Event

NTU Team Releases MemSkill: Learnable and Evolvable Memory Skills for Agents

2 reports · 2 sources

A research team from Nanyang Technological University and other institutions published MemSkill at NeurIPS 2026, packaging memory operations into learnable, reusable, and evolvable Memory Skills: a Controller selects skills, an Executor constructs memory, and task feedback is used to optimize skill selection, revising and expanding the skill library based on failure cases. Experiments used LLaMA and Qwen as base models and covered LoCoMo, LongMemEval, and ALFWorld. The team states that the results on Qwen come from direct transfer of skills learned on LLaMA, and that LongMemEval reused the skill library learned on LoCoMo, both without retraining; the skill library evolved on LoCoMo was also transferred directly to HotpotQA to test generalization from multi-turn dialogue to document-based question answering, likewise without retraining specific to HotpotQA. With the default selection of 7 skills, the team states that MemSkill outperforms MemoryOS and A-MEM across three context lengths.