ClawKit Logo
ClawKitReliability Toolkit
Back to Registry
Official Verified

Ntm Memory System

Skill by 1580021414-afk

skill-install — Terminal

Install via CLI (Recommended)

clawhub install openclaw/skills/skills/1580021414-afk/ntm-memory-system
Or

name: ntm-memory-system description: 基于 Neural Turing Machines 的外部记忆系统,让 AI 具备可读写的外部记忆能力 metadata: openclaw: emoji: "🧠" category: "AI-Core" version: "1.0.0" author: "小钳" paper: "Neural Turing Machines (Graves et al., 2014)" price: 0 contact: "微信 17612824848" tags: - 外部记忆 - 神经图灵机 - 可微分计算 - 算法学习

Neural Turing Machines Memory System

基于 DeepMind 的 Neural Turing Machines 论文,为 AI 提供可读写的外部记忆能力。


一、核心概念

1.1 什么是 Neural Turing Machine?

NTM 将神经网络与外部记忆资源耦合:

┌─────────────────────────────────────────────────────┐
│                  Neural Turing Machine              │
├─────────────────────────────────────────────────────┤
│                                                     │
│   ┌─────────────┐      读写头      ┌─────────────┐  │
│   │             │ ◄─────────────► │             │  │
│   │  Controller │                 │   Memory    │  │
│   │  (神经网络)  │                 │   Matrix    │  │
│   │             │                 │  (N × M)    │  │
│   └─────────────┘                 └─────────────┘  │
│         ▲                               │          │
│         │                               │          │
│         └───────────────────────────────┘          │
│                   注意力机制                        │
└─────────────────────────────────────────────────────┘

1.2 关键特性

特性说明
可微分端到端训练,梯度下降优化
外部记忆突破神经网络隐藏层大小限制
注意力寻址内容寻址 + 位置寻址
算法学习能学习复制、排序、关联回忆等算法

二、记忆架构

2.1 记忆矩阵

interface MemoryMatrix {
  // 记忆矩阵: N 个位置,每个位置 M 维
  memory: number[][];  // N × M
  
  // 权重向量: 读/写头的注意力分布
  weights: number[];   // N 维,和为1
  
  // 读写头状态
  heads: {
    read: ReadHead;
    write: WriteHead;
  };
}

2.2 寻址机制

interface AddressingMechanism {
  // 1. 内容寻址 - 根据相似度
  contentAddressing(key: number[], memory: number[][]): number[] {
    // 计算余弦相似度
    const similarities = memory.map(row => 
      cosineSimilarity(key, row)
    );
    // softmax 得到权重
    return softmax(similarities.map(s => s * beta)); // beta = 关注强度
  }
  
  // 2. 位置寻址 - 根据位置偏移
  locationAddressing(weights: number[], shift: number): number[] {
    // 循环移位
    return circularShift(weights, shift);
  }
  
  // 3. 混合寻址
  gatedAddressing(
    contentWeights: number[], 
    locationWeights: number[], 
    gate: number
  ): number[] {
    return contentWeights.map((w, i) => 
      gate * w + (1 - gate) * locationWeights[i]
    );
  }
}

2.3 读写操作

interface ReadHead {
  // 从记忆中读取
  read(memory: number[][], weights: number[]): number[] {
    // 加权求和
    return memory[0].map((_, j) => 
      weights.reduce((sum, w, i) => sum + w * memory[i][j], 0)
    );
  }
}

Metadata

Stars4473
Views1
Updated2026-05-01
View Author Profile
AI Skill Finder

Not sure this is the right skill?

Describe what you want to build — we'll match you to the best skill from 16,000+ options.

Find the right skill
Add to Configuration

Paste this into your clawhub.json to enable this plugin.

{
  "plugins": {
    "official-1580021414-afk-ntm-memory-system": {
      "enabled": true,
      "auto_update": true
    }
  }
}
Safety NoteClawKit audits metadata but not runtime behavior. Use with caution.