# Compress to Remember: Learning Compact Memory via On-Policy Distillation for Long Video Generation

arXiv: https://arxiv.org/abs/2609.36364 · HTML: https://mlsanity.com/paper/2609.36364/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Multimodal distillation, one of 12 directions in Knowledge distillation.
- Builds on 23 of its 26 references that are in the map.
- First posted on arXiv 28 Sep 2026.

## Placed in

- Knowledge distillation / [Multimodal distillation](https://mlsanity.com/distill/multimodal-distillation/)

## Builds on (most cited first, 23 of 23)

- [Scaling Laws for Neural Language Models](https://mlsanity.com/paper/2001.08361/) (2020)
- [Wan: Open and Advanced Large-Scale Video Generative Models](https://mlsanity.com/paper/2503.20314/) (2025)
- [VBench: Comprehensive Benchmark Suite for Video Generative Models](https://mlsanity.com/paper/2311.17982/) (2023)
- [Movie Gen: A Cast of Media Foundation Models](https://mlsanity.com/paper/2410.13720/) (2024)
- [Self Forcing: Bridging the Train-Test Gap in Autoregressive Video Diffusion](https://mlsanity.com/paper/2506.08009/) (2025)
- [Learning to Compress Prompts with Gist Tokens](https://mlsanity.com/paper/2304.08467/) (2023)
- [LongLive: Real-time Interactive Long Video Generation](https://mlsanity.com/paper/2509.22622/) (2025)
- [Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation](https://mlsanity.com/paper/2602.02214/) (2026)
- [Infinity-RoPE: Action-Controllable Infinite Video Generation Emerges From Autoregressive Self-Rollout](https://mlsanity.com/paper/2511.20649/) (2025)
- [Deep Forcing: Training-Free Long Video Generation with Deep Sink and Participative Compression](https://mlsanity.com/paper/2512.05081/) (2025)
- [Rolling Sink: Bridging Limited-Horizon Training and Open-Ended Testing in Autoregressive Video Diffusion](https://mlsanity.com/paper/2602.07775/) (2026)
- [MemRoPE: Training-Free Infinite Video Generation via Evolving Memory Tokens](https://mlsanity.com/paper/2603.12513/) (2026)
- [Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models](https://mlsanity.com/paper/2606.25473/) (2026)
- [Relax Forcing: Relaxed KV-Memory for Consistent Long Video Generation](https://mlsanity.com/paper/2603.21366/) (2026)
- [MBench: A Comprehensive Benchmark on Memory Capability for Video World Models](https://mlsanity.com/paper/2606.00793/) (2026)
- [PackForcing: Short Video Training Suffices for Long Video Sampling and Long Context Inference](https://mlsanity.com/paper/2603.25730/) (2026)
- [LongLive-RAG: A General Retrieval-Augmented Framework for Long Video Generation](https://mlsanity.com/paper/2606.02553/) (2026)
- [WorldKV: Efficient World Memory with World Retrieval and Compression](https://mlsanity.com/paper/2605.22718/) (2026)
- [Echo-Infinity: Learning Evolving Memory for Real-Time Infinite Video Generation](https://mlsanity.com/paper/2606.04527/) (2026)
- [Long-Horizon Streaming Video Generation via Hybrid Attention with Decoupled Distillation](https://mlsanity.com/paper/2604.10103/) (2026)
- [DySink: Dynamic Frame Sinks for Autoregressive Long Video Generation](https://mlsanity.com/paper/2605.21028/) (2026)
- [SlotMemory: Object-Centric KV Memory for Streaming Long-Video Generation](https://mlsanity.com/paper/2605.31033/) (2026)
- [Pyramid Forcing: Head-Aware Pyramid KV Cache Policy for High-Quality Long Video Generation](https://mlsanity.com/paper/2605.13111/) (2026)

## Abstract (arXiv)

Standard video generators do not natively compact historical context into reusable memory tokens. As generation continues, the growing history makes it increasingly difficult to retain information from earlier frames due to long-context degradation. Key-frame-based approaches address this challenge by retaining selected past frames, but can discard information needed for future generation. Rather than relying on frame selection alone, we study whether a frozen video generator can supply the supervision needed to learn a compact representation of the history. We propose Prediction-Aligned Context Compaction (PACC), which uses a learned compressor to aggregate information across past frames into compact memory tokens. We train the compressor through on-policy distillation, using the same frozen generator both as a student when conditioned on compressed memory and as a teacher when conditioned on the full history. The student generates continuations, while the teacher provides targets for the same noisy inputs at each denoising step. Only the compressor is updated to align the student's predictions with these targets. We evaluate PACC on MBench, which jointly measures memory-event coverage and consistency. PACC outperforms the strongest baseline by 6.63 points on Causal-rCM and 3.19 points on Causal Forcing. Evaluation on VBench-Long using MovieGen prompts further shows that PACC produces minute-long videos with generation quality competitive with baselines. Together, these results show that learning to compact historical context can improve long-video memory without modifying the underlying generator.