{"id":"2026492755430474002","url":"https://x.com/arscontexta/status/2026492755430474002","text":"","author":{"name":"Heinrich","username":"arscontexta","avatarUrl":"https://pbs.twimg.com/profile_images/2012958446891536384/neq1Tu46_200x200.jpg"},"createdAt":"Wed Feb 25 03:01:37 +0000 2026","engagement":{"replies":52,"retweets":125,"likes":1448,"views":836760},"article":{"title":"Company Graphs = Context Repository","previewText":"everything is a context problem\nwhen people say AI cant do real work, what theyre actually saying is they gave it bad context\n@alexalbert__ said 2026 will transform knowledge work (read this after you","coverImageUrl":"https://pbs.twimg.com/media/HCAPud7XcAEt1M8.jpg","content":"everything is a context problem\n\nwhen people say AI cant do real work, what theyre actually saying is they gave it bad context\n\n@alexalbert__ said 2026 will transform knowledge work (read this after you finished this article)\n\ni think hes right and the fundamental mechanism for this is structured context graphs that agents can traverse\n\ncoding was solved first because the structure was already there. codebases were text files with relationships between them\n\nthe agent reads the code, follows the imports and understands the architecture. that was natural because this is how programmers already worked with code\n\nknowledge work doesnt have that structure yet. mostly outdated knowledge bases or wikis nobody reads\n\nbut some people were building it anyway. the obsidian and tool for thought nerds spent years figuring out how to structure knowledge\n\nnotes that link to each other, ideas that are atomic and composable and maps of content that give you the topology of an entire domain\n\nturns out they accidentally engineered the perfect architecture for LLMs\n\n(this is what the methodology ars contexta, the art of context, is built around)\n\n## the company graph\n\nthe problem isnt that the context doesnt exist, its that its scattered everywhere and nothing can traverse it\n\nslack threads from 8 months ago that nobody can find, google docs with 12 versions, notion pages updated once and never again and most of it just lives in peoples heads, where it disappears when they leave\n\n@balajis put it well\n\nhes describing the pain\n\nstructuring knowledge properly is not an easy problem. when scaling knowledge it gets complex really fast\n\n(but this is exactly what the arscontexta plugin helps with, more on that later)\n\nwhat every company needs is a well structured company knowledge graph. thats the perfect context repository for your entire organization\n\nthese are real notes (md files) that capture: every decision with alternatives and reasoning attached, every meeting, not just the recording but extracted claims, decisions, action items and strategic shifts, everything you have published, every research session and every competitive analysis\n\nand of course, you can (and should) throw your code repositories inside this graph as well\n\none domain = one network of composable files\n\nskill graphs made this pattern graspable, but what im saying is this applies to every kind of context/knowledge/thoughts that can be structured as a traversable markdown graph\n\n## tacit knowledge\n\nthe hardest knowledge to capture isnt in documents, its in peoples heads\n\nwhen your CTO decides on postgres over mongo, maybe the decision gets written down. but the reasoning, the tradeoffs she considered, the context that made it obvious to her but invisible to everyone else is lost\n\ni wrote about this before: yapping is work now\n\nmeetings used to be the ultimate time sink, but now you can record conversations, an agent mines them exhaustively and the tacit knowledge locked in peoples heads becomes structured graph state\n\nthis is not about meeting summaries nobody reads, its active synchronization with your thinking and the externalized representation of your thinking\n\nyour agent works with the externalized version and thats why it needs to represent EVERYTHING you know\n\nmeetings are how you keep the graph in sync\n\n## agents as CEO\n\nwhat does a CEO actually do\n\nimagine running a company thats moving in fifty directions at once. the engineering department is building three products, marketing is positioning against competitors, sales is closing deals and the strategy shifts with every conversation\n\nthe CEOs job is to hold all of that\n\nthey should notice when the roadmap contradicts a decision from last quarter and when a competitor move changes what to build next\n\nthats a context problem\n\nwe started crafting a company graph for our own projects\n\nthe company note network holds everything as composable markdown files related to each other through wikilinks\n\nsidenote: