SCHEMA // AGENTIC BRAIN

Memory is only useful when it is typed, dated and traceable.

Nothing is stored as an opaque embedding blob. Every unit of knowledge is a row you can inspect, query and revert. The graph is the product.

Node types

TYPEDESCRIPTIONCOUNT · 50-PERSON DEMO
ENTITYService, system, team, person or vendor412
DECISIONA choice made, with the tradeoff recorded168
FACTAn atomic assertion with confidence + age1,842
ARTIFACTThe raw source record it was derived from9,317

Edge types

  • derived_fromfact → artifact

    provenance, always required

  • supersedesfact → fact

    newer assertion replaces older

  • contradictsfact ↔ fact

    flagged for human resolution

  • owned_byentity → person

    recall routing when a fact expires

Freshness decay

Every fact carries a half-life derived from how often its source class changes. A deploy config decays in weeks; a company value decays in years.

Contradiction detection

When two facts about one entity disagree, both are surfaced side by side with authorship and date. The graph never silently picks a winner.

Time travel

The graph is append-only. Query it as of any timestamp to see exactly what the organisation believed on that date.

SAMPLE // FACT RECORD
{
  "id": "fact_9f31c",
  "entity": "svc.payments-gateway",
  "assertion": "Rate limit is 1000 req/min per tenant",
  "confidence": 0.94,
  "state": "live",
  "asserted_at": "2026-08-03T09:12:00Z",
  "half_life_days": 90,
  "derived_from": ["repo/gateway.yaml#L42", "slack/C0X92/p17512"],
  "contradicts": []
}

Demo tenant. Illustrative record, no live connection.

Made by Sorvanis

We build data infrastructure that answers the questions a business keeps re-asking itself. Agentic Brain is one of them.

dev@sorvanis.ai