Beyond Clicks: Designing AI-Driven User Memory for Personalization
Today’s home shoppers don’t just want a catalog of listings. They expect platforms to understand and guide them toward a more efficient home-shopping experience. To do that, the personalization system must remember what they liked, anticipate what they’ll want next and, most importantly, adapt as their needs evolve.
At Zillow, we’re seeing this shift firsthand. A buyer might start out browsing condos in Seattle, but three weeks later they’re saving townhomes in a different state. Static personalization systems struggle to keep up with these dynamic journeys, which is why we’re rethinking personalization, not merely as a product feature but as a living memory.
The future of real-estate discovery will be shaped by personalization systems like these that not only remember your preferences, but also adjust to your shifting priorities and help you make complex, months-long decisions with confidence. To get there, we need more than better models. We need richer, more intelligent user memory.
What is “user memory” at Zillow?
User memory is the evolving, context-rich understanding of what each person values during their home shopping journey. It’s not just a log of clicks or page views; it’s a layered, AI-driven profile of preferences, priorities and intent — all evolving over time.
At Zillow, we’re building this memory across three dimensions:
- Recency and Frequency: What users care about right now vs. what they’ve consistently cared about
- Flexibility and Rigidity: Which preferences are must-haves vs. where they’re open to exploring
- Prediction and Planning: How their preferences are likely to shift over time
This memory is shaping search, recommendations and other personalized experiences on Zillow through real-time preference modeling, affordability estimations and user embeddings.
Powering personalization through deep consumer understanding
Understanding what users care about, and how those preferences evolve, is key to helping them navigate the home shopping journey with confidence. At Zillow, we’ve built systems that not only interpret behavioral signals, but also preserve meaningful patterns over time, enabling continuity and relevance across sessions and surfaces.
From recency-aware preference profiles to affordability-aware quantile models and deep user embeddings, these components form the foundation of a memory layer that powers personalized, responsive and consistent experiences.
We’re continually evolving this memory infrastructure to feel more intuitive and human. That means:
- Capturing the right level of nuance (what’s a strong signal vs. a soft preference)
- Adapting gracefully to changes in behavior and intent
- Preserving helpful context across different touchpoints of the consumer’s journey while avoiding stale assumptions
That user understanding is what powers personalization at Zillow, and it’s central to how we’re helping millions of buyers and renters feel seen, supported and confident as they shop for their next home.
Building blocks of long- and short-term user profiles
Designing personalization at scale means asking a key question: how do we represent what users care about, and how that changes over time?
Prospective buyers and renters typically exhibit both long- and short-term behavioral patterns. For instance, while a consumer’s primary long-term goal might be to purchase a home in Seattle, they may still browse listings in other cities. Over time, some of these exploratory short-term actions may solidify into durable, long-term preferences as the consumer reconciles market conditions with personal affordability constraints.
Batch user-profile pipelines are well suited for modeling long-term preferences because they operate over extended observation windows and can employ higher-latency, more complex algorithms. Since long-term preferences evolve slowly, certainly not hour by hour, it is efficient to generate these profiles once per day via batch jobs.
On the other hand, pipelines that model short-term profiles must detect transient shifts in behavior, so they benefit from low-latency, near-real-time processing. Streaming architectures are therefore better choices for capturing these rapid changes.
At Zillow, we approach this by combining batch with real-time data pipelines to capture both long-term intent and short-term signals. These systems power everything from homepage personalization, to push notifications, to search ranking.
The foundation: Preference profiles
We use preference profiles to summarize what kinds of listings each user engages with, by price, location, number of bedrooms, home type and more. Each profile is built from recent user interactions and captures the likelihood that a user prefers a certain value.
For example, a user might interact most with listings between $750K–$900K, or show strong interest in homes with 3+ bedrooms in a specific ZIP Code. These signals become input features for recommendation models, powering personalized ranking and targeted experiences.
Recency weighting
Most user preferences shift over time. That’s why we use recency weighting to give more recent consistent interactions higher weight. This makes our models responsive, surfacing townhomes in Oakland if that’s what a user started browsing this week, even if they were looking in San Jose last month.
The user profile decay is currently based on absolute recency, but can be tuned at multiple levels based on position in the interaction sequence. These tunable decay strategies let us construct both short-term and long-term preference representations, which are used in parallel today across different personalization surfaces. As we continue evolving our memory systems, we see opportunities to more deeply blend these temporal signals, creating unified profiles that dynamically balance stability with recency.
