Why every company needs a company brain

Most companies are not short on tools. They are short on memory. The decision from Thursday lives in a Slack thread. The exception lives in someone’s head. The ticket shows the outcome and hides the why. A new person — or a new agent — starts cold.⁹

That gap now has a name. A company brain is a persistent context layer: it gathers what the company already creates and makes that context usable by people and by AI agents. It is not a wiki with a chat box. It is not enterprise search. Its job is to keep the why attached to the work.⁴

How the category formed

The phrase is new. The pieces are not. Personal “second brain” practice — most visibly Tiago Forte’s method — taught a generation to capture and connect their own notes. Enterprise search, a decade earlier, made scattered files findable. Neither kept a living model of how the company actually decides.¹⁵⁴

On 22 December 2025, Jaya Gupta and Ashu Garg at Foundation Capital argued that the next systems of record would not own objects — customers, employees, invoices — but decisions. Salesforce, Workday, and SAP store the outcome. They do not store the exception, the approval, or “we did it this way last time.” Those traces live in Slack, on a Zoom, in a DM. Agents that only see the object layer cannot replay judgment.²

The next generation of defining platforms won’t be systems of record for objects. They’ll be systems of record for decisions.

They called the accumulated structure a context graph: a living record of decision traces stitched across entities and time, so precedent becomes searchable. A month later they wrote that the idea had become one of the most-discussed bets in enterprise AI — the layer above canonical data, the reasoning that connects data to action.²³

In 2026 the threads met. Andrej Karpathy described a persistent layer between raw sources and retrieval — an LLM wiki that reads incoming material and synthesizes interlinked pages. Garry Tan open-sourced GBrain so his own agents could work from meetings and mail. Y Combinator’s Summer 2026 Requests for Startups included an entry by Tom Blomfield titled Company Brain: “Garry’s G-Brain, but for every business in the world.” Not company search. Not a chatbot on a wiki. A new primitive.⁴[[ycRfs]][[ycArchive]]

Every company has critical know-how scattered across people’s heads, old Slack threads, support tickets, and databases, and AI agents can’t operate like that.

YC’s public note is blunt: every company will need a living map of how it works, turned into an executable skills file for AI. The models exist. The company-brain layer, they argued, does not yet.[[ycLinkedin]][[ycArchive]]

The value was already on the books

Before anyone said “company brain,” McKinsey Global Institute measured the drag. The average interaction worker spends an estimated 28 percent of the week on email and nearly 20 percent looking for internal information or tracking down a colleague who can help. A searchable record of knowledge can cut the time spent searching by as much as 35 percent. Used well, social and knowledge tools could raise interaction-worker productivity by 20 to 25 percent — and two-thirds of that value sits inside the company, not in marketing.¹

That is one day a week, per person, spent hunting for something the company already knows. The report is from 2012. The tools multiplied. The hunt did not go away.

Turnover makes the same hole wider. Atlan’s 2026 synthesis of institutional-knowledge research attributes a $1.3 trillion annual U.S. cost to knowledge-worker turnover (Deloitte, 2024), replacement costs of 50–200 percent of salary (SHRM, 2023), and an 8–12 month ramp to full productivity in knowledge-intensive roles. It also reports Gartner’s 2024 estimate that 70–80 percent of enterprise knowledge is tacit — never written down. Those figures are compiled, not original to us; the pattern is stable across the sources.¹²

Gyld puts the daily version simply: a new engineer spends the first two weeks asking questions that were answered six months ago in a Notion doc nobody can find. Data is what got recorded. Memory is what can be recalled in context and used to decide. Most companies have the first. Few have the second.⁹

Why it became necessary

People are good at working around missing context. You ask the person next to you. You remember the exception from last quarter. You know which Slack channel is actually true.

An agent does none of that. It only knows what you hand it, and it forgets between sessions unless something remembers on its behalf. Slite’s 2026 guide is exact on this: context sprawl was a nuisance; agents turned it into a dependency.⁴

Vectorize, writing for engineering teams, calls the same failure an organizational context problem, not a model problem. GPT-class models can reason. They cannot know that your prod environment is named us-west-prod-3, that deploys need the release coordinator, or that last quarter’s outage changed the runbook. Without a shared, enforceable memory, agents fail in production for reasons that look like “the model is dumb.”⁸

Tessera, covering Hyper’s “self-driving company brain,” states the strategic version: the bottleneck in agentic work has shifted. Models are getting smarter. They still arrive knowing nothing about your customers, your pricing history, or the sales motion that actually closes. The model is a commodity. Context is the moat.¹⁰

Princeton’s CoALA paper (2023) already named why a single chat log is not enough. Language agents need more than the current prompt: working memory, episodic traces, semantic facts, and procedural skill. A company that only stores documents has one of those layers. A company brain has to hold all four — or the agent will retrieve a page and still miss the commitment, the exception, or the procedure.¹⁴¹¹

What has to be true

The 2026 guides converge on a short list. If a product misses these, it is search or a wiki wearing a new name.⁴¹¹⁸

• It gathers from the tools the company already uses — chat, tickets, meetings, code, mail — not only from a handbook someone remembered to update.⁹
• It stores decisions and the why, not only the final object in the system of record.²
• It stays current. Stale pages are how wikis die. A brain has to notice when a fact changes.⁴
• It respects who may see what. A personal second brain has one owner. A company brain has roles, projects, and agents with different scope.⁴
• It is usable by people and by agents in the same memory — including through standard context pipes such as Anthropic’s Model Context Protocol.¹³⁹
• Writes back to tools wait for a human when the action is real. Precedent without control is just a faster mistake.

What Morrow takes from this

Morrow is built as that layer for teams and agents. Meetings, tickets, and chats become this morning’s report: who owns what, what slipped, what the company already decided. Memory is organized the way a company works — people, projects, decisions, tickets — not as a junk drawer of embeddings.

Local-first is the default. Cloud is optional. Mor and the other agents propose; nothing is written to a plan or an integration until someone confirms. That is the control Foundation Capital implied when they said decision traces must be captured in the execution path — and the governance Slite said a company brain needs once agents can act.²⁴

The trend is not a slogan. McKinsey already priced the search tax. Foundation Capital named the missing object: the decision. YC asked for the product. Agents made the workaround impossible. A company that wants agents to do real work needs a brain they can both read — and a human still on the write.