hand gets abandoned. A wiki an LLM updates as a side effect of normal use does not carry that failure mode.

Phase one, already built: an ecosystem, one person at a time

This is not a fringe pattern. A direct search of GitHub for projects built on it turns up dozens of active repositories, several with real adoption. Khoj alone has drawn more than 35,000 stars for a free, self-hostable AI second brain (Khoj, 2026), and a cluster of projects that cite Karpathy's pattern by name each carry thousands more, actively maintained (nashsu/llm_wiki, 2026). Every one of them is single-user. Some make that a selling point outright: one markets itself with "zero sharing, your notes stay yours."

Search for the team-shared version instead, and the results nearly disappear. The one genuine attempt found is a real, well-designed system — a master brain file, persistent memory layers, executable skills, built to work with any AI tool. It has zero stars, zero outside contributors, and a single creator (accountsFON/second-brain-system, 2026). Someone has already tried to build this, thoughtfully, and it has not caught on.

The gap: general-purpose team knowledge, not yet built for a small team

The Single-Player Ceiling surveys what people have actually built on Karpathy's specific pattern — real, popular, actively maintained projects. Every one of them targets only a single person's own machine and accounts, whatever topic that person feeds it.

The team layer itself is not hypothetical. Granola already runs it for one narrow input: a team can chat across every meeting transcript in a shared folder, free, on Granola's own Basic tier (Granola, 2025; Granola, 2026). Meta is piloting an internal team-level layer it calls Third Brain, where each employee's own workspace feeds a shared one across dozens of teams (Analytics at Meta, 2026). Dr. Tali Rezun has proposed a specific architecture for exactly this: each person's own brain stays private by default, and the shared layer holds only what a person explicitly chooses to contribute (Rezun & Kapusta, 2026).

What's still missing is the general case, self-serve, for a small team without an engineer: a shared wiki built from everything a team's AI use touches, not just meeting transcripts, and not an internal pilot at a company that already employs its own AI infrastructure team. This bundle covers that general case: a team layer on top of Karpathy's pattern, built for whatever a small team actually feeds it, not one input type.

The same idea, turned inward

What OKF Is makes a structural argument for why organizational knowledge needs a format at all: "Most organizational knowledge lives in places that do not travel well: wikis, slide decks, scattered documents, and people's heads." That argument is about publishing knowledge outward, for other people and other agents to read. This bundle is the same argument turned inward — a small team applying that same discipline to its own working knowledge, before any question of publishing anything externally ever comes up. The format problem and the team-practice problem are the same underlying claim at two different scopes.

How to use this bundle

Read The Single-Player Ceiling next for an honest look at what's already built and where it stops being useful for a team. Then Building Yours Without an Engineer sketches a way a small team could set this up and keep it alive, without hiring anyone to run it — an untested design, not a proven one.

Sources

  • Andrej Karpathy — LLM Wiki (personal gist), 2026. incrementally build and maintain a persistent wiki — a structured, interlinked collection of markdown files that sits between you and the raw sources. View source · verified 2026-07-05 · primary
  • Andrej Karpathy — LLM Wiki (personal gist), 2026. Personal... Research... Business/team... anything where you're accumulating knowledge over time. View source · verified 2026-07-05 · primary
  • Andrej Karpathy — LLM Wiki (personal gist), 2026. Humans abandon wikis because the maintenance burden grows faster than the value... The wiki stays maintained because the cost of maintenance is near zero. View source · verified 2026-07-05 · primary
  • Khoj (khoj-ai) — khoj-ai/khoj (GitHub repository), 2026. Your AI second brain. Self-hostable. Get answers from the web or your docs. View source · verified 2026-07-05 · primary
  • nashsu — nashsu/llm_wiki (GitHub repository), 2026. LLM Wiki is a cross-platform desktop application that turns your documents into an organized, interlinked knowledge base — automatically. View source · verified 2026-07-05 · primary
  • Five One Nine Marketing (accountsFON) — accountsFON/second-brain-system (GitHub repository), 2026. A plug-and-play system for building a shared 'second brain' for any organization using plain markdown files. Works with any AI tool. View source · verified 2026-07-05 · primary
  • Granola (Chris Pedregal, CEO) — Granola 2.0, 2025. your company's new collective brain. View source · verified 2026-07-05 · primary
  • Granola — Granola Pricing, 2026. AI chat within and across meetings... Shared folders for collaboration. View source · verified 2026-07-05 · primary
  • Analytics at Meta — How We Built an AI Second Brain for 60K Knowledge Workers, 2026. A team-level shared context system, internally called 'Third Brain', lets team members' individual workspaces feed into a shared knowledge layer. View source · verified 2026-07-05 · primary
  • Dr. Tali Rezun and Dražen Kapusta — The Shared Brain: When Second Brains Start Thinking Together, 2026. Private brains remain private. Shared intelligence is built only from explicit contributions. View source · verified 2026-07-05 · primary