AI Engineer World's Fair 2026
From Systems of Record to Systems of Context
Loading slides…
Read the talk
From Systems of Record to Systems of Context
Sidekick’s architecture combines a durable model of how someone works with fresh signals about their day, preparing context before an agent is asked to prioritize.
From a talk by Omri Bruchim and Tomer Ast
What should I focus on right now?
Ask an assistant what to focus on right now, and it may return a confident list with little connection to your work. Gemini, GPT and Claude can all be put to this test. In Omri Bruchim’s anecdote, Claude suggested going to the gym—a memorable example of advice that missed the workplace question.
Connecting the assistant to boards, tasks, email and Slack does not automatically solve that problem. It can have access to the records without knowing how they fit together or what they imply for your priorities.
Access to information is different from understanding its significance. That distinction is the starting point for monday.com’s architecture: the missing step is interpreting the available evidence, not merely retrieving more of it. Here, understanding means being able to connect work to the person, relationships and priorities that give it meaning.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
An assistant where the work happens
Bruchim and Tomer Ast introduce themselves as engineering managers at monday.com, where projects, tasks, decisions, meeting notes and action items accumulate as teams work. The product ambition extends beyond storing those records: a sales team might use an SDR agent to contact prospects, while finance or marketing teams might need help conducting research. The assistant has to contribute to the business outcome.
The platform divides that ambition among four offerings:
| Offering | Intended role |
|---|---|
| monday sidekick | Assist with the user’s work |
| monday vibe | Build software |
| monday agents | Create agents within the platform |
| monday workflows | Run more deterministic flows |
Sidekick is the focus here: a personal assistant intended to understand the user and their business, think and execute alongside them, and adapt to their working style while keeping them in control. Those promises make context an architectural requirement.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Why meaning must be built before the request
Putting one assistant above Slack, meeting notes and the rest of the workplace creates a wall of records. The first difficulty is the agent gap: an agent can execute a clearly specified task without knowing which task deserves attention. Ask it to draft a reply to a particular customer escalation and it has a concrete problem to solve. Ask what to focus on first and it must also understand who you are and what matters to you. Remembering something about you does not necessarily establish those priorities.
The second difficulty is that a record rarely contains its full meaning. Consider recovering the reason for a line of code:
- Read the line and any nearby comment.
- Use
git blameto locate the commit and inspect its message. - Follow the commit to the pull request and read its description.
- Follow the pull request’s connection to a monday.com item.
- Discover that the work originated in a customer complaint.
The explanation emerges from the connections between records. The line tells you what the software does; the chain can explain why someone changed it.
The third difficulty is timing. Reconstructing that meaning only when someone asks a question leaves too much work for request time. The proposed Monday world model builds understanding in advance: who the user is, why a fact matters, when to help and what not to do. It is a model of the work and its relationships, rather than simply a larger prompt or a longer context window.
In this diagnosis, provider connections and MCPs already make the data accessible. The remaining challenge is interpreting how entities connect. That scopes the claim about retrieval: obtaining records alone is insufficient; it does not mean the architecture eliminates retrieval.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Three things the agent can reason over
Ast describes collecting thousands of data points about a user across items, status changes, activity logs, messages and meetings. The architecture turns those observations into three kinds of context:
- Work structure: Key entities and their relationships—what depends on what, how a Slack message connects to a task, and who is blocking whom.
- Current snapshot: Live signals over those entities—overdue work, critical urgency, and the coworkers the user is actively working with, including why.
- Durable profile: What decisions, outcomes, work patterns and recurring cadences reveal about the user over time.
These outputs serve different reasoning needs. Relationships explain the structure of the work, signals describe its current condition, and the profile gives the agent a persistent account of the person doing it.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
One engine learns the person; another follows the day
Two engines construct this context using different time windows and schedules. The slow engine examines weeks of activity to identify the user’s role, routines, work rhythm, collaborators, goals and current projects. It distills those patterns into a durable profile. When a pattern continues to hold, the engine reinforces it, gradually refining its account of how the user works.
