The Missing Knowledge Layer in Cognitive Architectures for AI Agents
Summary: arXiv:2604.11364v1 Announce Type: new
Abstract: The two most influential cognitive architecture frameworks for AI agents, CoALA and JEPA, both lack an explicit Knowledge layer with its own persistence semantics. This gap produces a category error: systems apply cognitive decay to factual claims or treat facts and experiences with identical update mechanics. We survey persistence semantics across existing memory systems and identify eight convergence points, from Karpathy’s LLM Knowledge Base to the BEAM benchmark’s near-zero contradiction-resolution scores, all pointing to related architectural gaps. We propose a four-layer decomposition (Knowledge, Memory, Wisdom, Intelligence) where each layer has fundamentally different persistence semantics: indefinite supersession, Ebbinghaus decay, evidence-gated revision, and ephemeral inference respectively. Companion implementations in Python and Rust demonstrate the architectural separation is feasible. We borrow terminology from cognitive science as a useful analogy (the Knowledge/Memory distinction echoes Tulving’s trichotomy), but our layers are engineering constructs justified by persistence-semantics requirements, not by neural architecture. We argue that these distinctions demand distinct persistence semantics in engineering implementations, and that no current framework or system provides this.
Introduction
Cognitive architectures are essential for the development of advanced AI agents, yet significant gaps remain in their foundational structures. Recent research highlights a critical oversight in the most widely used frameworks, CoALA and JEPA, which lack a distinct Knowledge layer. This omission has profound implications for how AI systems manage and update factual information.
The Problem with Current Architectures
Both CoALA and JEPA apply uniform cognitive decay to various types of information, treating factual claims and experiential knowledge with the same update mechanisms. This leads to cognitive errors, undermining the reliability and accuracy of AI systems. By failing to segregate knowledge from memory, these architectures risk conflating distinct types of information, ultimately affecting performance and decision-making.
Persistence Semantics Survey
In our investigation of persistence semantics across existing memory systems, we identified eight convergence points that underscore the relevance of this issue. Key findings include:
- Karpathy’s LLM Knowledge Base highlights the need for structured knowledge management.
- The BEAM benchmark reveals alarming contradiction-resolution scores, suggesting poor handling of factual discrepancies.
- Current frameworks do not adequately address the unique requirements of knowledge persistence.
Proposed Four-Layer Decomposition
In response to these challenges, we propose a novel four-layer architecture that separates Knowledge, Memory, Wisdom, and Intelligence. Each layer is designed with distinct persistence semantics:
- Knowledge: Indefinite supersession allows for the most reliable factual claims to remain unaffected by decay.
- Memory: Ebbinghaus decay models the natural forgetting process of experiences over time.
- Wisdom: Evidence-gated revision incorporates new information while considering existing beliefs.
- Intelligence: Ephemeral inference focuses on transient reasoning based on current context.
Implementations in Python and Rust validate the feasibility of this architectural separation, demonstrating that distinct persistence semantics can be realized in engineering practice.
Conclusion
The lack of a dedicated Knowledge layer in cognitive architectures presents a significant barrier to the evolution of AI agents. By addressing this gap with a structured four-layer model, we can enhance the reliability and functionality of AI systems. Future research and development must prioritize these distinctions to ensure that AI agents can effectively manage knowledge and memory, ultimately leading to more intelligent and capable systems.