TriggerBench: Investigating Prospective Memory for Large Language Models

While Large Language Models (LLMs) have become adept at retrieving past information when explicitly asked (Retrospective Memory, or RM), they often struggle to act on latent constraints without direct prompts. This ability—remembering to perform a future action based on a specific trigger—is known as Prospective Memory (PM). TriggerBench is a new benchmark designed to evaluate this capability across five dimensions, including state-tracking, temporal grounding, and safe coding.

The study reveals that PM is fundamentally harder than RM. While models can maintain near-perfect accuracy for RM tasks even at 100K tokens, their performance on PM tasks drops sharply as context length scales. This suggests that without an explicit query to act as an attentional anchor, models struggle to maintain situational awareness over long interactions. Furthermore, the research identifies a precision-recall trade-off: highly proactive models often overfit to an "always-remind" heuristic, leading to high false-alarm rates in scenarios where no intervention is actually required.

Enhanced reasoning capabilities, such as those found in "thinking" models, significantly improve proactive recall. However, this comes at the cost of attentional fragility. When triggers are overloaded with concurrent user requests, model performance degrades substantially. The study also demonstrates that PM serves as a behavioral probe for spare reasoning capacity; successful reasoning trajectories correlate with higher PM accuracy, suggesting that PM tracks the cognitive budget that simple token counts often obscure.

As LLMs are increasingly deployed as autonomous agents in professional and daily workflows, the ability to remember latent constraints—such as safety protocols or temporary debugging shortcuts—is critical. Current RAG and memory-augmented systems often fail at these tasks because they rely on semantic similarity, which is frequently absent between a latent constraint and its future trigger. TriggerBench provides a rigorous framework to measure these failures and highlights the need for active working memory architectures that can maintain situational awareness beyond simple retrieval.