BINDR: Bottleneck Iterative Neural Dynamic Retrieval, a General-Purpose Sequence Model

Sequence data appear everywhere, from speech and machine sensors to engines, aircraft and energy systems. These signals arrive one step at a time, mix fast patterns with slow trends, and are often noisy. Most sequence models read each step once. This work asks whether reading the data twice improves performance and sample efficiency. We introduce BINDR (Bottleneck Iterative Neural Dynamic Retrieval), a ~50,000-parameter sequence model that reads the data twice (a GRU pass followed by a state-space pass), uses a multiscale causal convolution to capture patterns at different time scales, and has a bottlenecked shared exponential memory with FiLM conditioning. BINDR was tested against six parameter-matched rivals (GRU, LSTM, TCN, Transformer, DiagonalSSM, MambaLike) across 8 public datasets spanning 7 domains and 5 training-data amounts (200 settings, ~2,700 total runs including ablations and reruns). BINDR ranked first in 75 of 200 settings versus 41 for the next-best model at the primary comparison point, and 84 versus 29 by the end of training; its end-of-training wins over all six rivals remained significant after Holm correction. With 5–25% of the training data, rivals typically needed 1.2–2x more data to match it. On NASA C-MAPSS FD001 jet-engine data with 10% of the training set, BINDR's error was 30% lower than the next-best model. On the official BatteryLife benchmark, using its own splits, seeds and metrics, BINDR won 4 of 8 headline scores against the best published baselines. A pre-registered control experiment found that the second pass does not improve performance (82 wins vs 89 losses against a same-size one-pass BINDR), so the two-pass hypothesis was not supported. BINDR's advantage is not explained by depth: it beat a same-size 2-layer GRU 117–53. Ablations indicate its strength comes mainly from the multiscale mixer and FiLM conditioning, with the memory improving sample efficiency. Submitted to IRIS National Fair 2026–27 (Robotics & Intelligent Machines), where it was selected as a national finalist.