Continuity of Mind: Grounded, Memory-Driven Cognition and Functional Consciousness in Language Agents

ACL ARR 2025 May Submission7835 Authors

Abstract:

We introduce a cognitively inspired agent architecture that enables \textit{continuity of mind} in large language model (LLM) agents—maintaining coherent, grounded cognition over extended interactions through structured memory and modular control. At the core of this architecture is Episodex, a context management strategy (CMS) that segments reasoning into semantically coherent episodes and abstracts conceptual knowledge for symbolic reuse. Episodex operates within a meta-agent framework inspired by Global Workspace Theory (GWT), supporting dynamic attention regulation, memory retrieval, and deliberation. By integrating episodic and conceptual memory with grounded control loops and hallucination recovery mechanisms, the system supports memory-driven cognition beyond the token limits of transformers.

Evaluated in the ALFWorld environment, our framework achieves a success rate of 87.5% on out-of-distribution tasks using GPT-4o. Memory-driven retrieval reduces the average number of actions per successful task by over 30%, while conceptual memory clustering boosts performance from 50% to 82%. We argue that combining modular memory, perceptual grounding, and symbolic abstraction enables a practical form of functional consciousness in LLM agents—supporting scalable, goal-directed reasoning in interactive environments.

Paper Type: Long

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Research Area Keywords: global workspace theory, functional consciousness, cognitive modeling, episodic memory, conceptual memory, attention and working memory, cognitive control in agents, dual-process reasoning, psycholinguistic modeling

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Theory

Languages Studied: English

Submission Number: 7835