How deep large-model Agents can run in long-horizon interaction depends on memory; but moving the entire 72B teacher to the device side is neither practical nor economical. DuoMem on arXiv 2606.29961 gives a fairly engineering-flavored solution — not relying on a larger model, but splitting "procedural memory" into two spaces to compress to the student. Specifically dual-space distillation: in the context space, the procedurally generated memory from the teacher is directly prepended to the student's input, equivalent to giving the student a cheat sheet with answers; in the parameter space, the student then fine-tunes a lightweight LoRA on the teacher's successful trajectories — trainable parameters under 10M, with only a few MB of teacher memory added. The effect is quite illustrative. On the embodied-decision benchmark ALFWorld, the 4B student model's task success rate jumps from 4.3% to 77.9%, basically catching up to the 72B teacher's 87.1%; wall-clock is 3×+ faster than the 72B teacher, truly with the feasibility of real-time on-device deployment. The authors ran an ablation on 8 models (2B–72B), confirming that the two spaces complement each other, neither is dispensable. For mobile, in-car, robotics, and other on-device Agents, this path is more realistic than simply stacking parameters — it essentially splits "teacher's procedural knowledge" into attachable annotations + trainable muscle memory for delivery. The short board is also clear: procedural memory requires the teacher to pre-"scout" on similar tasks to generate; if scenarios diverge quickly, memory maintenance costs will balloon; and the 4B model's ultimate boundary still depends on the underlying instruction-following capability — DuoMem solves "how to move knowledge", not "how to fill capability".