The World Model.
Team0 is not a second brain you have to feed, or a prompt with a memory bolted on. Underneath is the World Model, a typed graph of everything your work says, and the Understanding Engine, which keeps a current view of it. This is how they work, in depth.
From a vocabulary to a living graph.
Knowledge engineering builds meaning in layers, from a controlled vocabulary up to a knowledge graph. Team0 implements each layer, then adds the four a working understanding needs and a static graph lacks: time, belief, maintenance and permission.
- 01Controlled vocabularyAgreed words for things and relationships.A registry of entity types and typed relationships, each with a cardinality and a category. An unknown relationship never enters as belief; it waits as a candidate.
- 02MetadataStandard fields on every record.Every statement carries its source, a reference to the exact message or meeting, a confidence, and the passage that supports it.
- 03TaxonomyConcepts arranged by kind.People, organisations, projects, deals, topics, threads and events. Each relationship category sets how long a statement stays relevant.
- 04ThesaurusAlternate names for the same thing.Aliases, several addresses per person, reviewed relationship synonyms, and the owner’s own verified addresses.
- 05OntologyRules about what things mean.Single-valued relationships supersede; events never replace events; a request is not a commitment.
- 06Knowledge graphThe connected facts.One entity-and-statement graph per workspace, isolated in its own database schema.
- 07TimeAdded by Team0.Two clocks on every statement: when it became true, and when Team0 learned it.
- 08BeliefAdded by Team0.Trusted, evidence and candidate lanes that cannot leak into each other, plus who holds each view.
- 09MaintenanceAdded by Team0.Beliefs are retired, superseded and corrected continuously, with every change typed, attributed and reversible.
- 10PermissionAdded by Team0.Who is asking decides what is retrieved, before retrieval. Nothing is redacted afterwards.
Where Team0 differs.
Most memory layers and second-brain tools store what they are told and retrieve what looks similar. These are the design decisions that make Team0 different, each explained in depth.
- 01
Evidence cannot become belief by accident
A message that mentions something is stored as evidence. It can be searched and cited, but deduplication or repetition can never promote it to a trusted statement.
Read more - 02
Nothing is ever deleted
Every retirement is a typed, attributed verdict that can be reversed, and it survives the same email being read again next cycle.
Read more - 03
Supersession keys on identity, never on wording
A newer statement replaces an older one only through a single-valued relationship or a typed attribute of the same record. Similar text is never enough.
Read more - 04
The model proposes; code decides
Where a language model judges, it picks from a closed set of verdicts and code applies them. Model confidence is never the deciding signal.
Read more - 05
Similar names never merge on their own
Matching addresses merge records. A similar name needs a model check, or one clear first-name case, before two records become one, and every merge can be undone.
Read more - 06
A second vendor checks the links
Relationships read from prose are re-checked every night against their own words. When Team0’s model and a model from another vendor both find no support, the link stops counting as fact.
Read more - 07
Truth and attention are separate questions
Whether something is supported, where it came from, how it relates to you, and whether you have seen it are four independent answers.
Read more - 08
Prevention, not redaction
Who is asking is declared, never inferred. Their scope is compiled into the query itself, so a model never receives what it may not say.
Read more
From raw work to a read an agent can trust.
Six stages, each with one job. Models are used where language needs understanding; everything that can be deterministic is deterministic.
- 01SourcesEmail, calendar, meetings, documents, work tools, and what connected agents save back.
- 02ExtractionEach item is broken into typed statements: who said it, to whom, when, the exact supporting passage, and where it came from.
- 03Trust tiersEvery statement lands as trusted, evidence, or candidate. Nothing an agent or a model says is trusted automatically.
- 04The graphOne entity-and-fact graph per workspace: people, companies, matters and events, joined by typed relationships.
- 05UnderstandingA maintained view of what is current, what changed, what is open and what is still unknown, refreshed where the change landed.
- 06ReadsEach request gets a scoped, sourced answer built from the records and dates it names.
The whole system, in one diagram.
Select any part to see what it does. Sources are consolidated into the graph, maintained continuously, and read under the asker’s scope.
System atlas
One living model. Two cognition loops. Every permitted surface shares the result.
Loop 01 · Continuous consolidation
Reality changes. Team0 changes what it believes.
This loop keeps working after the conversation ends: recognizing events, grounding claims and maintaining meaning before anyone asks.
