Bi-temporal answers when a fact is true. The segment vector answers for whom.

Current-generation temporal-graph systems (Zep, mem0, Graphiti) answer one question well: is this fact true right now? At KAAS scale the question that pays the bills is one step further. Is this fact true for whom, right now? A surface claim about a brand can be true at one tier and contested at another. The same claim can hold in EMEA boutiques and fail in Latam mass-market stores. Without a way to scope a fact to a slice of the workforce or the customer base, the answer flattens into a corpus-wide average that nobody actually believes.

The platform attaches a segment vector to every fact. The vector is a sparse projection over the segmentation axes that matter for the vertical (tier, geography, store format, customer segment, brand partner, demographic). A query intersects the vector against the caller’s scope. Facts whose segment vector does not intersect the scope are not returned. They are not hidden by a policy layer. They are not in the result set.

The shape used in early commercial deployments captures across four tiers. Tier-1 Direction (senior leadership) holds the brand’s internal story. Tier-2 Brand Operators (marketing, comm, merch, design) execute against that story. Tier-3 Front Line (store managers and franchisees) see what actually happens on the floor. Tier-4 Customers see the brand from the outside. The same surface claim takes different values at each tier.

“The brand is a value-fashion destination” can test true at Tier-4 today, be contested at Tier-2 today (where the brand operators are working a repositioning that has not yet landed at the customer), and have been true at Tier-1 in 2023, when the strategy was different. Same claim, three valid_at windows, three segment scopes. The platform holds all three states in the same corpus and returns the right one for the right query.

That property is what lets the Direction team see what their advisors and customers see, without flattening the four perspectives into a single average that obscures the gap.

At a large beauty-retail deployment, every brand partner on the shelf pays for a scoped API into the corpus. The scope is a projection of the segment vector. A brand partner’s API key projects onto brand_partner = <their_brand>. The query returns the facts that intersect that projection. Every other brand partner on the shelf gets the parallel scope, each restricted to its own slice.

One brand partner cannot see another brand partner’s facts. The segment vector ensures the cross-partner facts are not in the query result. There is no access control rule running on top of a shared dataset. There is a single dataset, and the query language scopes the result before any access control runs.

The same property holds at the customer-segment axis. “Shade X gets rejected most” can be true for Latam mass-market stores and false for EMEA boutiques. The fact carries both windows on its segment vector. The Latam-scoped query returns true. The EMEA-scoped query returns false. Both are correct.

Every vertical names its own axes. The general structure is the same: a sparse vector across a small number of axes (typically four to seven), where each axis carries a discrete or low-cardinality value set.

Retail (four-tier shape). Tier (direction, brand operator, front line, customer), region, store format, customer cohort.

Beauty retail at scale. Region, market, store format, brand partner, customer segment, advisor tenure band.

Consulting. Practice area, partner cohort, engagement scale band, client industry.

Restaurants. Concept, region, store format, shift band, day part.

The axes are documented per deployment. They are additive: a new axis can be introduced when the methodology requires it (a new customer cohort emerges, a new region opens), and the historical corpus is re-projected so the new axis is populated retroactively where the conversations contain the signal.

The temporal model and the segment vector cooperate when a contradiction lands. The platform does not silently invalidate a fact that has been overtaken by new evidence. It checks whether the new evidence applies corpus-wide or only to a slice. If only to a slice, the old fact is split: a new fact carrying the contradicting evidence is created with a tighter segment scope, the old fact loses that scope from its segment vector, and the old fact remains valid where it still applies.

This is the difference between a tool that forgets the past and infrastructure that holds the past honestly. The corpus does not pretend the contradiction did not exist. It documents the slice where the new claim applies and preserves the slice where the old one still holds.