Honeycomb: Constant-Size Scene Memory Representation for Video World Models

arXiv · Jack Wei Lun Shi, Kaichen Zhou, Haoyu Chen, Yufeng Weng, Keane Ong, Ruojin Cai, Hang Hua, Justin K. W. Yeoh, Mengyu Wang · 2026-10-04

This item is no longer on the current shelf. Its page remains available, and editorial corrections may be made. See what is featured now.

Honeycomb shows you can keep video world-model memory fixed in size without sacrificing long-horizon consistency.

What it does

Honeycomb introduces HexMemory, a low-rank scene-memory that stores features in exactly six spatial and spatiotemporal planes regardless of video length. A feed-forward writer maps each newly generated chunk onto fresh plane features, warps the existing planes to cover the expanded spatial-temporal range, and fuses them with confidence-weighted pooling plus a learned residual correction. A reader then pulls latents from HexMemory to condition the next generation step, so the model never needs to re-process the full history. Experiments on WorldScore and RealEstate10K demonstrate strong generation quality and robust revisit consistency while the memory footprint stays constant.

Why it matters

Constant-size memory eliminates the quadratic storage blow-up that has limited video world-models to short horizons. By avoiding per-scene optimization, training and inference become faster and cheaper, opening the door to production-scale video synthesis and interactive applications. The approach also shows that a compact latent store can preserve temporal coherence, a key hurdle for realistic video generation.

How it applies

Practitioners can drop HexMemory into existing video-generation pipelines as a plug-in memory module, reusing the same writer/reader logic. Because the writer only touches the newest chunk, integration adds minimal overhead and scales to arbitrarily long sequences. The released code makes rapid prototyping possible for VR content, simulation, or any domain that needs long-form visual generation.

The short version

Imagine a video model that never runs out of memory.

My take

The clever part is treating memory as a token-level latent that can be warped and merged rather than a raw frame buffer. That shift makes the problem more like language modeling—where a fixed context window works—than like raw video storage. For builders, it means you can focus on model quality instead of engineering ever-larger caches.

How it connects

• Aligns with the broader move toward low-rank, compressed representations in transformers.
• Echoes recent work on constant-size caches for long-context language models.
• Sets a template for other modalities (audio, 3-D) where persistent but bounded memory is needed.

Bottom line

Swap in HexMemory to keep your video world model scalable without blowing up memory.

Takeaways

• HexMemory uses a fixed set of six spatial/spatiotemporal planes as a constant-size scene memory.
• The writer processes only the latest video chunk, avoiding full-history recomputation.
• Confidence-weighted pooling and residual correction preserve consistency across revisits.
• WorldScore and RealEstate10K experiments confirm high video quality and stable long-term coherence.
• Open-source code enables immediate adoption in downstream video-generation systems.

Topics: video generation, world models, memory efficiency, latent representations