Context & Profiles API

Two related surfaces: context bundles (the token-budgeted, provenance-bearing payload you put in a prompt) and profiles (the non-temporal “who is this subject” snapshot).

Context bundle

POST /api/v1/projects/{projectId}/lattice/context

Aggregate a subject’s profile + active patterns + predictions + memories + observations into one token-budgeted bundle. Every claim carries source ids so any inference can be traced.

{
  "subject": { "kind": "customer", "externalId": "acct-42" },
  "include": ["profile", "patterns", "predictions", "memories", "observations"],
  "maxTokens": 2000,
  "memoryLimit": 20,
  "predictionLimit": 10,
  "excludeCategories": ["medical"]
}

include — which of the five sections to build. Omit for all five.
maxTokens — hard cap; sections truncate to fit (200–16000, default 2000).
excludeCategories — min-necessary access: drop categories from the bundle (e.g. keep medical / HR / PII out of a sales-context query against the same project).

The response carries profile, patterns[], predictions[], memories[], observations[], a provenance block (memory / pattern / observation ids), tokensEstimate, a truncated[] list of dropped sections, and a guidance directive to place at the top of the prompt so the model treats the bundle as already-known.

Batch context

POST /api/v1/projects/{projectId}/lattice/context/batch

Build bundles for many subjects (≤500) in one call — the bulk-load path that replaces N round-trips. The same options apply to every subject; bundles are returned in request order.

{
  "subjects": [
    { "kind": "customer", "externalId": "acct-42" },
    { "kind": "customer", "externalId": "acct-43" }
  ],
  "maxTokens": 1500
}

Batch context is the endpoint to reach for when loading memory for hundreds of customers at once (e.g. a campaign build) — one call, not one per subject.

Profile

POST /api/v1/projects/{projectId}/lattice/profile (also POST /api/v1/projects/{projectId}/memory-intelligence/profile)

The non-temporal counterpart to prediction — “who is this subject?” folded from their active behavior patterns.

{ "subject": { "kind": "customer", "externalId": "acct-42" } }

Response:

{
  "subject": { "kind": "customer", "externalId": "acct-42" },
  "rfmSegment": "loyal_customer",
  "recencyScore": 5, "frequencyScore": 4, "monetaryScore": 4,
  "topEntity": "tonys-pizza",
  "cadenceSummary": "every 7 days",
  "risks": [
    { "kind": "declining_engagement", "description": "…", "severity": 0.6, "sourcePatternId": "mem_…" }
  ],
  "contributingPatternIds": ["mem_…"],
  "generatedAt": "2026-07-13T00:00:00Z",
  "durationMs": 6
}

risks are the active signals (declining engagement, at-risk RFM segments, missed cadence) an agent should surface before a non-trivial action. Every field traces back through contributingPatternIds.

Behavior Discovery Cohorts & Estimators