Most sales data is stored as snapshots: the current stage, the current contact list, the latest note. That makes it hard to answer the questions that matter, like what changed, when and why. We designed CXAstra’s understanding layer around time from the start.
Entities, events and claims
The model has three primitives. Entities are the things that persist — accounts, people, products, competitors. Events are things that happened — a meeting, an email, a contract signed. Claims are statements extracted from events, such as “Priya is the economic buyer” or “budget approval is expected in Q4”. Each claim carries its source, timestamp and a confidence score.
Why claims, not fields
A field holds one value and overwrites the old one. A claim can be contradicted, confirmed or superseded by later evidence while keeping its history. When two conversations disagree about a buyer’s timeline, we store both, show the conflict and let recency and source authority decide which is shown by default.
“Overwriting data hides change. Recording claims over time makes change the thing you can see.”
Resolving identity
The same person appears as an email address, a calendar invite, a name spoken on a call and a CRM contact. Identity resolution merges these using deterministic matches first, then probabilistic signals such as co-attendance and email domain, with every merge reversible and auditable.
Serving it fast
- An append-only event log is the source of truth.
- Materialised views keep the current state of each entity ready to read.
- A retrieval index over claims grounds every answer Astra produces.
The result is a customer memory that can answer “what do we know?” and “how do we know it?” at the same time — which is what makes AI suggestions trustworthy.