Understanding constraints
While preference profiles work well for understanding what users value, they can miss nuance in the affordability and other constraints. At Zillow, we understand user’s preferences based on their interaction in the app/website, as well as what they can afford, to have a 360-degree view of their needs. We’re also exploring how other sources of context, such as nuanced preferences that shoppers share with their agents, can further enrich this understanding. Together, these signals allow us to move beyond static filters toward a more adaptive and personalized experience. This helps us answer questions like:
- Are they browsing within or beyond their likely budget?
- How wide is their price tolerance?
- What listings might stretch their range, but still feel relevant?
These signals are especially useful in constrained markets, helping us recommend listings that fit both the user’s preferences and their financial realities given market constraints.
Embedding models: Going beyond structured preferences
Users often have interests toward certain features that cannot be effectively captured by structured preference profiles, especially with nuanced features like their interest toward textured walls, or types of trees in the backyard. Embeddings complement preference profiles by filling in gaps, surfacing surprising-but-relevant options, and adding more semantic understanding to our personalization systems.
Raw interaction history: Supporting continuity and transparency
While models help us infer and generalize user preferences, there are moments where the most useful memory is simply an accurate replay of what happened. That’s where raw interaction history comes in.
We retain timestamped event sequences for each user, including views, saves, searches and filter applications. This raw behavioral history supports critical experiences that depend on continuity and transparency, such as recently viewed homes, allowing users to revisit without retracing their steps.
Powering memory through batch and real-time infrastructure
Personalization isn’t static. It’s a living system that needs to evolve as users explore new areas, change priorities, or return after a break. To support this dynamic behavior, Zillow uses a dual-layered architecture that blends batch processing with real-time updates. This hybrid setup enables us to maintain accurate, responsive user memory at scale. Our batch systems process engagement data daily to build long-term user profiles. These pipelines are designed to:
- Aggregate user activity across features like price, bedroom count and location
- Generate stable, preference-aware representations from multiple sessions
Real-time signals: Capturing shifts instantly
On the flip side, user preferences can change in a matter of minutes. A user might switch from browsing downtown condos to saving single-family homes in the suburbs, all in a single session.
To handle these shifts, Zillow’s near real-time infrastructure ingests behavioral signals (views, saves, searches, filter updates) and updates user state within seconds. These signals flow through:
- Online Data Platform tables for near-instant access
- Lightweight, in-session models that complement batch profiles
- Event-driven triggers used by recommendations, alerts and personalization surfaces
Together, this infrastructure ensures the system reflects what users are doing right now, not just what they’ve done before.
Blending both for a seamless experience
Rather than treating batch and real-time as separate modes, we leverage them side-by-side to power a consistent, responsive experience:
- Long-term profiles shape relevance across time
- Real-time signals tune personalization moment-to-moment
- The same modeling logic often runs in both pipelines with different tuning
While these signals are typically processed through separate pipelines, they are consumed together by downstream systems that enable us to preserve stability, while responding to shifts in behavior. This approach allows Zillow to personalize across all contexts, whether someone opens the app after a week away or refreshes a page five times in one session.
Trust-centered personalization
The foundation of effective personalization is user trust. At Zillow, we believe that giving consumers a more tailored experience should never come at the expense of their privacy, control or data rights.
That’s why our personalization infrastructure is designed to:
- Respect user choices and preferences around data usage
- Comply with privacy regulations like the California Consumer Privacy Act (CCPA)
- Protect identities through secure, privacy-aware data handling
- Support transparency and control for users who want to manage their experience
We are committed to building systems that are not only intelligent, but also responsible, compliant and user-first, so home shoppers feel both understood and respected at every step of their journey.
The path ahead: Personalization that feels truly personal
The next wave of personalization isn’t just about better algorithms, it’s about building systems that genuinely understand our consumers. At Zillow, we’re building a memory-first personalization infrastructure that adapts as users’ needs evolve, recognizing what they’re looking for and why.
It takes more than one type of data to get there. We’re blending structured user profiles, behavioral signals, semantic understanding and affordability insights to shape a home shopping journey that feels less like a machine and more like a guide.
This isn’t just a technical challenge, it’s a trust-centered mission. Whether someone is actively browsing, coming back after time away or shifting their priorities mid-search, our goal is to meet them where they are, with a system that adapts and supports them every step of the way.