Typed Attributes API

Typed attributes are the structured / numeric data source the engine reasons over for range queries, thresholds, and trajectories — instead of opaque metadata JSON. Numeric and anomaly predictions read from here.

All routes are under /api/v1/projects/{projectId}/memory-typed.

Attribute definitions

POST /memory-typed/attributes

{
  "attributeKey": "order_total",
  "dataType": "numeric",
  "unit": "usd",
  "minValid": 0,
  "maxValid": 100000,
  "required": false
}

dataType is numeric | categorical | temporal | boolean.

GET /memory-typed/attributes lists the registered definitions (optional attributeKey, limit, offset).

Observations

Ingest typed measurements. Each row is validated against its attribute definition — accepted or quarantined — and accepted numeric values are folded into per-subject running accumulators.

POST /memory-typed/observations (synchronous, returns the report)

{
  "observations": [
    {
      "attributeKey": "order_total",
      "subjectKind": "customer",
      "subjectExternalId": "acct-42",
      "valueNumeric": 48.5,
      "observedAt": "2026-07-12T18:03:00Z",
      "source": "pos"
    }
  ]
}

Exactly one value* field is meaningful per the attribute’s dataType (valueNumeric / valueText / valueBool / valueTs).

Response — the ingest report:

{ "accepted": 1, "quarantined": 0, "duplicates": 0, "quarantineReasons": {} }

For high volume, POST /memory-typed/observations/enqueue queues the batch and returns 202 { "enqueued": N }.

GET /memory-typed/observations queries raw observations by subject, attribute, time window (since / until), numeric range (minValue / maxValue), and status.

Accepted numeric observations also emit a typed.observation event, so memory-value alert rules can fire on them. See Events & Alerts.

Accumulator

GET /memory-typed/accumulator?subjectKind=…&subjectExternalId=…&attributeKey=…

Read the running statistics for one subject + attribute:

{
  "subjectKind": "customer",
  "subjectExternalId": "acct-42",
  "attributeKey": "order_total",
  "count": 12,
  "sum": 583.4,
  "min": 12.0, "max": 92.5, "last": 48.5,
  "cumulative": 583.4,
  "mean": 48.6, "variance": 410.2, "stddev": 20.3,
  "ewma": 45.1
}

mean / variance / stddev are derived on read from the stored moments; ewma is the exponentially-weighted moving average that anomaly detection uses as a baseline.

Metering

Typed observations count against the events-ingested meter and are gated by that entitlement.

Memory API Predictions