Artificial Cognitive Design: memory, time, interoception and continuity for long-lived AI agents
Year: 2026
Type: preprint
Abstract: We have automation bots that run for years with limited intelligence, and language models that solve hard problems but rarely last beyond a session. Between them is a gap that matters more each month as agents move from answering questions to doing jobs: agents that are intelligent and also last, working for months, acting on their own or controlling a physical body. Such an agent has to remember, understand time, sense its own condition, and remain itself. Today these are solved imperfectly, with patches. Upgrade the model underneath an AI agent and it may come back thinking and judging differently, with no established, thorough way to check, let alone ensure, continuity. I argue that these abilities should be deliberately designed, and I introduce artificial cognitive design (ACD): the design of cognition for artificial agents, starting from the goals they pursue and combining machine-native mechanisms with designs inspired by nature, human cognition above all, where nature does better. Intelligence is the foundation of this cognition, but the rest has to be designed. These abilities carry a responsibility. Agents that have memory, a sense of time, interoception and continuity, and that may even take part in their own design and growth, could come to be more than tools. ACD therefore considers ethics from the start, toward every intelligence, organic or machine.
Author: Su, Tian
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225348085