This page describes how the system holds state over time — one layer of the architecture. What that becomes on a shop floor is under Process AI; how processing works is under processing.
Why State Is the Difference
A language model works one request at a time and starts from nothing after each. For an operation that means every question is answered without knowledge of the previous one. This context layer is the part of the architecture meant to close that break — a query today should be answered in the context of what happened at the same site before.
The layer has two parts, and both are described in detail below: a reflection cycle of six steps that runs after each decision, and a memory architecture of five separate databases. The construction is settled. What is open is how far it already reaches in live operation today — that is being established with our engineering team. Which is why every statement on this page says whether it refers to today's operation or to the construction.
The Self-Reflection Cycle
After every decision, Henri runs a six-stage reflection cycle before moving on to the next perception:
- Perceive: Henri takes in his environment and recalls what he did in a comparable situation before.
- Process: New information is connected to existing contextual knowledge, not handled in isolation.
- Evaluate: Was the last decision right? How confident was the system? It checks its own work.
- Reflect: Gaps are identified — things that were misjudged or are not yet known.
- Anchor: What was learned is stored permanently in memory, not just held for the current session.
- Expand: The model of the operational environment grows, and the cycle begins again.
This cycle runs continuously in the background, independent of whether an active request is being handled — it's the mechanism by which repeated experience turns into genuine understanding of the specific installation.
Five-Database Memory Architecture
The context layer is split across five mutually independent databases, each with its own job: core identity (what the system knows about itself and its installation), signal statistics (the normal ranges of the connected measuring points), operational knowledge (what to do in a given situation), contextual awareness (which circumstances currently apply) and event patterns (what happened in comparable cases before).
The separation is deliberate: a fault in one database must not distort the others. Together they are meant to hold state beyond a single session. How far that carries across a restart today is exactly the question we are settling with our engineering team — the split itself is settled, its reach in operation is not.