Dice Question Streamline Icon: https://streamlinehq.com

Optimal number of Kinaema memory embeddings and dependence on training sequence length

Determine whether the observed saturation in performance around N≈20 memory embeddings for Kinaema is caused by training on sequences limited to length T=100, and establish if training on longer sequences shifts the optimal number and usage of memory embeddings.

Information Square Streamline Icon: https://streamlinehq.com

Background

The authors perform a sensitivity paper varying the number of memory embeddings while keeping embedding size constant and observe that performance scales until roughly N=20 embeddings.

They suggest this may be an artifact of the training sequence length and that longer training sequences might change the optimal memory configuration.

References

We conjecture that this is due to training with sequences of length $T{=}100$ and longer training could lead to bigger choice for optimal memory usage.

— Kinaema: a recurrent sequence model for memory and pose in motion (2510.20261 - Sariyildiz et al., 23 Oct 2025) in Section 5, Sensitivity study: memory capacity (Table 3: Kinaema: varying memory structure, Mem-RPE)