A research paper published on arXiv has identified a fundamental architectural gap in current cognitive frameworks for AI agents. According to the paper by Michaël Roynard, 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 omission, the author argues, produces a category error: systems apply cognitive decay to factual claims or treat facts and experiences with identical update mechanics.
Roynard surveyed persistence semantics across existing memory systems and identified eight convergence points pointing to related architectural gaps. These include Karpathy's LLM Knowledge Base and the BEAM benchmark's near-zero contradiction-resolution scores. The paper proposes 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.
The Gap in Current Architectures
The paper notes that current frameworks treat factual knowledge and experiential memories with the same update mechanics, leading to inconsistencies. For example, factual claims may undergo cognitive decay designed for episodic memories. The BEAM benchmark, cited in the paper, reports near-zero contradiction-resolution scores, highlighting the difficulty of maintaining consistent knowledge without a dedicated layer.
Proposed Four-Layer Decomposition
Roynard's model separates layers by their persistence requirements:
| Layer | Persistence Semantics | Description |
|---|---|---|
| Knowledge | Indefinite supersession | Facts are replaced but never decay; they persist until explicitly superseded |
| Memory | Ebbinghaus decay | Experiences fade over time based on a forgetting curve |
| Wisdom | Evidence-gated revision | Beliefs are updated only when new evidence reaches a threshold |
| Intelligence | Ephemeral inference | Inferences are temporary and discarded after use |
The author borrows terminology from cognitive science (echoing Tulving's trichotomy) but emphasizes that the layers are engineering constructs justified by persistence-semantics requirements, not by neural architecture.
Implementation and Feasibility
To demonstrate that the architectural separation is feasible, companion implementations were provided in Python and Rust. The paper argues that these distinctions demand distinct persistence semantics in engineering implementations, and that no current framework or system provides this.
The research has implications for developers building AI agents that need to distinguish between encyclopedic facts, personal experiences, evolving beliefs, and transient reasoning. Without an explicit Knowledge layer, systems risk confusing these categories, leading to unreliable behavior in applications like customer support, research assistance, or autonomous decision-making.
