Developing an on-device, cross-application mobile companion agent that transforms unstructured phone usage data into structured user profiles—under resource-constrained conditions—is a critical step toward evolving intelligent systems from conversational "instant tools" into true "companion partners." However, current mainstream AI agents perform well in instruction following but lack the ability to understand implicit user intent or model long-term memory for specific individuals. This project pursues two key technical objectives: (1) on-device adaptive long-term memory construction, which extracts high-value structured information from massive unstructured usage data to build durable user profiles; and (2) hybrid retrieval and intent understanding based on long-term memory, which enables the agent to accurately interpret vague user intentions and retrieve relevant information to support complex tasks. By overcoming key limitations in cross-application memory and retrieval, this project aims to shift the paradigm from tool-based agents to genuinely companion-like agents, while addressing users' growing demands for personalized, memory-aware intelligent assistance.