MemoryCD: Benchmarking Long-Context User Memory of LLM Agents for Lifelong Cross-Domain Personalization

A research team led by Weizhi Zhang has introduced MemoryCD, a groundbreaking benchmark designed to rigorously evaluate the long-context memory capabilities of Large Language Model (LLM) agents. Unlike previous benchmarks that relied on short, synthetic dialogues, MemoryCD is built from the massive Amazon Review dataset, tracking authentic user interactions across years and multiple product domains. This creates the first large-scale, user-centric testbed for lifelong personalization, moving beyond scripted personas to real-world behavioral data.

The benchmark employs a comprehensive evaluation pipeline, testing 14 state-of-the-art LLM base models—like GPT-4, Claude 3, and Llama 3—alongside 6 established memory method baselines. These are assessed on 4 distinct personalization tasks across 12 diverse domains, from books and electronics to clothing and home goods. The analysis reveals a significant gap: current memory methods are "far from user satisfaction" in accurately simulating real user behaviors, especially in cross-domain settings where an agent must connect preferences from different areas of a person's life.

Published as a workshop paper at Lifelong Agent @ ICLR 2026, MemoryCD establishes a crucial new standard. It provides developers and researchers with the tools to measure how well an AI agent can build and utilize a persistent, evolving memory of a user, which is essential for creating truly personalized assistants, recommendation systems, and customer service bots that learn and adapt over time.