AI
MemFold: Learning Compact Soft Memory for Long-Context Personalization via On-Policy Optimization
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 03/10/2026, 08:16:50 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 21:18:24 UTC.
Briefings
MemFold is a method for learning compact soft memory for long-context personalization via on-policy optimization. It compresses a query-conditioned textual memory into K continuous vectors forming an interface, trained on the reader's own rollouts using group-relative rewards and confidence-gated on-policy distillation. The teacher is frozen during training, ensuring supervision remains on the student's current distribution. Across three Qwen backbones, MemFold achieves the highest accuracy on PersonaMem-32K and PersonaMem-128K, with performance margins widening at longer history lengths. Results transfer to PrefEval and LongMemEval without target-domain training. Ablations attribute most gains to the reward term and a smaller additional gain to the teacher signal, while memory interventions show the reader depends on the instance-specific content of its soft memory.
Key points
- MemFold compresses a query-conditioned textual memory into K continuous vectors that form the reader's memory interface.
- The method optimizes a fixed-budget soft memory by the behavior it supports, using group-relative rewards and confidence-gated on-policy distillation.
Context and limitations
Results are limited to three Qwen backbones and PersonaMem-32K/128K benchmarks; performance on PrefEval and LongMemEval requires transfer without target-domain training.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 08:16:50 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifqfm5g3tuuvakbmqktlisddyexvyy7glai5detgpdegkca6qutzm