The fast engine takes a short recent window and frequently recomputes live signals: what is overdue, what has suddenly become urgent, and which coworkers have become involved.
| Engine | Evidence window | Output |
|---|---|---|
| Slow | Weeks of activity | Durable profile and recurring patterns |
| Fast | Short recent window | Current state and live signals |
The split separates relatively stable knowledge about the person from changes that may alter today’s priorities. A durable routine and a newly urgent obligation both matter, but they should not be treated as the same kind of evidence.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Rapid updates and gradual consolidation
Ast connects this temporal split to two precedents. The first is complementary learning systems: rapid learning associated with the hippocampus complements slower consolidation in the neocortex. For Sidekick, the useful inspiration is the coexistence of recent experience and durable learning, not a claim that its engines reproduce the brain.
The second is Lambda architecture. A batch layer recomputes views over historical data, while a speed layer handles recent data not yet incorporated into those results. Queries combine the two into a served view. Sidekick applies a related idea to its context model: combine slower processing with a faster account of recent changes. The profile-versus-signals division is an architectural analogy, not an assertion that the implementations are identical.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Precompute context, then refresh selectively
The serving path connects those engines to the assistant:
- Collect workplace data from monday.com, Slack, email and calendar.
- Turn the data into structures, signals and patterns.
- Run both engines offline, before the user engages with Sidekick.
- At engagement time, recompute a thin slice of logic for recent activity and serve the resulting context to the agent.
- Let Sidekick decide when and how to traverse the data model and retrieve further context.
Precomputation prepares the agent to reason; selective retrieval still helps it investigate. The request does not have to initiate the entire process of interpreting the user’s work.
This design also separates freshness from availability. Ast describes isolating sources so that a bad feed does not break the others. At serving time, the thin logic layer verifies part of the context against live data; elsewhere, the system retains the last verified context. The intended behavior is graceful degradation, with some information remaining older rather than the whole context becoming unavailable. This is a resilience design claim, not a measured availability guarantee.
Urgency is another intended consequence of the model. If the assistant can interpret the significance of a fact, it has a basis for deciding whether to notify the user proactively or remain silent. The goal is not simply to detect more activity, but to recognize when that activity merits an interruption.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A growing model still trails reality
Continued observation is intended to make the context more useful: daily capture fills in the layers, sharpens the profile and produces a model specific to the individual’s work. Ast describes adding a source as deliberately inexpensive and additive, extending the evidence available to the engines.
The architecture nevertheless has three explicit limitations:
- Lag: The represented world always trails what is actually happening.
- Cold start: New users do not yet have enough reliable history to support the same reasoning.
- Bias and noise: Signal definitions encode their designers’ assumptions, and distinguishing important evidence from noise remains difficult.
The proposed advantage is an architecture that can be enriched, tested and improved over time—not a claim that these problems have been solved.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Turning Omri’s breadcrumbs into today’s context
Return to the opening question: what should Omri focus on right now? The worked example begins with what Bruchim calls breadcrumbs: his current board, the past month’s emails, meeting transcripts and the action items from those meetings. Offline calendar processing adds recurring patterns. A regular meeting with VPs, for example, helps establish the significance of another meeting tomorrow. Recent Slack messages can also contain relevant statements that never appeared in the team’s daily stand-up.
The slow engine turns that history into a profile: Omri Bruchim is an engineering manager in Tel Aviv, working on Sidekick and Notetaker, with particular working hours and daily routines. The fast engine supplies obligations from the past day or few days. In the example, those include three commitments to other people and an email from a VP that he has not answered.
The two outputs contribute different parts of a prioritization answer. The profile establishes whose work is being prioritized; the recent signals identify pending obligations that matter today. The example illustrates the evidence available to Sidekick, rather than a measured prioritization result or completed action. The commitments remain commitments, and the unanswered email remains something to address.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Connect the work before asking the agent to act
For Sidekick, the proposed architectural shift is to process relationships and user context before the agent needs them. A capable model with access to records still needs a basis for understanding how those records connect to this person’s work. The Monday world model supplies that basis in advance, so a question about priorities can begin with an account of the user’s situation instead of a fresh pile of records.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Instructions for opening Sidekick, connecting work apps, using contextual assistance and managing personalization and memories.
Further reading
A guide to finding Sidekick, Vibe, agents, workflows and Notetaker within monday.com.
The foundational account of rapid hippocampal learning and gradual neocortical consolidation through interleaved experience.
Explains how precomputed historical views and a realtime layer covering recent data combine to answer queries.
Updates since the talk
Omri Bruchim describes Sidekick's architecture for permission-aware retrieval, specialized subagents, bounded tools, sandboxed execution and evaluation.
Full transcript
Loading transcript…