Second Brain

Summary: An external system holds your knowledge so your head, or your team’s, is free to think. The idea is the same at every scale: a person’s second brain (what I learn), a personal AI OS (how I work), and an org’s company brain (what we know and why). People and organizations have always failed to maintain one. Notes rot and every company’s Confluence decays. LLMs change that by doing the upkeep: they capture, sync, connect and resolve contradictions, so each new task starts with full context. Agents are increasingly the main readers, so the bigger the scope, the more value comes from the “missing why” and from governed actions, and the more expensive syncing and conflict resolution gets.

Key ideas

Personal scale: what I learn

• Humans may simply not be built to be knowledge maintainers; well-filed knowledge still rots (osim-tochna—second-brain-and-llm-wiki, dana-maman).
• With a second brain, coming back to a topic you start at expert level instead of from scratch (osim-tochna—second-brain-and-llm-wiki, amit-bendor).
• Scattered capture (WhatsApp link groups, LinkedIn saves, bookmarks) bloats and goes unused; a single place with processing fixes it (osim-tochna—second-brain-and-llm-wiki).
• Use it for self-reflection: compare yourself to an expert’s framework (“where am I against this?”) and close gaps. Recorded meetings revealed things she’d missed in the moment (osim-tochna—second-brain-and-llm-wiki, dana-maman).
• Obsidian’s graph view shows growth over time and exposes orphan notes, topics connected to nothing (osim-tochna—second-brain-and-llm-wiki).
• Decide where to draw the line: Dana deliberately left out ~5,000 stale contacts (osim-tochna—second-brain-and-llm-wiki).

Work scale: how I work (personal AI OS)

• Extending the brain from “what I learn” to “how I work”: meetings, projects, apps, suppliers, writing, design system, even dreams. Dana Maman calls hers “Dana-OS”, a one-woman “factory”; the wiki is just its R&D department (osim-tochna—second-brain-and-llm-wiki, dana-maman).
• All context must live in one place to give the most value, as long as it’s curated, not bloated (osim-tochna—second-brain-and-llm-wiki).
• Built once, reused forever: each lesson (a security checklist, a dev concept) becomes part of how everything is done next; a design system makes quotes and HTML decks look great instantly; a social agent comments in her voice (osim-tochna—second-brain-and-llm-wiki).
• She measures the factory with dashboards (system health, loops, token use, time-to-result, app health) because she doesn’t read code (osim-tochna—second-brain-and-llm-wiki).
• Relevant beyond developers: any business owner or writer can use it (osim-tochna—second-brain-and-llm-wiki).
• A concrete instance: a meeting-centric assistant that briefs, records, summarizes, files tasks, and drafts emails, learning from every correction (langtalks—72-personal-assistant-agent).

Org scale: what we know (company brain)

