{ "video": { "id": "Btk8wDUVs74", "title": "From Systems of Record to Systems of Context — Omri Bruchim & Tomer Ast, monday.com", "duration": 958, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "Btk8wDUVs74", "title": "From Systems of Record to Systems of Context — Omri Bruchim & Tomer Ast, monday.com", "one_liner": "monday.com's engineering managers argue the bottleneck for work assistants was never retrieval but understanding, and show how they precompute a \"Monday world model\" offline with a slow engine that learns who you are and a fast engine that learns your day.", "summary": "Omri Bruchim and Tomer Ast describe shifting monday.com from a system of record — every task, message and status logged — to a system of context that understands how those records connect. Their framing: an assistant can have all your boards, emails and Slack messages and still answer \"what should I focus on right now?\" with disconnected bullets, because it has all the data and zero understanding. They present the Monday world model behind Sidekick: thousands of data points per user distilled into work structure, a live snapshot, and a durable learned profile, built by two engines on different time windows — a slow one over weeks and a fast one over the last day or two — precomputed offline, with only a thin slice recomputed at serve time.", "key_points": [ "The core distinction of the talk is understanding, not context, memory or retrieval: \"the problem was never the missing of data, the retrieval\" — they already have every provider connection and MCP they want.", "Three reasons this is hard: the \"agent gap\" (agents nail a defined task like drafting a reply to a customer escalation, but guess when asked what to prioritise), records carry no meaning, and meaning cannot be built at runtime the moment a question is asked.", "The meaning-from-records analogy: a line of code → git blame → the commit message → the PR description → the linked Monday item, which reveals the customer complaint that caused it. That chain is what they're trying to build.", "The data model constructs three things the agent can reason over: how the user's work is structured (entities, dependencies, who's blocking whom, how a Slack message connects to a task); a current snapshot of live signals (what's overdue, suddenly urgent, which coworkers you're pulled in with); and what's learned over time (decisions, outcomes, work patterns, cadences) distilled into a durable profile.", "Two engines on different schedules: a slow engine over weeks mining persona, routines, rhythm, collaborators, goals and projects — a profile is reinforced every time it holds — and a fast engine over a short recent window recomputing live signals.", "They did not invent the split: it mirrors complementary learning systems in neuroscience (hippocampus captures instantly, neocortex distills durable lessons) and lambda architecture in data processing (speed layer over a real-time window, batch layer over full history, merged into one served view).", "Both engines precompute offline from Monday, Slack, email and calendar; when a user engages Sidekick only a thin slice of logic is recomputed for recent activity, then Sidekick decides when and how to traverse and retrieve from the data model.", "Two properties fall out of the architecture: resilience — sources are isolated so a bad feed can't break the rest, and the serve-time layer verifies part of the context against live data while the rest falls back to the last verified context — and urgency awareness, knowing when to be proactive versus stay silent.", "Stated limitations: the model always trails the live world, new users have no reliable data to reason from, signals carry built-in biases, and the hardest part is telling the important parts from the noise.", "Demo: \"breadcrumbs\" (current boards, months of email, meeting transcripts and action items, full calendar, recent Slack) produce a slow profile — engineering manager, two projects Sidekick and a notetaker, based in Tel Aviv, hours per day — and a fast list: three commitments promised to other people and an unanswered VP email." ], "takeaways": [ "Stop treating \"what should I focus on?\" failures as a retrieval problem — adding more connectors, MCPs or a longer context window won't fix an agent that doesn't understand the relationships between records.", "Build understanding ahead of time, offline. Anything that has to be derived at the moment the user asks the question is already too late; keep serve-time to a thin recompute over recent activity.", "Split your context pipeline by time window rather than by source: a slow batch engine that learns the person over weeks, and a fast engine that recomputes live signals over the last day — one knows you, one knows your day.", "Isolate data sources and make the serve path degrade gracefully — verify a portion of context against live data and fall back to the last verified context for the rest, so a bad feed never takes the assistant down.", "Make adding a new data source cheap and purely additive, so the model compounds: more sources sharpen the profile instead of requiring rework." ], "topics": [ "agents", "context-engineering", "memory", "personal-assistants", "data-architecture", "lambda-architecture", "productivity-tools" ], "tools": [ "monday.com", "Monday Sidekick", "Monday Vibe", "Monday Agents", "Monday Workflows", "Monday world model", "Claude", "Gemini", "GPT", "Slack", "GitHub", "MCP" ], "quotes": [ { "text": "It has all the data. Um, but it has zero understanding.", "at": "01:58", "url": "https://www.youtube.com/watch?v=Btk8wDUVs74&t=118s" }, { "text": "Actually, I tested last week and uh Cloud asked me to go to the gym. I don't know if it's a compliment or not, but uh this is what he suggested.", "at": "01:31", "url": "https://www.youtube.com/watch?v=Btk8wDUVs74&t=91s" }, { "text": "So understanding, understanding of the context has to be ahead of time. Um, you need to build it much before someone asks the question.", "at": "07:02", "url": "https://www.youtube.com/watch?v=Btk8wDUVs74&t=422s" }, { "text": "One knows you and the other one knows your day.", "at": "09:22", "url": "https://www.youtube.com/watch?v=Btk8wDUVs74&t=562s" } ], "words": 2877 }, "summary_url": "/#Btk8wDUVs74", "transcript": { "html": "/transcripts/Btk8wDUVs74.html", "txt": "/transcripts/Btk8wDUVs74.txt", "vtt": "/transcripts/Btk8wDUVs74.vtt" } }