Agents need a brain they can write to

Karpathy was right

In April, Andrej Karpathy crystallized something a lot of people had been circling. An LLM-maintained markdown wiki, he wrote, is a “persistent, compounding artifact.” The readers who picked up his thread reached for the same metaphor: RAG re-derives everything per query like an interpreter; a wiki compiles the synthesis once, persists it, and reuses it. The post hit sixteen million views.

A wave of personal-brain setups followed. Obsidian as the canonical version, Mem as an AI-native sibling that auto-organizes notes, and a long tail of similar tools. GitHub shipped a repo-based agent memory system for Copilot. Tiago Forte declared that personal context management was eating personal knowledge management. YC’s Spring 2026 RFS called out the missing primitive by name: a “company brain” with access controls, audit, and write access for the agents who ought to be maintaining it.

The pattern is single-player

Obsidian-as-second-brain works great if you’re the only one tending the garden. Some teams do run a shared Git repo as a company-wide brain (we have customers who do exactly this), but it asks every contributor to be fluent in pull requests and merge conflicts, and the editing tools are no match for a real wiki. The patterns mostly assume one curator, one agent, one knowledge graph. That’s fine for individual research. It collapses the moment work involves anyone else.

Your teammate needs to find the fact you wrote down. The agent learns something during a run and has nowhere to record it. A customer thread needs a price that nobody owns the source of truth for. You want the answer to “what is our position on X” to be the same whether a person or an agent asks it.

The personal-brain patterns aren’t wrong. They’re under-scoped. The thing that makes a knowledge base useful for one person, markdown and hierarchy and linking, is not the thing that makes it useful for a team. That requires identity, permissions, multi-author edits, and a surface the agent can write to in real time.

What we shipped

The brain is a writable knowledge layer inside this+that. Every team gets a team space, every person gets a personal space, and pages nest hierarchically. Markdown editor with @-mentions for people, tasks, and other pages. Page-level versioning, conflict detection when two people or agents edit at once, attachment support. Every page and every version is signed by its author. When an agent writes a page, the history also records the teammate who delegated the work, so you can always see who authorized what.

The not-obvious part is that agents can write to it. Humans reading and writing a wiki is the baseline; every wiki since the 90s has supported that. What’s different about the brain is that agents are equal-class participants in both directions.

That gives you two patterns worth designing around:

Workflows that read the brain. An agent or workflow reads existing team knowledge to act. The canonical example is a sales agent: pricing questions arrive over email, the agent answers using the canonical price sheet maintained in the team brain, schedules a follow-up call if the prospect is interested, and logs the interaction as a task that links back to the source thread. The agent’s reply is grounded in what your team actually decided, not in what a base model guesses.

Workflows that write the brain. An agent or workflow produces durable knowledge over time. The canonical example is a competitive analysis agent: each run produces a page in the brain — what the agent found, what was new since last time, what changed. The next run reads the previous one as context. Over weeks the brain accumulates a history you can search instead of a one-shot report that fades.

These two patterns compose. A read-and-write workflow handles a teammate’s inbound question: the assistant searches your brain (only you see it), drafts a reply grounded in what you know, and on send, writes the answered question back into the brain so the next person to ask it can find the answer without bothering you.

Outward, when you want it

Brain content is private by default. But you can wrap a slice of it in a public artifact: an AI-authored page, landing page, presentation, or microsite with its own URL that anyone outside your team can visit and interact with. The page can query your brain in real time through a controlled action, so a customer asking a question gets an answer grounded in what your team has actually written down.

A public FAQ that’s never out of date because it reads from the same pages your team is editing. A landing page whose product details stay current. A microsite for a launch that compounds with the work the team is doing internally. You decide what gets exposed and through which surface. The default is private, the path outward is controlled and explicit, the underlying knowledge is the same.

And the public surface isn’t read-only. The same artifact can show your live calendar availability and let a visitor book a demo, so a launch page can answer questions from the brain and close a meeting without sending the visitor anywhere else. You choose the brain subtree the public artifact can answer from; the rest of the brain stays private.

And it goes the other direction too. Some teams already have a knowledge base that works for them: a Notion workspace, a Confluence, a shared Git repo. Our agents can write into those through MCP, the same way they read from them. The brain inside this+that is the default, but if your team has already settled on a different home for knowledge, our agents can keep that home up to date instead.

Knowledge, not transactional data

The brain is for the durable thinking your team does. The reason your pricing is the way it is. Your positioning relative to competitors. The company’s vision. The kind of thing people used to write down once and then forget to update.

It’s not a CRM. It’s not a marketing analytics product. Transactional data lives in the systems built for it (Salesforce, HubSpot, GA4, your data warehouse) and stays there. The brain is for the things that don’t fit cleanly into rows.

A cautious third pillar

We’ve been describing this+that as two things: task extraction (what needs doing) and workflows (what gets done automatically). We’re cautiously adding a third, the brain (what you and your agents know). Each one makes the others more useful. The brain grounds the workflows. The workflows put new knowledge in the brain. The tasks reference the brain for context.

They work as one product because the messages, the tasks, and the brain all live inside the same trust envelope. The agent drafting your reply has the permissions you have. The page the agent writes inherits the team’s access controls. Nothing leaks across boundaries because there aren’t any to leak across.

Karpathy nailed the why

His core point, that knowledge work compounds when it persists and the marginal cost of maintaining that persistence drops to zero with LLMs, is exactly right. The piece missing from the personal-brain framing is that real work is rarely a single-player game. Half the value of writing something down is so someone else, or an agent serving someone else, can find it next week.

We built the multiplayer version because that’s where work actually happens.