MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization
MemFold trains a fixed-size latent memory by whether it helps the assistant answer correctly, not by whether it can reconstruct the stored text.
The paper compresses query-conditioned textual memory into `K` continuous vectors used as the reader model’s memory interface. Training uses the model’s own rollouts, with task rewards plus a confidence-gated teacher signal from a frozen textual-memory model. The teacher is used only for scoring sampled tokens during training and is removed at inference. The authors report the best measured accuracy across three Qwen backbones on PersonaMem-32K and PersonaMem-128K, with larger margins on longer histories, plus transfer to PrefEval and LongMemEval without target-domain training. HF Daily Papers' note
The paper compresses query-conditioned textual memory into `K` continuous vectors used as the reader model’s memory interface. Training uses the model’s own rollouts, with task rewards plus a confidence-gated teacher signal from a frozen textual-memory model. The teacher is used only for scoring sampled tokens during training and is removed at inference. The authors report the best measured accuracy across three Qwen backbones on PersonaMem-32K and PersonaMem-128K, with larger margins on longer histories, plus transfer to PrefEval and LongMemEval without target-domain training. HF Daily Papers' note
score 5