Large language model (LLM) agents rely on external memory and past experience to handle long‑horizon tasks, yet most only retrieve records or summaries and lack causal understanding of the world.
Inspired by human causal cognition, we introduce CauProphet. It maintains a Causal Cognition store that accumulates causal facts across tasks. For each new task, the agent selects relevant facts, combines them with current evidence, and constructs a Causal Belief—a structured representation of causal factors and the effects of possible actions.
The belief guides decision‑making; after each observation the agent updates both the belief and the underlying Causal Cognition, allowing causal knowledge to persist beyond a single episode.
Evaluations show that CauProphet progressively acquires causal knowledge and transfers it to future decisions. In a financial benchmark, it reuses discovered causal patterns, achieving the highest mean rank correlation with next‑day returns—a 25.4% improvement over a baseline agent—and raising prediction accuracy from 21.3% to 28.6% in a dynamic setting.
Additional experiments on scientific reasoning and everyday scenarios confirm reliable causal analysis and action selection, highlighting causal experience learning as a promising direction for next‑generation LLM agents.
Review