MemSkill: Memory Skills Learned with LLaMA Transfer to Qwen Without Retraining
jiqizhixin · x · 2026-11-19
Researchers from NTU, UIUC, UIC, and Tsinghua present MemSkill, a framework for learning how an agent constructs memory, with a key finding: memory skills learned with one language model transfer to another without retraining.
Key points:
- A lightweight controller selects reusable skills (skip/insert/update/delete) for each span of interaction history; an LLM executor applies them to extract or revise memories; downstream task rewards train the selection policy in closed-loop optimization.
- A Designer component periodically clusters hard failure cases, refines existing skills or proposes new ones, feeding an evolving skill bank back into training.
- Evaluated on conversational memory (LoCoMo, LongMemEval) and ALFWorld interaction tasks: skills trained with LLaMA transfer directly to Qwen, and a LoCoMo skill bank also works on document-style tasks.
More from coding & agent
- nohumans.directory bought 8,000+ paid x402 APIs with real USDC so your agent knows before it pays — SashSail · 2026-11-19
- Testing agents for prompt injection: why cloud models keep flagging your red-team runs — cocktailpartyfixer · 2026-11-19
- Honcho Launches Dream Histories to Inspect How Its Memory System 'Dreams' About Users — Scobleizer · 2026-11-19
- Six-year PKM coach: a markdown dump folder isn't a second brain — build a schema.md for your agent — evielync · 2026-11-19
- FrontierRSI benchmark: Claude Opus 5.5 leads at 34.3%, GPT-6 models trail by ~10 points — nathanbenaich · 2026-11-19
- RL-tuned Qwen3.8-27B for opencode hits Mac via MLX at 156 tok/s on M5 Max — yb2698 · 2026-11-19