Super Self Improving

超级自我优化智能体 - 多模态记忆、反馈循环、元学习、置信度校准 / Super Self-Improving Agent - Multi-modal Memory, Feedback Loops, Meta-Learning, Confidence Calibration

Security Scan

The skill's description and runtime instructions broadly match a self‑improving memory agent, but the SKILL.md refers to persistent files, CLI commands, and token/usage monitoring without declaring config paths, binaries, or credentials — an internal inconsistency that warrants caution.

Purpose & Capability

The declared purpose (self‑improving agent with memory, feedback loops, meta‑learning and confidence calibration) is consistent with the prose in SKILL.md. However, the skill references persistent storage under ~/.super-self-improving and CLI commands (e.g., super-self-improving stats) while the registry metadata declares no required config paths or binaries. That mismatch is unexpected: a memory/monitoring skill legitimately needs a place to store data and potentially tooling, but those requirements are not declared.

Instruction Scope

SKILL.md instructs storing and reading memory files (hot.md, preferences.md, etc.) under the user's home directory and defines workflows that imply reading session/context and tracking token consumption and performance metrics. The instructions do not explicitly ask for unrelated system files or credentials, but they do assume persistent local file I/O and implicit access to usage/token metrics. The guidance 'do not store sensitive information' is advisory only — there is no enforcement or mechanism described to prevent sensitive data from being written or transmitted.

Install Mechanism

This is an instruction‑only skill with no install spec and no code files packaged. That minimizes immediate supply‑chain risk because nothing is downloaded or executed by an installer. However, the README references a python script (super_self_improving.py) and CLI commands which are not included; if a user later installs code from an external source to satisfy these commands, that introduces risk not captured here.

Credentials

The skill declares no required environment variables or credentials, yet it describes token monitoring, cost estimation, and agent scheduling/auto‑scaling features that in practice often require access to provider APIs, billing info, or orchestration credentials. The absence of declared env vars/config paths but the presence of features that normally need them is an inconsistency. Also, persistent memory in the user's home could unintentionally capture secrets even if the skill 'promises' not to.

Persistence & Privilege

The skill intends to persist data under ~/.super-self-improving (detailed directory layout and files). The registry metadata lists no required config paths, so this persistence is an implicit privilege not declared up front. The skill does not request always:true or attempt to change other skills/settings, which reduces the privilege concerns, but users should be aware it expects to create and maintain local persisted state.

What to consider before installing

This skill conceptually fits a self‑improving memory agent, but there are mismatches you should resolve before installing or running it: 1) The SKILL.md expects persistent files in ~/.super-self-improving and CLI commands (e.g., super-self-improving) yet the package declares no config paths or binaries — ask the publisher which files will be created and inspect them. 2) Token monitoring and scheduling features imply access to usage/billing data or orchestration credentials; confirm what data sources it will read and whether any credentials are required. 3) The skill promises not to store sensitive data, but that is not enforceable — avoid letting it store any secrets, or run it in a sandboxed account. 4) Because no code is included, do not install third‑party code to satisfy referenced commands unless you review its source. If you decide to proceed, run in an isolated environment, inspect any created files under ~/.super-self-improving, and verify there are no unexpected network endpoints or credential reads.

超级自我优化智能体 / Super Self-Improving Agent

基于原有self-improving的增强版，增加多模态记忆、元学习、置信度校准等功能。

Enhanced version with multi-modal memory, meta-learning, confidence calibration and more.

🆕 相比原版新增功能

1. 多模态记忆 / Multi-modal Memory

• 📝 文本偏好 (Text preferences)
• 💻 代码模式 (Code patterns)
• 🎨 风格偏好 (Style preferences)
• 🔧 工具使用习惯 (Tool usage habits)
• 📊 性能指标 (Performance metrics)

2. 反馈循环 / Feedback Loops

• ✋ 显式反馈 (Explicit feedback) - 用户直接纠正
• 👁️ 隐式反馈 (Implicit feedback) - 从行为推断
• 🤖 合成反馈 (Synthetic feedback) - 自我评估

3. 元学习 / Meta-Learning

• 学习如何学习 (Learn how to learn)
• 识别最佳策略 (Identify best strategies)
• 动态调整方法 (Dynamic method adjustment)

4. 置信度校准 / Confidence Calibration

• 预测准确度追踪 (Track prediction accuracy)
• 校准评分 (Calibration score)
• Uncertainty quantification

5. 错误分析 / Error Analysis

• 错误分类 (Error categorization)
• 根因分析 (Root cause analysis)
• 预防模式 (Prevention patterns)

---

📁 目录结构 / Directory Structure

~/.super-self-improving/
├── memory/
│   ├── hot.md           # 始终加载 (<100行)
│   ├── preferences.md    # 用户偏好
│   ├── patterns.md      # 行为模式
│   └── metrics.md       # 性能指标
├── projects/            # 项目级记忆
├── domains/             # 领域级记忆
├── archive/             # 归档
├── feedback/
│   ├── explicit.md      # 显式反馈
│   ├── implicit.md      # 隐式反馈
│   └── synthetic.md     # 自我评估
├── errors/              # 错误分析
│   ├── categories.md    # 错误分类
│   ├── root_causes.md  # 根因分析
│   └── prevention.md   # 预防模式
└── meta/
    ├── strategy.md      # 学习策略
    ├── calibration.md  # 置信度校准
    └── stats.json      # 统计信息

