Beyond Flat Vector Retrieval: Mitigating the Dual Collapse and In-Context Prompt Poisoning of RAG via Structured Valid-Time Cognitive Architectures

Retrieval-Augmented Generation (RAG) connects large language models (LLMs) to external non-parametric memory through dense bi-encoder indices. However, it relies on a flawed assumption: semantic proximity in vector space reliably indicates factual validity over time. We identify three systemic failure modes arising from this assumption. Temporal Inversion occurs because transformer bi-encoders compress token representations into fixed-length vectors without encoding temporal ordering. Consequently, historical and updated assertions yield near-identical cosine similarities (∆ sim_cos < 0.012), causing obsolete retrievals in 41.7% of state transitions. Dual Collapse emerges at enterprise scale (10,000 passages; 500 entities), where entity identifiers receive less than 5% of self-attention weight, driving cross-entity cosine similarities above 0.98 and reducing top-1 accuracy to 9.0%. InContext Prompt Poisoning occurs when stale retrievals enter instruction-tuned decoders, which reproduce obsolete facts despite conflicting parametric priors, yielding hallucination rates of 33.3%–40.0% on public entities. To address these failures without model retraining, we introduce HERM4AI (Human-like Evolving & Reasoning Memory for Artificial Intelligence), a model-free cognitive architecture combining deterministic valid-time intervals [tfrom, tto), a bounded 7-slot working memory with salience eviction, an associative knowledge graph, spreading activation, sleep-phase consolidation,and ACT-R workflows. Across seven benchmark suites, four cloud GPU experiment tracks, and a curated 12-case public-entity benchmark informed by TempLAMA and FreshQA, HERM4AI achieves 100.0% accuracy, 0.065 ms query latency, and under 6 MB RAM on a standard CPU. The public-entity evaluation is not a direct evaluation on TempLAMA or FreshQA.