little example about the framing of our CLAUDE.md:\n\nbtw humans have externalized knowledge for thousands of year, this is what really enabled progress\n\neach medium (like cave paintings, parchment paper, books, digital information...) let the next generation build on what came before instead of starting from scratch\n\nwe are standing on the shoulders of giants\n\nagents live in context windows like humans live in lifespans. they are temporary, bounded and forget everything when the session ends\n\nthey need externalized knowledge for the same reason we needed writing: to transcend the limits of individual memory\n\na company graph is an agents library. every session it picks up the accumulated knowledge of the entire organization and operates from there\n\n## you dont read the graph\n\nyou dont need to read everything in the graph. thats the whole point\n\nsame as hiring mckinsey for a strategic analysis. twenty analysts spend thousands of hours on your problem\n\nyou dont read all their internal documents, you just want the deliverable\n\nthe company graph is the research base, you interact with it and get what you need: a competitive briefing for tuesday, a pitch deck for the conference or dynamically rendered components or views inside an AI-native knowledge work app\n\nyou can kick off background jobs that go acquire more knowledge, synthesize different perspectives or craft deliverables for different audiences all simultaneously\n\nits like 10 employees aka subagents doing all the ground work so you can derive your next deliverable or action from it\n\nbut the structure matters A LOT. an unstructured dump of notes doesnt scale\n\nyou need wikilinks as semantic connections, atomic composable markdown notes, maps of content notes for navigation / attention management, metadata for queries or progressive disclosure and a few other \"kernel functions\". these are the primitives that make a graph traversable\n\narscontexta relies on a methodology graph to help you build systems like this. it constructs yours through semantic metaprogramming, you describe what you want and the system derives the architecture\n\nthis isnt just for companies, but for any field of knowledge work like your university studies, personal research or whatever\n\nand whats most important: you own your own memory because its just instructions, hooks and markdown files\n\nthe methodology graph of arscontexta also makes the whole thing antifragile. when youre exploring new structures or unsure if the architecture fits, you can always consult it and it gives you research-backed guidance for your individual setup\n\ngive it a try\n\n(little warning: rushed the release and debugging/testing \"semantic metaprograms\" and knowledge graphs is genuinely hard but were on it. please write an issue if you find something broken)\n\nbtw were building an app for this too. think cursor for knowledge work, or networked note-taking apps but rethought AI-native, deeply interwoven with arscontexta, the art of context\n\n## the graph improves itself\n\ncompanies have wanted one place where all knowledge is stored forever, but all \"solutions\" died the same death:\n\nmaintenance costs (imo this is also why tools for thought never went mainstream)\n\nsomeone had to keep it updated\n\nagents dont get bored of maintenance and they dont skip the update because theyre late for a meeting\n\nthe thing that killed every wiki is the exact thing agents are built for\n\nyou can build MUCH more complex methodologies and structures that would be unmaintainable for a human (actively exploring rn)\n\nalso a company graph with an agent operator is fundamentally different\n\nthe agent notices when two notes contradict each other and flags the tension\n\nit notices when the spec / PRD graph is out of sync with your codebase (yes. please apply that graph concept everywhere)\n\nwhile working, friction signals accumulate automatically and when enough observations pile up the agent proposes structural changes to the system itself\n\nit refactors its own instructions and it evolves its own architecture when the current one creates too much drag\n\nalso its easy to make this knowledge economically actionable and valuable\n\nstart putting your company into a graph\n\nthe structure doesnt need to be perfect on day one, future models will refactor the architecture easily and the arscontexta plugin creates the basic individual structure for you to get started\n\nwhat happened to software development with vibe coding is about to happen to knowledge work\n\n2025 was agents writing code, 2026 is agents disrupting knowledge work and steering companies\n\none last thing: your company is already a graph. the question is whether you can see it\n\nheinrich"}}