Ant International Research Highlights How Readable AI Memories Can Unlock More Personalised, Privacy

Human sales and customer representatives have institutional memories to serve clients: their preferences and choices. How can we build a memory mechanism for AI services to create similarly personalised experiences?

Ant International’s research on this subject recently received the Senior Area Chair (SAC) Highlight Award from the Best Paper Committee at EMNLP 2026, a leading global AI conference.

Why do we need “readable AI memory”?

In this paper titled 'Text as a universal interface for transferable personalisation', we look into one fact: until now, AI memory of a user has usually been generated and stored as dense numerical code only the original system can interpret.

Two long-standing problems with that: First, today's user profiles are "black boxes" — opaque code locked inside a single app, so activity in one service teaches nothing to another, and every product has to re-learn you from scratch.

Second, profiles are welded to the model that created them, so they don't transfer when products or models change.

How do we create “readable AI memory”?

Our approach treats natural language as a universal interface: the preference summary is plain text, so any model or service can read, audit and act on it.

Training a model to write good preference summaries is the key here. To make those summaries genuinely accurate, we trained the model in two stages. First, generating candidate profiles and stress-testing them against users' actual future choices, keeping only those that prove correct. Second, reinforcing summaries that stay useful as they're updated over time.

We developed AlignXplore+, a universal preference reasoning model, which is trained to log a user’s preferences through short, plain-language summaries, continuously updated as new interactions happen, or a “readable AI memory”.

What does this research mean for AI and fintech?

This research shows that these preference summaries can transfer across different tasks and models effectively. Moreover, at 8bn parameters, our model is relatively compact, but achieved an average score of 75.10% across nine benchmarks – a significant advantage over an open-source model more than twice its size. This shows that the quality of the representation (how well the summaries are constructed) matters more than the size of the model.

The findings also present several practical upsides for data privacy, fairness and consumer trust.

With a universal data interface, personalisation no longer depends on collecting or moving raw behavioural data between systems, and only the distilled, readable summary travels. That supports personalisation across an ecosystem, from payments to mini-programmes to merchant platforms, in a privacy-considerate way. Because the profile is expressed as text, it can also be audited, explained and corrected, which makes it more transparent and trustable.

Second, it means smaller, more efficient models can deliver excellent personalisation when paired with a strong memory, rather than relying on the largest possible model. That has real implications for cost and speed for businesses.

What are some practical applications of these insights?

First, a merchant or wallet could offer genuinely seamless, personalised experiences across every touchpoint — recommendations, product offers, an AI assistant — driven by one coherent, portable view of the customer.

Second, a readable preference profile supports explainability and governance, because humans can review and edit it.

Third, it offers a more efficient route to personalisation on smaller models, which matters for latency and operating cost at scale.

Imagine you’re planning a trip. Today, even with AI assistants, that may involve explaining your needs and preferences repeatedly to different platforms for flights, hotels and tours.

A readable and secure AI memory of you simplifies and connects these needs. Booking flights and hotels? Your AI assistant logs your preferences in plain language — "prefers window seats, is a member of these frequent flyer programs, prefers hotels within 40 minutes ride to the airport." As the journey progresses, this readable profile updates automatically.

When you hail a ride via another app, ride types, promotions and AI’s suggestions pop up tailored to your preferences, even though the profile was built in a different service by a different model. No behavioural data was shipped around, and no giant models required. The consumer simply feels understood from the first tap across connected services.

All in all, a transferrable, secure AI memory mechanism provides richer, more trusted and more inclusive experiences for consumers.