Beyond Clicks: Designing AI-Driven User Memory for Personalization

Anil Roy30 Jul 2025

Problem

Home shoppers search for months and change their minds along the way. A buyer may browse condos in Seattle, then three weeks later save townhomes in another state. A profile built once keeps recommending last month's interests. Recommendations, search ranking, notifications and emails must follow the shopper's current intent without forgetting long-term interests.

Example from the article: a user whose long-term goal is a home in Seattle still browses other cities, and some of those exploratory views later become durable preferences.

Solution in brief

Zillow treats personalization as a "user memory" made of several layers: recency-weighted preference profiles, affordability estimates, user embeddings and the raw event history. Long-term profiles are rebuilt daily in batch. Short-term state is updated within seconds from an event stream. Downstream systems read both.

Questions this case answers
  • How to represent user preferences that drift over weeks while still reacting to changes within a single session
  • How to serve both slowly changing long-term and fast short-term user profiles to ranking, notification and email systems
  • How to keep a user's preference profile current when their goals shift from week to week?
  • How to combine long-term and short-term user signals in one personalization system?

Tools and methods

  • Preference profiles — per-user likelihood that the user prefers a given value of price, location, bedrooms, home type
  • Recency weighting — gives recent interactions a higher weight, so profiles follow the current search
  • Affordability-aware quantile models — estimate budget and price tolerance
  • User embeddings — capture tastes that structured attributes miss
  • Raw event history — timestamped views, saves, searches and filter applications
  • Batch plus near-real-time pipelines — separate long-term and short-term profile stores
Continue reading — it's free

Google or a one-time email code. No password.