Abstract
Memory management remains a fundamental challenge for autonomous AI systems. Existing approaches employ operating system metaphors, flat storage with LLM-driven management, or note-taking paradigms, but none incorporate principles from cognitive neuroscience—despite memory science offering over a century of empirically validated models.
We present ZenBrain, a multi-layer memory architecture for AI agents that integrates twelve established neuroscience models into a unified system. ZenBrain implements seven distinct memory layers—working, short-term, episodic, semantic, procedural, core, and cross-context—orchestrated by twelve algorithms including Hebbian learning dynamics for knowledge graph co-activation, Ebbinghaus forgetting curves with FSRS spaced repetition scheduling, sleep-time memory consolidation, Bayesian confidence propagation with 95% confidence intervals, and emotional valence tagging.
We evaluate ZenBrain across seven experiments on three retrieval benchmarks—LoCoMo (1,986 QA pairs), MemoryAgentBench (5 memory capability dimensions), and MemoryArena (cross-session dependencies)—plus controlled retention, consolidation, and algorithm-level experiments. Multi-layer routing outperforms flat storage by 21.6% in F1 on LoCoMo (p = 0.005), with the largest gains on temporal queries (+176%), and by 19.5% on MemoryArena's cross-session dependencies (p = 0.015, +53.5% on dependency chains). Our 3-phase sleep consolidation (SWS/REM/SHY) achieves a 37% stability improvement (p = 0.005, Cohen's d = 9.17) with 47.4% storage reduction through synaptic homeostasis. Hebbian knowledge graph dynamics produce retrieval precision@5 of 0.955 from raw co-activation patterns (vs. 0.200 uniform, p = 0.005), and Bayesian confidence propagation separates true from false facts with AUC improvement from 0.533 to 0.797 (p = 0.009). ZenBrain is open-source, production-deployed, and distributed as composable npm packages with 9,500+ automated tests.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Bering, Alexander, "ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/9720