DuoMem introduces a dual-space distillation framework for deploying memory-augmented agents on constrained devices. It combines context-space distillation, which prepends teacher-generated procedural memories to the student input, with parameter-space distillation using lightweight LoRA adapters trained on successful teacher trajectories. On ALFWorld, the method raises a 4B model’s task success rate from 4.3% to 77.9%, approaching the 72B teacher’s 87.1%. It uses fewer than 10 million trainable parameters, requires only a few megabytes of precomputed memories, and reportedly completes tasks more than three times faster than the teacher.
No heat snapshots are available in the last 24 hours.