Lattice & Patterns

The lattice is the layer above raw memories. Where memories record what happened, the lattice models how things relate and what tends to happen next — it’s where MemMesh’s prediction lives.

From observations to patterns

As events accumulate against an entity, the engine distills recurring behavior into patterns — compact descriptions of a regularity (e.g. “this account renews roughly every 12 months”). Patterns are first-class objects on the lattice that can be inspected, scored, and retired.

Prediction

Given a subject, MemMesh can forecast the next likely event from its patterns and return it with a confidence score. A prediction looks like:

subject: sarah-pizza
event:   renewal_due
date:    2026-07-01
confidence: 0.82

Calibration

A confidence score is only useful if it’s honest. MemMesh calibrates predictions against what actually happened — when reality diverges from a pattern, the pattern is recalibrated or deactivated. The goal: when it says 80%, it means it.

You can inspect this directly. GET /api/v1/projects/{projectId}/lattice/calibration returns the active patterns bucketed by stated confidence, with the realized hit-rate of their past predictions per bucket — a read-only reliability report. The closed loop that feeds it (decisions → outcomes → recalibrated confidence) is documented in Outcomes & Calibration.