Always working · source-aware
- ConversationsEmail threads · replies · attachments
- Time togetherMeetings · calls · transcripts
- Work in motionCalendar · tasks · projects · CRM · GitHub
- Connected worldDocuments · tools · payments · public web · your agents
Observe and establish authority
Turn activity into grounded evidence.
- 01RecognizeFind the meaningful event inside the activity
- 02BindResolve who, what, authorship and relationships
- 03GroundAttach source, evidence and both kinds of time
- 04PlaceKeep beliefs, evidence and candidates in separate trust lanes
Durable, grounded knowledge
The living World Model
People, organizations, projects, events, decisions and commitments remain connected as the world changes.
Maintain and derive
Keep the model coherent without erasing history.
- 01ConnectForm episodes, storylines and meaningful relationships
- 02ReconcileConfirm, challenge, supersede or settle from evidence
- 03PreserveRetain history without carrying stale beliefs forward
- 04Derive and retractMaintain conclusions as their evidence changes
This is a loop, not an import: new evidence can confirm, contest, supersede, settle or retract what the model currently carries.
Loop 02 · Just-in-time understanding
The maintained world becomes the situation now.
The World Model supplies durable truth. Fresh Awareness, purpose and permissions determine what it means, and what deserves attention, right now.
Govern the request
Know what may be understood now.
- Purpose and scope
- What is being asked, what it actually needs, and what may be read
- Live Awareness
- Your synced calendar, inbox, meetings and tasks, with a live read when the copy isn't enough
- Channel and authority
- Identity, permissions, voice and action limits
One shared read
Retrieve the relevant world once.
Exact, semantic, temporal and structural retrieval resolve into one frozen evidence set, not a second memory system.
- 01Current truth
- 02Open commitments and asks
- 03Provenance neighbours
- 04Recent relevant context
- 05Explicit evidence and annotations
Understanding Engine
Build one Current Situation.
- Truth
- Supported · contested · unknown
- Work
- Actionable · settled · waiting
- Coverage
- Fresh · partial · stale · unavailable
- Reasoning
- Evidence, identity and derivation remain attached
Executive attention
Decide what the moment deserves.
Truth does not automatically become an interruption. Utility, timing, channel and action authority remain separate decisions.
- Say
- Act
- Ask
- Defer
- Stay silent
One understanding · every agent you connect
Your agents turn it into work, without fragmenting it.
Same situation · any agent
- Every agent you connectStarts from the same understanding
- Meeting preparationHistory, people and open work
- Picking up workContinue in any agent where another stopped
- Building and researchGoals and constraints carried in
- Before actingWhat changed, checked first
- Each agent’s own accessOnly what you allowed it to see
- Replies and corrections
- Delivery and use
- Completed actions
- Failed actions
- Outcomes, dismissals and restraint
Statements, not summaries.
The World Model stores individual statements, each tied to its subject, the records it concerns, who said it, the passage that supports it, and two clocks: when it became true and when Team0 learned it.
- subject
- Mara Levin
- The record the statement is about.
- predicate
- owns
- A typed relationship from a controlled vocabulary.
- object
- Northpeak renewal
- Another record, which makes this an edge in the graph, or a plain value.
- concerns
- Mara · Daniel · Northpeak
- Every record the statement touches, so it is found from any of them.
- holder
- Daniel, in an email to you
- Who said it. A claim someone made is not the same as a fact.
- evidence
- “Mara is taking Northpeak from here.”
- The exact supporting passage, kept with its source.
- valid from
- Wednesday
- When it became true in the world.
- learned
- Wednesday 8:14
- When Team0 learned it. The two clocks are kept separately.
- status
- trusted · current
- Trusted, evidence or candidate; current or superseded.
- supersedes
- Daniel owns Northpeak renewal
- The older statement, kept in history rather than overwritten.
- Why two clocks
- A meeting on Monday can be learned on Wednesday. Keeping both lets Team0 answer “what was true then” and “what did we know then” separately.
- Why nothing is overwritten
- When ownership moves from Daniel to Mara, the old statement is superseded and kept. History is part of the understanding.
- Why typed relationships
- A controlled vocabulary of relationships means a new source adds statements, not new tables. The model extends without changing shape.
One person, one record, carefully.
