From Sensor Streams to Bounded Machine Memory: Architecture, Negative Transfer Results, and Domain Specialists

Industrial and edge equipment produce continuous telemetry that cannot be kept in full on flash storage, while first-in-first-out buffers overwrite rare fault episodeswith routine data. We study μEngram, a bounded stateful memory that compresses short sensor windows into noise-whitened low-dimensional vectors, tracks operating regimes with a fixedbudget of semantic, candidate and episodic slots, quarantines unverified shifts, and pins raw waveform slices for rare episodes. We evaluate it and several encoder variants on a syntheticgenerator and ten public benchmarks: CMAPSS, MetroPT-2 and MetroPT-3, SKAB, SCANIA Component X, OS-Metrics, EngineAD, MFPT, CWRU and TSB-AD-U. On the synthetic stream the memory cutswrite operations by 86.2% over 10 seeds while retaining every injected event. Across the benchmarks, alarm behavior is governed primarily by threshold design, calibration coverage andalarm-counting rules rather than rank-ordering metrics: models reaching 1.0000 AUROC on CWRU still alarm throughout clean normal operations. Time-series foundation models and theirdistilled student models shifted alarm timing without consistently improving operational detection, and role-separated domain specialists outperformed global pooling on multi-subsystemtelemetry. A crash-image study across 6,599 simulated recovery states shows that detector state must be checkpointed to prevent spurious alarms and missed fault openings across restarts.We release the complete implementation, benchmark runners and reproduction instruction