The Missing Knowledge Layer in Cognitive Architectures for AI Agents

Researchers identify a critical architectural gap in leading AI agent frameworks (CoALA and JEPA), which lack an explicit Knowledge layer with distinct persistence semantics. The paper proposes a four-layer decomposition model with fundamentally different update mechanics for knowledge, memory, wisdom, and intelligence, with working implementations demonstrating feasibility.

Current cognitive architecture frameworks treat all information with identical persistence mechanics, creating fundamental engineering problems. CoALA and JEPA apply cognitive decay indiscriminately to both factual claims and experiential memories, conflating systems that require opposite behaviors—facts should persist indefinitely unless superseded, while experiences should fade according to Ebbinghaus decay curves. This architectural confusion manifests across eight convergence points in existing memory systems, from Karpathy's knowledge base proposals to BEAM benchmark contradiction-resolution failures.

The research addresses a longstanding oversight in AI agent design stemming from frameworks that prioritize unified cognitive models over engineering requirements. As AI systems scale toward practical deployment, the inability to distinguish knowledge persistence from memory decay becomes increasingly problematic. An agent facing conflicting information sources needs different resolution strategies for factual contradictions versus experiential updates.

The proposed four-layer model—Knowledge (indefinite supersession), Memory (Ebbinghaus decay), Wisdom (evidence-gated revision), Intelligence (ephemeral inference)—offers practical separation of concerns. Knowledge layers maintain factual consistency, memory layers simulate human forgetting, wisdom layers balance new evidence against accumulated experience, and inference layers support temporary reasoning without persistence. Python and Rust implementations validate this isn't purely theoretical.

For AI development teams building production systems, this framework provides actionable architecture guidance. Organizations developing autonomous agents, reasoning systems, or long-horizon decision-making tools currently lack standardized approaches to information persistence. The research creates vocabulary and engineering patterns for distinguishing when information should update permanently, decay gradually, or remain ephemeral. This becomes critical as systems require consistent factual grounding while maintaining appropriate uncertainty bounds.

• →Leading AI agent frameworks conflate knowledge and memory, applying identical update mechanics to information requiring opposite persistence behaviors
• →Eight convergence points across existing systems identify architectural gaps from knowledge bases to contradiction-resolution benchmarks
• →Proposed four-layer decomposition separates indefinite supersession, exponential decay, evidence-gated revision, and ephemeral inference into distinct engineering constructs
• →Working implementations in Python and Rust demonstrate architectural separation is technically feasible without requiring unified cognitive models
• →Production AI systems need explicit persistence semantics to avoid cognitive errors where factual claims degrade like memories