Operator-LSTM: Content-Conditioned Time-Decay Dynamics for Irregularly Sampled Streams

Irregularly sampled streams force recurrent models to decide how much stale state to retain after variable elapsed time. We study Operator-LSTM, a discrete event-time cell that separates measuredtime exponential carry from a masked bounded content write, with retention recomputed from current content and recurrent state. We use it as a controlled testbed, not a generally superior architecture. An initial three-seed noisy time-to-live (TTL) test gives a +4.39-point mean paired difference over a learned global per-dimension timescale baseline. A separate pre-specified 10-seed follow-up has 8/10 positive paired differences and a +5.81-point median; two retained near-chance comparator outcomes enlarge the +10.47-point mean, so we do not treat it as uniform. Learned timescales do not recover the ground-truth TTLs, and write-free interval subdivision produces small but nonzero state deviations rather than an exact semigroup. Under a frozen PhysioNet 2012 protocol, the synthetic advantage is not corroborated and GRU-style baselines perform better.