Why do LLM agents remember preferences but not act on them?

The paper separates two things that memory benchmarks usually run together. It administers "paired Know and Act tests to the same user preference": a recall test, and a behavioral scenario where that preference should shape the response. Across 16 systems, five memory architectures, and 1,000 preferences embedded at three levels of expression strength, the authors report "a large gap between Know and Act outcomes." Agents "often pass the recall test" but "fail to reflect that same preference in the paired behavioral scenario." Memory architectures reduce the gap without closing it, and utilization "remains especially weak for health and therapy-related preferences," which the authors flag as the cases where failing to act carries the greatest real-world stakes.

The paper's name for the underlying problem is a "knowledge utilization problem": a model may fail to act on a preference "even when [it is] fully present in context." When a response is not tailored, an evaluator cannot tell whether the model failed to remember or remembered and failed to use, and the paired design exists to tell those apart. The failure attribution in the conclusion then splits the errors into three stages, retrieval, comprehension, and application, and reports that "comprehension failures dominate overall." On this account the bottleneck is less about getting the preference into context than about interpreting what it means for the present request. The authors conclude that personalization systems should ensure memories are "selectively retrieved, correctly interpreted, and reliably translated into context-appropriate behavior."

This sits close to Why do LLM agents ignore condensed experience summaries?, which also finds that information available to an agent does not reliably change what it does. What this paper adds is a design that isolates the point of loss: same preference, two tests, so recall accuracy and behavioral uptake can be compared directly. It also extends the argument in How should we actually evaluate agent memory systems? by adding a stage after retrieval, where stored content has to be applied. And it qualifies Does retrieved memory quality depend on its functional role?: getting the right memory retrieved is necessary, but the response can still ignore it. The reduction the paper reports from memory architectures should not be read against Do memory systems actually help language models learn continuously?, since the excerpt does not say which architectures narrow the gap or how they compare with the long-context baselines it also evaluates.

The excerpt is silent on most of what would make the result actionable: how large the Know-Act gap is in numbers, how it moves across the three expression levels beyond "varies," which architectures help most, how comprehension failure was operationalized, and what share of errors falls in each of the three stages beyond comprehension being the largest. It also does not say whether health and emotional preferences are harder to interpret or simply harder to translate into behavior. What it does support is narrower and still useful: a passing recall score is not evidence that a personalized agent will act on what it recalled, so behavioral tests belong beside recall tests, with particular care in health and therapy contexts.