When AI Memory Knows Your Topic, Not Necessarily You

A test with two moving parts

The preprint Not Your Diary, Just Your Topic: Grading the Wrong-User Control, by Karina Romanova and colleagues, investigates a common personalization test: give a model another user’s history and measure how much its performance falls.

The difficulty is that replacing the person usually replaces the topics too. A history about politics and one about sport differ in more than ownership. Worse performance with the second history does not, by itself, establish that the model used information unique to the first person.

The researchers therefore made replacement histories progressively harder to distinguish by content. Their question was not simply whether history mattered, but how much advantage remained when someone else’s history covered similar ground.

Making the comparisons fairer

The main experiment used an older news-categorization version of the LaMP-2 benchmark: 15 categories, 1,605 validation examples and 323 validation users. Each history was restricted to its first six items.

A frozen Qwen2.5-3B-Instruct language model received history information compressed into eight numerical vectors. Trainable components produced and supplied those vectors, while the language model itself remained unchanged. The main control experiments used three training seeds and held each trained model fixed while changing its history input.

Comparisons included the user’s own history, a random other user’s history, a similar user’s history, synthetic histories assembled from other users’ articles, and histories masked out during evaluation. Some synthetic histories also matched the category at each original history position. Their construction did not use the current query’s answer.

Crucially, the team separately trained a model without history. That distinguishes a system’s dependence on an expected input from the benefit of providing that input at all.

A large swap penalty, a smaller benefit

With the correct history, average accuracy was about 79.4%. Replacing it with a random user’s history lowered accuracy by 10.1 percentage points. Yet the advantage over the separately trained no-history model was only 2.3 points.

Masking history only at evaluation produced a larger apparent benefit of 4.8 points. The distinction matters: a model trained to expect history may struggle when that input disappears, whereas one trained without it can learn to operate differently.

Matching replacement content narrowed the gap. A similar real user’s history left a 3.6-point advantage for the original history. Synthetic histories closely matched in embedding space—where numerical representations capture content similarity—reduced it to about 1.2 points. Matching categories as well left roughly 0.9 points.

Category information alone also carried substantial signal. Histories containing only category labels reproduced 86% of the accuracy gap between the user’s own history and a random donor’s. This does not mean categories explain 86% of personalization causally: these interventions do not separate topic, coherence and preferences cleanly.

Even the remaining 0.9-point difference was sensitive to how uncertainty was estimated. Its bootstrap interval excluded zero, but an interval based on the three seed-level results included zero.

Familiar users versus new users

The original news split substantially overlapped users between training and validation. The researchers therefore divided users into two groups, trained on each group separately and evaluated both familiar and held-out users across three seeds.

The advantage over a similar user’s history averaged 3.2 points for users seen during training, but only 0.65 points for unseen users. The uncertainty interval for unseen users included zero. Strong synthetic replacements likewise left no clear advantage there, although random replacements still caused a substantial penalty.

That is inconclusive evidence for a small benefit, not proof that user-specific information is absent. These fold models also trained on fewer users, and no separate no-history model was trained within each fold.

A one-seed pilot on the current movie-tagging version of LaMP-2 found no observed advantage for the correct history over similar or synthetic histories. Four other LaMP tasks showed small contrasts comparable to variation between training seeds under the tested setup.

What the study can support

A separate analysis warned against claiming that personalization benefits concentrate in a few users. With about five evaluation items per person, apparently concentrated effects were largely compatible with sampling noise, although exploratory evidence of variation between users remained.

The team also found that adding gated cross-attention to a prefix-based memory model improved accuracy under its training recipe. Those models retained the prefix and had different parameter budgets, so the result does not establish a universally superior architecture.

The central limitation is scope: an older news task, short histories, a small number of seeds and imperfect content matching. Deployment usefulness, latency and privacy were not tested. The practical lesson is nevertheless clear: a wrong-user penalty alone is insufficient evidence of user-specific memory. Similar-content replacements, unseen-user evaluation and a model trained without history provide more informative comparisons.