Sketch is the
second brain of your company.
That's a big claim, and you should interrogate it. So this page is the interrogation: you ask how the thing is actually built, Sketch answers — and behind this conversation, a brain is forming. Watch the background.
ok — "second brain." what does that actually mean?
Known means: every person, client, project, and promise exists as a record a machine can reason over — so your AI never starts from zero. The term is 90 years old and almost everyone who used it built the cabinet, not the brain. The full history →
UNDER THE HOOD · a context graph + timelines + a deterministic trigger layer + a written rulebookbut the intelligence already exists. what did YOU build?
The model is Claude — identical for every company on earth, so it can't be anyone's edge. The engineering problem is everything around it: a memory that fills itself, reflexes that fire without prompting, and rules learned from real mistakes. That's the build. By the end of this page, you'll see the whole brain assembled.
UNDER THE HOOD · rented frontier model · the graph, triggers, and rulebook are the proprietary layera memory that "fills itself"? every wiki I've owned died because nobody filed.
IT FILES ITSELF · OR GRAB A CARD AND THROW IT
Each record is captured as-is, from the tool that owns it — chat, meetings, boards, CRM, docs. The SharePoint-era wikis died because staying current cost human filing effort. That constraint is gone: the librarian reads everything itself, so the catalogue stops rotting when people stop filing.
UNDER THE HOOD · five structured source families, read continuously — nobody files anythingso where do those filed records GO? a database? a pile of embeddings?
Not a pile of documents — a map of your company. One card per real thing: a person, a client, a project, a promise. And the cards hold hands: Meera works at Juniper, Juniper owns the redesign, the redesign carries a promise. A question walks the edges and reads all your tools at once.
- Entities, not documents. The unit of memory is the real-world thing, not the file it was mentioned in.
- Edges are the answers. "Where does Juniper stand?" is a walk, not a search.
- The graph is inspectable. Open any card, see every fact, see every receipt.
and when it reads wrong? AI is famous for making things up.
Our early pipeline extracted freely. It invented a company called "Gmail" — 57 people had gmail addresses — and gave a 3-product client 203 products. The rebuild is the actual technology: facts enter only when a structured tool vouches for them, the model is spent only on what no tool wrote down, and every fact keeps its receipt. Ask any answer "says who?" and it points to the exact message.
UNDER THE HOOD · structured-source vouching · corroboration gates · per-fact provenance · extraction prompt on version fourteenmy client is @meera on slack, a bare +91 number on whatsapp, an email address. how do you know that's one human?
To every tool you own those are three strangers. The graph cross-references the evidence — shared threads, signatures, who replies to whom — and resolves one person. From then on every message, call, and promise lands on her one card. And it is careful by design: two similar names stay two people, because one wrong merge poisons every answer about both.
UNDER THE HOOD · entity resolution across channel identifiers · the matcher is forbidden from trusting reordered namesfine, people. but projects? my teams call the same project three different things.
The words never match — the cast does. When the same people keep showing up around a thing, it's the same thing. Text similarity can't see that; the room can. Here is the real one it did to us:
Five spellings, one node. Nobody told it. So you can ask in your words — "the pricing thing" — and land on the right project with its whole history.
UNDER THE HOOD · de-duplication by social fingerprint, where embedding similarity failsa knowledge base is a snapshot. snapshots rot. why won't yours?
Identity lives in the graph and rarely changes. Status lives on timelines clipped to each card and never stops. The map never shakes; the news never goes stale. Answers are about now.
UNDER THE HOOD · identity in the graph, time on timelines — principle 2 of the context-graph note
Nature solved this design long ago: reflexes don't route through the expensive brain. Your blink never consults your genius. Sketch runs the same split — and this is why the automations actually work at company scale:
Fires on triggers, not prompts. Runs all day on cheap intelligence, because shelving doesn't need a genius.
When a question deserves the best brain money can rent, that brain sits down at the same catalogue and reads.
And every mistake either lane ever made is now a written rule the system cannot break again — about 80 rules, each with a name, a reason, and the incident that created it. Learning you can open and read.
UNDER THE HOOD · deterministic trigger layer + a legislated rulebook — genius is for reading, not shelvingok. prove the whole stack works. one real story.
Our founder thought the decision through with a frontier model. First it gave good general counsel — knowledge of the world, rentable by anyone. Then it sat down at Sketch's reading desk and pulled our complete history with one prospect: the first demo call in March, a WhatsApp half-promise from ten days earlier, an open pull request — in under a minute. The advice stopped being general wisdom and became: you made this person a commitment, address it. Same model, same conversation. It didn't get smarter. It got informed — by the part of the brain we had to build, because nobody would rent it to us.
UNDER THE HOOD · the full story →A brain was forming.
Here it is.
Every card, person, promise, and project from this conversation — filed, merged, de-duplicated, and connected. This is the neural map of one small week. Yours will hold thousands.
last question. is this brain locked inside Sketch?
Think of Sketch as middleware: the live context manager of your entire company. Its own agent uses the graph all day in Slack and WhatsApp. But the graph speaks MCP — so any frontier brain can sit down at the same reading desk. Claude in your IDE. Codex in your pipeline. The agents in your workflows. Every one of them stops starting from zero, because the context is managed once, in the middle.
Models get replaced every year. The map outlives every model that reads it — and every answer, from any brain, gets sharper the day a better model ships.
UNDER THE HOOD · the graph is served over MCP — one catalogue, many readersRented mind. Built map.
Wired reflexes. Earned rules.
One brain, in the middle of everything — you just interviewed it. Built from records you already keep. Every answer carries receipts. The research is published: the company-brain note · the context-graph note.
hire sketch