• Definition: a shared knowledge layer (a Y Combinator “call for startups” idea) every agent connects to (often via MCP/API), reads, and writes memory into; success = how well agents get context, and value is measured in money through the loops it enables (ai-engineering-podcast—company-brain, roi-zalta).
• Not solved even at team level: the technology mostly exists, but continuously syncing sources and resolving contradictions is expensive (ai-engineering-podcast—company-brain).
• Example: Cerebras built an organizational “knowledge layer” every employee’s agent reaches via MCP, doing agentic RAG over all sources and even routing to human SPOCs (ai-engineering-podcast—company-brain).
• Teams that indexed everything (all sources, code as source of truth) found it too general. Moving to team-level context, then to layers, made the big jump (ai-engineering-podcast—company-brain).
• Layer 1, factual memory: meetings, Slack/Teams, email (more formal, so more trusted), docs (official > personal) (ai-engineering-podcast—company-brain).
• Layer 2, the “missing why”: decisions, commitments, and reasons, often made in hallways and meetings, rarely recorded. Capturing them is the context-building agent’s job (ai-engineering-podcast—company-brain).
• Layer 3, context-graph reasoning: an ontology of actors, interactions, time, rationales, and dependencies (see knowledge-graphs) (ai-engineering-podcast—company-brain).
• Layer 4, governed actions: loops that create value, e.g. onboarding a new C-level in weeks, not six months, or telling each customer which shipped feature they asked for two years ago (ai-engineering-podcast—company-brain).
• Ontology lenses: the same event means a feature gap to a PM and a churn signal to finance; the brain should inject role context into each agent’s queries (ai-engineering-podcast—company-brain).
• Humans are knowledge nodes: the brain should know that some knowledge only lives with “the IT guy” and route the agent to ask him (ai-engineering-podcast—company-brain).
• Authority weighting: naive approaches trust C-level over junior employees; better to weight by information type (a developer outranks the CEO on code) and by usage (ai-engineering-podcast—company-brain).
• Maybe the center shouldn’t be the company but its customers (ai-engineering-podcast—company-brain).
• Domain-expert knowledge, not frontier models, is the durable edge (Satya Nadella, as cited) (ai-engineering-podcast—company-brain).
• Wonderful’s internal “Wonder” agent connects to Drive, calendars, public Slack, Salesforce, the product, and Snowflake so nobody asks colleagues factual questions. It ranges from “why is prod broken?” to cross-data research (e.g. FDE background vs delivery speed); next step: taking actions (ignore-instructions—24-wonderful-road-to-100m, roy-lazar).
• Early-stage teams generate “shadow data” (customer meetings, Slack and WhatsApp threads, hallway and late-night architecture debates) that matters most precisely when there’s no product yet (startup-for-startup—353-self-updating-team-brain).
• monday’s Harmony team brain: every member’s machine runs a Claude skill every ~30 minutes that pulls new items via their own MCPs (note-taker, Slack, WhatsApp, docs), updates a shared git wiki, and pushes, so each person is a node and the brain stays alive without per-person workflows (startup-for-startup—353-self-updating-team-brain).
• Delivery surfaces: a Wikipedia-style site for morning reading, the repo inside Cursor or Claude for builders, Figma MCP plus the brain for designers, and a WhatsApp agent for anyone, including people on vacation or reserve duty (startup-for-startup—353-self-updating-team-brain).
• Results: a vision deck for the CEO in hours instead of weeks; Google Ads messaging written by Claude from the brain, which knows more than any single member; alignment across product and go-to-market without extra syncs. Next: autonomous per-discipline agents built on top of the brain (startup-for-startup—353-self-updating-team-brain).
• Culture shift: meetings without the note-taker became “unheard of”, and the team even records in-person conversations (startup-for-startup—353-self-updating-team-brain).
• Claroty’s agentic assistant “Clar” was built after an internal tour collecting best practices from different departments, so the model reflects the company’s cumulative expertise and can take operational actions, not just report (hidden-layers—claroty-ben-mashiach, ben-mashiach).

Disagreements & open questions

• Total transparency, bug or feature? Everything recorded or written becomes visible to the whole team; the guest now sees it as a feature (“everyone knows everything”), while acknowledging people still want to gossip in meetings and must switch the note-taker off to do so (startup-for-startup—353-self-updating-team-brain).
• Who builds it? Agents will build the brain that gives agents context, so whose context builds it? Humans may lose the ability to read or audit the org’s context (ai-engineering-podcast—company-brain).
• Index everything or curate? Dana keeps one place for all context but deliberately excludes stale data (osim-tochna—second-brain-and-llm-wiki); company-brain builders found indexing everything too general and moved to team-level, layered context (ai-engineering-podcast—company-brain).

Takeaways

• Use Obsidian’s graph view to find orphan notes (osim-tochna—second-brain-and-llm-wiki)
• Before building, sketch your life domains on paper (circles + arrows) to decide what goes in (osim-tochna—second-brain-and-llm-wiki)
• Ask the AI to interview you to personalize the system (style, interests, work) (osim-tochna—second-brain-and-llm-wiki)
• Turn each newly learned practice (security, privacy checklist) into a reusable skill or rule (osim-tochna—second-brain-and-llm-wiki)
• Start from a concrete business use case and map its data sources before building a company brain (ai-engineering-podcast—company-brain)
• Build team-level context before attempting company-wide (ai-engineering-podcast—company-brain)
• Capture the “missing why”: record decisions with their reasons (ai-engineering-podcast—company-brain)
• Weight sources by type and authority (email > chat; official docs > personal notes) (ai-engineering-podcast—company-brain)
• Prototype a team brain locally (note-taker + Slack MCPs into a small git wiki) before automating it (startup-for-startup—353-self-updating-team-brain)

Related

llm-wiki · knowledge-rot · knowledge-graphs · rag · memory-consolidation · personal-ai-assistants · ai-guardrails · enterprise-ai-adoption