Tension-based decision architecture with active memory for incipient ruptures
M.E.M.I. is not a classical controller and not pure statistics. It is an architecture that reads the development of the state rather than merely reacting to individual events — and asks not only "is something wrong now?" but "are we moving into something that resembles a known rupture?"
M.E.M.I. has begun to handle a classic balance mechanically rather than through threshold adjustment: the system has shown three things in the most important tests samtidig —
→ Markedly better detection of slow drift
→ Opening of previously blind value-drift scenarios
→ Elimination of false alarms in oscillating scenarios
The three error types normally pull in opposite directions. Handling them together is the documented breakthrough.
M.E.M.I. should at this point be understood as: an experimental, tension-based decision architecture with active memory for incipient ruptures. It is not merely a model that classifies. It is a system that attempts to keep track of three things simultaneously.
The system compares the current state development with stored episodes from past ruptures — not as pattern matching, but as structural recognition.
Trajectory-awareness: the system measures direction and acceleration in the state, not only the current position. A system can be far from the threshold but moving rapidly towards it.
Not merely "what happened last time", but what was structurally different about the rupture, and whether that pattern should increase or decrease sensitivity now.
Development has not proceeded through broad parameter search, but through a series of precise diagnoses and targeted patches. Every step is built from observed failure mechanisms.
Stronger drift episodes must weight more than weak ones when comparing against memory. Linear weighting underestimates systemic movement.
Early alarm is not only about thresholds, but about how a rupture is stored as an episode. Incorrect storage leads to delayed or absent recognition.
Static episodes cannot track slow drift. They become stale and must be updated via trajectory-aware refresh to remain relevant.
Oscillation cannot be separated from real drift via goal-distance alone. The system requires separate gating logic to avoid false alarms in oscillating scenarios.
Sustained pressure can generate episode spam that contaminates memory. Cooldown mechanisms reduce this without harming detection.
Den første version der samlet set gav bedre detektion, åbnede blinde scenarier og eliminerede falsk alarm — uden at skabe nye regressioner. v6.8 fungerer nu som referencepunkt for næste spor.
Trajectory-aware refresh ensures episodes do not become stale during gradual system changes.
The system now detects value-drift patterns that previously went unnoticed because they did not match known goal-distance patterns.
Oscillation gating prevents the system from interpreting rhythmic variations as incipient ruptures.
The most important open problem is no longer early detection — but de-eskalation.
M.E.M.I. has become good at detecting and holding the significance of a rupture. The next research question is: when and how should the system release an alarm state, without forgetting too early?
This is a harder problem than classical recovery, because the system must not become blind again immediately. The work points towards mechanisms for controlled relaxation of memory and risk, without damaging the documented detection.
As part of the M.E.M.I. research, an alternative controller concept has been developed. Sukker_Fabrikken is designed as a counterpoint to classical PID regulation and symbolic rule-based control — focusing on context-sensitive adaptation rather than fixed rule application.
A controller architecture that works with tensions and context rather than fixed rules. Builds on M.E.M.I.'s trajectory logic but as an independent regulation mechanism.
Status: conceptually developed, awaiting empirical testing within the M.E.M.I. framework.
Classical controllers correct deviations. Sukker_Fabrikken attempts to understand retningen af afvigelsen — og justere reaktionen afhængigt af om systemet er på vej ind i et brud eller på vej ud af det.
Relevant to the de-escalation problem: can a context-sensitive controller help the system release the alarm state more controllably?
M.E.M. (Meaning · Experience · Model) is the theoretical foundation. M.E.M.I. is the technical expression of the same underlying structure: a system that works with tensions — not answers — and that understands itself as part of the field it navigates.
M.E.M.I. has passed an important milestone: documented failure modes, reproducible improvements, and a baseline that holds. We are looking for professionals and organisations willing to think alongside us in the next phase — particularly around de-escalation and Sukker_Fabrikken.