We propose EPIC, a preference-aligned memory construction framework for on-device RAG that selectively retains preference-relevant information and aligns retrieval toward user preferences under tight memory budgets.

We present PastForward, a system that accelerates on-device GUI agents by reusing computational experience accumulated during ordinary task execution: it retrieves prior output sequences as multi-token proposals verified in a single VLM forward pass, and starts next-step inference from prior GUI transitions while the current action executes, achieving 1.63–2.36× action-step latency speedups on device while maintaining task success rates.