The Missing Knowledge Layer in Cognitive Architectures for AI Agents

The article discusses the absence of a dedicated Knowledge layer in the cognitive architectures CoALA and JEPA, which leads to mismanagement of factual claims and experiences. It proposes a four-layer architecture (Knowledge, Memory, Wisdom, Intelligence) with distinct persistence semantics, such as indefinite supersession and Ebbinghaus decay, addressing the identified architectural gaps. This framework emphasizes the need for differentiated persistence mechanisms in AI systems, highlighting a critical area for improvement in current architectures that could enhance the reliability and efficacy of AI agents.