Dual-Axis Bitemporal Neuro-Cognitive Architecture and Tripartite Epistemic Governance for Long-Horizon Autonomous Agents
Technical Report & Preprint (VTI-RES-2026-02) — Bilingual Edition (English & Portuguese) + Executive Briefs + LaTeX Sources. Large Language Model (LLM) agentic systems suffer systematic degradation over extended horizons due to three cognitive failure modes: long-context attention decay (Lost in the Middle), statistical override of newly enacted rules by high-frequency historical priors (Override Failure), and memory pollution from transient runtime errors (Contaminated Mindware). This paper formalizes the Dual-Axis Bitemporal Neuro-Cognitive Architecture implemented in MemoryOS (VTI OS System), coupling five functional memory tiers plus a quarantined aversive store (Axis 1) with a tripartite executive governance framework inspired by Keith E. Stanovich (Axis 2). Evaluated on a controlled N = 50 bitemporal conflict benchmark, the post-RRF epistemic reranker increased NDCG@5 from 0.5764 to 1.0000 (+73.49%) with 0/50 revocation override failures at P95 = 0.34 ms.
Authors
- Caio Gonçalves
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24290283
- Primary Topic
- Psychiatry, Mental Health, Neuroscience
- Type
- preprint