Predictions

MemMesh isn’t only a memory store — it’s a general prediction engine. You declare what you want predicted and the engine predicts it from the subject’s observation history. You never pick a model; the target type selects one for you.

const p = await tf.lattice.predictTarget(
  { kind: "customer", externalId: "acct-42" },
  { kind: "event_occurrence", eventType: "subscription_cancelled" },
  { horizonDays: 90 },
);

Declare a target

A prediction target is a small, typed spec — not a hard-coded endpoint. The kind drives model selection:

Target kind	Answers	Example
`event_occurrence`	Will it happen within the horizon?	churn, reorder, pay-late
`numeric`	How much?	next order amount, next lab value
`event_time`	When next?	next visit, replenishment date
`anomaly`	Is the latest reading an outlier?	fraud, sensor drift

event_occurrence / event_time use eventType; numeric / anomaly use attributeKey.

The result is calibrated

Every prediction carries an interval, not just a point estimate — and a stated 80% means roughly 80% in reality (conformal coverage, not vibes).

Calibration is queryable: GET /api/v1/projects/{projectId}/lattice/calibration returns confidence buckets mapped to the realized hit-rate of past predictions in each band — the honest answer to “of the things we rate ~0.8, do ~80% actually happen?”. See Outcomes & Calibration.

if (!p.abstained) {
  console.log(
    `churn risk ${(p.probability * 100).toFixed(0)}% ` +
    `[${(p.probabilityLower * 100).toFixed(0)}–${(p.probabilityUpper * 100).toFixed(0)}%]`,
  );
}

It abstains — a first-class answer

When there isn’t enough signal — too little history, out-of-distribution input, an interval too wide to be useful — MemMesh returns an abstention instead of a confident guess.

⚠️

An abstention means unknown, never “no” or “low risk”. Always check abstained before reading a value. This honesty is what makes the engine safe for regulated use.

if (p.abstained) {
  console.log("not enough signal —", p.abstentionReason);
}

Every prediction is explainable

A prediction carries its provenance: the ids of the observations it was derived from and a human-readable derivation, so any single prediction can be explained or audited.

console.log(p.explanation);          // "12 purchases, every ~21 days"
console.log(p.evidenceMemoryIds);    // the observations behind it

How it relates to patterns

Mined behavior patterns become features and priors for these predictions — they give a brand-new subject a useful, calibrated, low-confidence answer immediately, sharpening as real history accrues.

See the TypeScript SDK for the full predictTarget surface, and Behaviors for discovering the patterns that feed predictions.

Lattice & Patterns Behavior Discovery