The Understanding Engine.
The World Model holds what your work says. The Understanding Engine turns it into something useful: a maintained picture of what matters now, and, for every request, a short, scoped, sourced read that an agent can act on.
Meaning, built while nobody is asking.
Beyond individual statements, the engine maintains structure that makes the graph readable.
- Moments
- Conversations and meetings become titled episodes you can open, with who took part and what came out of them.
- Storylines
- Related statements across sources are grouped into matters that move over time. A storyline only counts as moving when its evidence re-verifies and it has recent activity.
- Who matters
- Person importance is scored deterministically from several components, so the people you actually work with rise above everyone who ever emailed you.
- Tie strength
- Reciprocity and frequency distinguish a real relationship from a one-way stream of messages.
- Embeddings
- Statements are embedded for meaning-level search, inside our own cloud account.
- What you care about
- Nine kinds of explicit action (pinning, confirming, editing, acting on, dismissing and more) are folded nightly into ranking weights with a ninety-day half-life. Corrections fade; they never scar.
One read, from question to packet.
Every read follows the same path, whether it serves your own workspace or a connected agent.
- 01ResolveWork out which people, companies, matters and dates the request names, and who is asking.
- 02ScopeCompile the asker’s scope into the query, so nothing outside it can be retrieved.
- 03GatherDraw candidates from eight independent pools at once.
- 04Rank and deduplicateScore by recency, confidence, your learned preferences and person importance. Each item appears once, with how it was found.
- 05Deliver and recordReturn a short packet with sources and statuses, and keep the exact request and response.
anchor statements about the records the question names neighbourhood one hop out along typed relationships lexical tiers exact and near-exact wording keyword trigrams partial words, codes, product names typo-tolerant fallback misspelled names still land recall window what happened around the dates asked about recency what is newest across the workspace semantic meaning-level similarity from embeddings
What an agent actually receives.
An agent read answers the request that was made, not a dump of the workspace.
A question about a person returns that person, what is open with them and what changed. A question about a day returns that day’s schedule from the synced calendar, plus anything dated in it. Times carry their time-zone offset, so a meeting at 11:00 in Jerusalem is never read as 08:00.
Statements that were heard but not confirmed are included and labelled as such. Broad requests get the people who matter now; narrow ones don’t. Every read is frozen with an identifier, so it can be replayed later under the agent’s current access, and any statement in it can be opened to its exact evidence.
request "What's open with Northpeak before Friday?"
resolved Northpeak · Mara Levin · to Fri 2 Oct (+03:00)
current Mara owns the renewal trusted · email Wed
Limited June pilot agreed trusted · your email Tue
open Renewal terms due Fri promised by Daniel · meeting Mon
heard "Procurement signed off" unconfirmed · meeting Tue
unknown Who sends the terms now
read id frozen · replayable · 3 sourcesA current picture, kept current.
Alongside individual reads, the engine maintains an overview of what matters now: the situations, promises and changes worth your attention.
- Refreshed where the change landed
- A new statement reopens only the parts of the overview that cited what changed. A meeting tomorrow reopens its own concern and nothing else.
- Bounded by design
- Full rebuilds are capped, and focused updates are rate-limited, so cost follows how much your work changed rather than a timer.
- Honest when it can’t
- If a refresh can’t run, the reason is recorded and the previous overview stays. A failed attempt never overwrites a good one.
Measured on real questions.
Numbers from our own workspaces while building. They describe specific changes on specific data, not a benchmark.
- 21.8K → 5.9K characters
- Average packet per agent read on 23 real questions, after reads were built from what the request names. “Mostly noise” answers fell from 21 to 5.
- 7–19K → ~1.1K characters
- “What do I have tomorrow?” style questions, which now return the day’s events, including a meeting the older read missed.
- 5 of 8 overviews
- Live overviews that could refresh without a full rebuild after the change-scoped update: one needed no model call, four updated only the touched items.
- OverviewThe World Model
- Data modelStatements, two clocks, belief lanes
- Entity resolutionOne person, one record, carefully
- Truth maintenanceHow beliefs are retired, never deleted
- Trust and scopeWho said it, who may see it
- Memory vs understandingWhat a memory layer does, and what understanding adds
- The 31 problemsEverything you have to solve, in one list