From Systems of Records to Systems of Context
Tomer Ast
Omri Bruchim
What Should I
Focus on Right Now?
Your AI assistant has every board, task, email, and message you've ever touched.
And it still can't really answer.
Hey, do you remember to send me the slide for the meeting tmw morning?
03
The Challenge
Not missing data.
Missing understanding!
Omri Bruchim
Tomer Ast
monday.com is a global software company that helps customers scale business impact and expand what teams can deliver.
Our mission has always been to help teams achieve their business outcomes
AI work capabilities
Your intelligent AI personal assistant that understands your work, thinks and executes with you
Knows you and your business
Accelerates work at every level
Works the way you do
Keeps you in total control
Delivers answers you can trust
One assistant —
all of your work.
10Sharp at doing tasks.
Lost at finding them.
Point an agent at a problem and it's brilliant. But “what should I focus on?” isn't a problem to solve — it's asking the agent to find what the problems even are.
“What should I work on first?” → guesses.
11A log says what happened.
Never what it means.
A record is like a line of code someone wrote months ago. It tells you what changed — never why. To recover the meaning you trace it back, one hop at a time.
The why lives a hop away, every time:
git blame → the commit message → the PR description → the monday task that asked for it → the customer who complained in the first place.
12At runtime,
it's too late.
Turning records into meaning takes layers of processing — signals, patterns, connecting the dots. An LLM can't reconstruct your whole work history the instant you ask.
13Not a bigger prompt.
The 
world model.
Context that follows your work: who you are, what you're working on, what needs action — a model the agent can reason over to answer what it can't derive on its own.
14It isn't retrieval.
The problem was never getting the data — hand an agent every board, every doc, every message, and it's still lost. So the model isn't a smarter search over your work. It's a working understanding of it.
15Three faculties of understanding.
Items, people, meetings, boards — and the relationships that carry the meaning: what blocks what, who's waiting on whom, where a thing got decided.
Live signals read off your current state: what's overdue, what's piling into a crunch, who you're suddenly in the loop with.
Episodes, decisions and their outcomes — distilled into durable patterns: how you actually work, not how a job description says you do.
16Two engines, on different clocks.
17Two engines, on different clocks.
18Two engines, on different clocks.
19This split isn't ours.
Learning Systems
Architecture
Two different fields, the same shape. We applied it to the one place no one had:
the world model.
20The brain already runs slow + fast.
Learning Systems
21So does the data stack.
answer
Architecture
22Heavy below, live on top.
A thin live slice — the agent reads a model that's already built.
23Behavior we never had to bolt on.
It degrades —
it doesn't fail.
"But precomputed means stale." So it's built to bend, not break.
It knows when to stay quiet.
The agents that don't understand can't tell — so they spam you. Restraint isn't a feature we add; it falls out of understanding.
In the model it's just another signal: worth interrupting for, or leave them alone.
24world model
And it compounds.
Every day you work, the layers get richer and the profile gets sharper. Adding a source is deliberately cheap — so the surface only grows.
The more it sees, the more it understands. The more it understands, the more you lean on it.
25The hard parts.
Eventual consistency — the model is always a little behind the live world. A feature, but you design for it.
Cold start — new users, before there's anything to learn from.
Signal from noise — the hardest one, and the same one underneath the whole talk. That's the work.
26What should I focus on right now? Sidekick can answer that now.
Not because the model got smarter — because it finally has a model of your work.
Omri Bruchim · Eng Group Lead, Sidekick & Notetaker (Tel Aviv) — ~5 hrs of meetings/day, 0.39 focus ratio, 104 open action items.
- 3 commitments past due — oldest from the Acme sync (9 days ago): send the integration spec and confirm the pilot scope.
- Reply to your VP about the pivot — the Q3 roadmap draft is blocking 4 people.
- Protect focus tomorrow — 6 back-to-back meetings; move or decline 2.
27
The bottleneck was never raw capability.
It was understanding.
The most capable agent in the world, pointed at your work without it, is still just guessing.
28
The world model.
That's what we're building — the understanding. Sidekick won't just answer your questions — it'll understand your work.
It doesn't just answer you… it knows you.
29