ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems
Date
May 2026
Publication form
Open-access preprint (arXiv) and open-access publication with DOI (Zenodo)
Summary (short form)
This work presents a seven-layer memory architecture for autonomous AI systems whose design principles are derived from cognitive neuroscience. The architecture integrates Working Memory, Short-Term Memory, Episodic Memory, Semantic Memory, Procedural Memory, Core Memory, and Cross-Context Memory into a unified system. Consolidation follows sleep-replay mechanisms with Hebbian-driven strengthening and Ebbinghaus-driven decay. The architecture is available as a modular open-source library under Apache 2.0 and is deployed in production applications. In evaluation on LongMemEval-500 with three independent LLM judges, it wins all twelve head-to-head comparisons against Letta, Mem0, and A-Mem, reaching 91.3% of long-context-oracle accuracy at 1/106 of the token budget (Bonferroni-corrected, p ≤ 6.2·10⁻³¹).
The full scientific abstract and the formal presentation of methods and results are available in the preprint and DOI versions.
Access
What this means in practice: read the analysis on our blog →