受脑科学启发的七层记忆架构，让AI长期记忆更稳定高效
ZenBrain: A Neuroscience-Inspired 7-Layer Memory Architecture for Autonomous AI Systems
- 模仿大脑机制构建七层记忆系统，整合15种神经科学原理
- 在长时记忆任务中表现超越现有模型，仅需极低计算成本
- 适合需要持续学习与跨会话推理的自主AI系统
ZenBrain是一种基于神经科学的七层记忆架构，专为大语言模型智能体设计，统一了十五项机制，包括双因素突触巩固和模拟-选择式睡眠循环等，由单一记忆协调器管理：九个基础算法加六个预测性记忆组件。现有系统中无一集成超过两项此类机制。单独剔除每项机制后，出现称为‘协同掩蔽’的现象：在中等负载下，十四项剔除看似无损（几乎如无用冗余），但将衰减率提高至0.25/天、持续60天后，九项机制变得关键（最大ΔQ达-93.7%；威尔科克斯检验，10组种子），五项从0%直接下降至低于-89%。这表明各机制构成协同生存网络，轻载测试会严重低估其贡献——此警示或适用于更广泛架构。所有消融实验均能在笔记本上一分钟内完成（npm install，无需API密钥）。在LongMemEval-500上，ZenBrain在九次一对一答案质量对比中全面胜出（3个竞争模型 × 3个LLM评委），经邦费尼修正显著（α=0.05/18，p_min=6.2e-31，d∈[0.18,0.52]），以1/106的每查询令牌成本达到全上下文最优者91.3%的二元判断准确率（47.7% vs. 52.2%）。多层路由相比单层基线提升20.7% F1（LoCoMo），无衰减消融显示有原则遗忘代价仅为Delta-P@5=0.002。坦诚说明：LoCoMo的基于子串聚合的F1指标偏向词法检索（BM25），不反驳此设定。ZenBrain优势在评委评分的答案质量和跨会话推理中尤为突出。
原文摘要 · Abstract (English)
ZenBrain is a seven-layer, neuroscience-derived memory architecture for LLM agents that unifies fifteen mechanisms - from Two-Factor synaptic consolidation to a Simulation-Selection sleep loop - under a single MemoryCoordinator: nine foundational algorithms plus six Predictive Memory Architecture components. No system among those we survey integrates more than two of them. Ablating each mechanism separately exposes an effect we call cooperative masking. Under moderate load, fourteen of the fifteen ablations look costless - the architecture reads as mostly dead weight. Raising decay to 0.25/day over 60 days, with no change to the mechanisms, makes nine of the fifteen individually critical (Delta-Q up to -93.7%; Wilcoxon, 10 seeds), five of them moving from exactly 0% to below -89%. The mechanisms form a cooperative survival network, and mild-load ablation systematically underestimates architectural contributions - a caution we conjecture applies beyond ZenBrain. Every ablation table reproduces in under one minute on a laptop (npm install, no API keys). On LongMemEval-500, ZenBrain wins all nine head-to-head answer-quality comparisons (3 competitors x 3 LLM judges) against Letta, Mem0 and A-Mem under Bonferroni-corrected significance (alpha=0.05/18, p_min=6.2e-31, d in [0.18, 0.52]), and reaches 91.3% of a full-context oracle's binary-judge accuracy at 1/106th of the per-query token cost (47.7% vs. 52.2%). Multi-layer routing beats a flat single-layer baseline by +20.7% F1 on LoCoMo; a NoDecay ablation shows principled forgetting costs only Delta-P@5=0.002. Honest scope: LoCoMo's substring-based aggregate F1 favors lexical retrieval (BM25) by metric design, and we do not contest this. ZenBrain's advantages are most pronounced on judge-graded answer quality and cross-session reasoning.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。