Building an AI That Remembers You: My Internship Journey at AIGility Cloud Innovation

Four months ago, I walked into AIGility Cloud Innovation(智悦云创), an fast-moving AI startup. Today, as my internship comes to a close, I’m proud to say we’ve built something that not only works but cares.

From February to June 2025, I served as a Large Model Algorithm & Backend Engineering Intern, embedded directly in the founding team of our flagship product: the Resume Optimization Assistant. We also proposed TiMem (Temporal-Hierarchical Memory Consolidation).

Inspired by how human memory consolidates experiences—from fleeting moments to stable self-concepts, we designed a five-layer temporal memory tree (TMT) architecture. At its core, TiMem automatically organizes raw dialogue fragments (L1) into session summaries (L2), daily behavioral patterns (L3), weekly trends (L4), and finally, a coherent user persona profile (L5). Crucially, retrieval is adaptive: simple factual queries pull from lower layers; identity-aware questions activate the high-level L5 representation.

The result

• 75.3% accuracy on the LoCoMo long-memory benchmark
• 52.2% reduction in context token usage vs. flat RAG or Mem0

Most importantly, TiMem became the memory backbone of our product. Users now get consistent, personalized advice across weeks of interaction—something that felt impossible just months ago.

None of this would have been possible without the incredible team at AIGility. Thank you to my mentor, Xuanqing Yu, for trusting me.