Description
Long-horizon agents mostly use hand-written memory rules that don't fit new tasks. MemSkill proposes a framework that learns, refines and reuses memory skills from task feedback, letting agents evolve how they remember.
A data-driven loop replaces static memory operations for more adaptive memory across settings; published at NeurIPS 2026.
Memory skills:Learned and reused.
Evolution:Improves from feedback.
Long horizon:Many settings.
A data-driven loop replaces static memory operations for more adaptive memory across settings; published at NeurIPS 2026.
Features
Memory skills:Learned and reused.
Evolution:Improves from feedback.
Long horizon:Many settings.