Getting identity wrong poisons everything downstream, so merging people is the most conservative part of the system.
- Deterministic keys first
- The same email address is the same person; the same domain is the same company. Mail, meetings and calendars about one person collapse into one record.
- Your own addresses
- Addresses Gmail has verified as yours are treated as you. A look-alike address is asked about once, never assumed.
- Similar names are checked
- Similar names never merge on their own. A merge needs a matching address, a model check that the two are the same person, or one clear case: a lone first name and the only person with that name you deal with. Every merge can be undone.
- A second opinion on links
- Relationships read from prose, like who works where or who introduced whom, are re-checked every night. If Team0’s model and a model from another vendor both find the text does not support a link, it stops counting as fact.
Rules the understanding follows.
These are built into how statements are reconciled, not written into a prompt and hoped for.
- A request is not a commitment
- “Can we launch in June?” is stored as a request. Only an agreement makes it a promise.
- Newer does not always win
- Recency alone never overrides authority. A later message from the right person replaces an earlier one; a rumour does not.
- Silence proves nothing
- A promise nobody mentioned again is not marked done. Completion needs evidence.
- Heard is not confirmed
- Something said in a meeting stays “heard” and reaches agents labelled that way until it is confirmed.
- Truth is not attention
- Whether something is true and whether it deserves your attention are separate judgments. A stranger writing “urgent” does not make it urgent.
- Corrections travel
- Correct a statement once, and everything that rested on it is revisited. Every connected agent reads the fix.
Kept current where the change landed.
The Understanding Engine maintains an overview of what matters now. When something changes, it reopens only the items that cited it, instead of rewriting everything on a schedule.
- Change-scoped refresh
- A new fact reopens the parts of the overview that depend on it. A new meeting tomorrow reopens its own concern and nothing else.
- Bounded by design
- Full rebuilds are capped, and focused updates are rate-limited, so cost tracks how much your work changed, not how often a timer fires.
- Failures are named
- When a refresh can’t run, the reason is recorded and the previous understanding stays in place. A failed attempt never overwrites a good one.
What an agent actually receives.
Agents don’t get a dump of everything. Each read resolves the people, matters and dates the request names, applies the agent’s access, and returns a short packet with sources, labelled by status.
- Retrieval that isn’t just vectors
- Keyword, fuzzy, structural and semantic retrieval work together over the graph. Semantic search is one signal, never proof.
- Deterministic packets
- Opening your workspace and building an agent’s packet read from maintained records. They don’t ask a model to improvise the context.
- Times in your time zone
- Every time in a packet carries its offset, so an 11:00 meeting in Jerusalem is never read as 08:00.
- Every read on the record
- The exact request and the exact packet are kept, so you can see what any agent worked from.
request: "Help me plan Friday's Northpeak call" resolved: Northpeak · Mara Levin · Fri Oct 2 current Mara owns the renewal trusted · email Wed Limited June pilot agreed trusted · your email Tue Renewal terms due Friday trusted · meeting Mon replaced Daniel owns the renewal superseded Wed unknown Who sends the terms now sources: 3 · scope: this agent's grant
Measured, not claimed.
Numbers from real questions on our own workspaces while we built this. They describe specific changes on specific data, not a benchmark.
- 21.8K → 5.9K
- Average characters handed to an agent per read, after reads were built from what the request names. 23 real questions.
- 21 → 5
- Of those 23 answers, how many were “mostly noise” before and after the same change.
- ~1–4K vs ~34–37K
- A synthesized briefing over every source, compared with the older approach of pasting per-source summaries.
- 5 of 8
- Live overviews that, after one change, could be refreshed without a full rebuild: one needed no model call, four updated only the touched items.
Where it stops.
The honest edges, as they stand today.
- It can be wrong
- Extraction is done by models and can misread. That is why every statement shows its source and can be corrected in one place.
- It is not instant
- New work is picked up as your sources sync, not the second it arrives.
- Reads are bounded
- A packet has a size ceiling. Relevance decides what goes in, and a related conversation can still be missed.
Go deeper.
- ArchitectureFollow one fact from email to every agent
- 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
- Understanding EngineFrom graph to a read an agent can use
- Memory vs understandingWhat a memory layer does, and what understanding adds
- The 31 problemsEverything you have to solve, in one list
Your agents will change. Your understanding shouldn’t.
Invite-only private beta.