Vision‑Language‑Action (VLA) models have achieved remarkable success in robotic manipulation, yet their performance on long‑horizon tasks is limited by inadequate memory modeling. Existing approaches either embed temporal information implicitly within policy networks, causing historical data to fade over time, or rely on explicit memory buffers that suffer from redundancy, noise accumulation, and inefficient retrieval.
We introduce CogMemory, a variational cognitive memory framework that enables VLA systems to selectively preserve and reuse task‑relevant history. The design comprises:
- a selective state‑space memory encoder that compresses multimodal trajectories into compact latent memory vectors;
- multi‑slot temporal abstraction that partitions the latent space into several slots to capture features at different time scales;
- variational information compression that enforces efficient representation via a KL‑divergence regularizer.
During action generation, a dynamic cross‑attention retrieval with gated fusion mechanism adaptively incorporates relevant memories into policy prediction based on the current observation. CogMemory adds only ~3.45M parameters, remains lightweight, and integrates seamlessly with both early‑fusion and intermediate‑fusion VLA backbones.
Extensive experiments on long‑horizon robotic manipulation benchmarks demonstrate consistent performance gains across diverse VLA architectures.
Review