Abstract Large language model (LLM) agents reason and act in dynamic environments but often lack an explicit account of what they know or believe at each step. This paper introduces a lightweight epistemic state management framework inspired by public announcement logic (PAL) and graded modal logic. We treat observations, tool results and internal inferences as announcements that narrow an agent’s knowledge state. Each announcement updates a minimal knowledge base that serves as a filter on the agent’s possible worlds, maintaining consistency while avoiding full model enumeration. Our concrete update mechanism is a consistency-preserving knowledge-base operation that draws on PAL’s model-restriction semantics as conceptual motivation. We survey graded and plausibility-based extensions for representing uncertainty and discuss how qualitative plausibility orderings can model belief revision under bounded rationality. The framework enables explicit, inspectable reasoning in LLM agents, improving self-consistency, interpretability and epistemic awareness. We provide implementation sketches with working code and discuss evaluation strategies for benchmarking epistemic reasoning in LLM agents.
Mullins et al. (Mon,) studied this question.