Incarnate agents: keeping the experience in an external memory returns much more than distilling it in weights

Pre-recorded study on a vision-driven challenge: reconnecting recovery memory to a LoRA policy adds 34 points, distillation alone barely 1 point.

Should we integrate the experience of a robotic agent into its weights, or keep it separate and consult it when acting? Cheng-Yin Tu settled the question with a methodology that is rare in robotic learning: hypotheses, experimental arms, tests and rules of analysis were pre-recorded before any training. A "teacher" — a vision-language model that finds the lessons learned from its past attempts, with their images — produces trajectories on a vision-guided attainment task; LoRA "student" policies are then distilled from it.

+1 point vs. +34 points

Distilling this behavior into the weights improves the score by only one point compared to a control, without significance; as the test turns out to be undersized, the author withdraws this hypothesis rather than declaring it refuted. On the other hand, reconnecting the teacher's frozen memory to the students' weights adds 34 points (z = 5.15) and reaches the teacher's level, around 49%. The same memory buffer brings the basic model, without any adapters, to 48%, and increases three other adapters from 19-26% to 53-55%: once the memory is plugged in, the adapter no longer provides anything measurable.

The resulting instruction is to externalize memory. On a second, more difficult task, the advantage goes in the same direction (+6 points) without reaching significance, which the author links to insufficient memory coverage.