---

🔄 工作流程 / Workflow

用户输入 → 意图识别 → 上下文匹配 → 执行 → 反馈收集
                  ↓                        ↓
            记忆检索 ←──────────────── 自我评估
                  ↓
            模式学习 → 策略更新 → 置信度调整

---

📊 性能指标 / Performance Metrics

追踪以下指标：

指标	说明
task_completion_rate	任务完成率
user_satisfaction	用户满意度
error_rate	错误率
response_time	响应时间
pattern_accuracy	模式识别准确率
calibration_score	置信度校准分数

---

🎯 核心机制 / Core Mechanisms

1. 反馈收集 / Feedback Collection

# 收集反馈
def collect_feedback(context):
    explicit = detect_explicit_correction(context)  # 用户直接纠正
    implicit = detect_implicit_feedback(context)    # 行为推断
    synthetic = self_assessment(context)            # 自我评估
    
    return combine_feedback(explicit, implicit, synthetic)

2. 模式识别 / Pattern Recognition

# 识别重复模式
def recognize_patterns(memory, threshold=3):
    # 统计出现频率
    # 识别关联规则
    # 生成模式建议
    return patterns

3. 策略更新 / Strategy Update

# 基于反馈更新策略
def update_strategy(patterns, metrics):
    # 分析什么有效
    # 调整方法
    # 更新置信度
    return updated_strategy

4. 置信度校准 / Confidence Calibration

# 校准置信度
def calibrate(prediction, actual_outcome):
    # 记录预测 vs 实际
    # 计算校准分数
    # 调整未来预测
    return calibrated_confidence

---

📋 触发条件 / Triggers

显式纠正

• "不对"
• "应该是..."
• "我告诉过你..."
• "我不喜欢..."

隐式信号

• 用户重复问题
• 长时间沉默
• 跳过回答
• 转换话题

自我评估触发

• 完成复杂任务后
• 收到模糊反馈
• 遇到新场景

---

🏆 升级规则 / Promotion Rules

层级	使用频率	确认次数
HOT	每次	3次确认
WARM	相关上下文	5次使用
COLD	显式查询	归档

---

🔒 安全边界 / Security Boundaries

1. 不存储敏感信息 (No sensitive data)
2. 不访问未授权文件 (No unauthorized file access)
3. 不修改系统配置 (No system config changes)
4. 定期清理过期数据 (Regular cleanup)

---

📈 使用示例 / Usage Examples

# 查看记忆统计
super-self-improving stats

# 添加显式反馈
super-self-improving feedback --explicit "不要用markdown表格"

# 查看性能指标
super-self-improving metrics

# 导出记忆
super-self-improving export

# 置信度校准
super-self-improving calibrate

---

⚡ 与原版对比 / Comparison with Original

特性	原版	增强版
记忆类型	文本	多模态
反馈来源	显式	显式+隐式+合成
学习方式	被动	主动+被动
错误处理	记录	分析+预防
置信度	无	完整校准
性能追踪	无	完整指标

---

📝 记录格式 / Logging Format

显式反馈

## 2026-03-05
- 用户纠正: "不要用表格，用列表"
- 原因: 用户偏好
- 状态: 已确认

隐式反馈

## 2026-03-05
- 行为: 用户重复提问3次
- 推断: 上次回答不够清晰
- 动作: 改进回答方式

自我评估

## 2026-03-05
- 任务: 复杂代码调试
- 评估: 第一次尝试失败
- 改进: 添加更多调试信息

---

🎯 最佳实践 / Best Practices

1. 频繁小改进 > 偶尔大改进
2. 量化跟踪 > 主观感觉
3. 预防优先 > 事后纠正
4. 透明可解释 > 黑箱学习
5. 用户控制 > 自主推断

---

💰 Token监控 / Token Monitoring

功能 / Features

• 📊 实时token消耗追踪 / Real-time token consumption tracking
• ⚠️ 异常消耗预警 / Abnormal consumption alerts
• 📈 使用趋势分析 / Usage trend analysis
• 💵 成本估算 / Cost estimation

指标 / Metrics

指标	说明
session_tokens	当前会话消耗
total_tokens	总会话消耗
cache_efficiency	缓存效率
avg_tokens_per_turn	每轮平均消耗
cost_estimate	成本估算

告警规则 / Alert Rules

- 超过平均2倍 → 警告
- 超过平均3倍 → 严重告警
- 缓存效率<50% → 优化建议
- 接近限制(80%) → 提醒

---

🤖 Agent调度优化 / Agent Scheduling Optimization

功能 / Features

• 🎯 智能任务分配 / Intelligent task allocation
• ⚡ 负载均衡 / Load balancing
• 🔄 自动扩缩容 / Auto scaling
• 📊 性能最优化 / Performance optimization

调度策略 / Scheduling Strategies

策略	适用场景
round_robin	均衡负载
shortest_queue	最少等待
skill_match	技能匹配
cost_efficiency	成本优先
performance_based	性能最优

优化规则 / Optimization Rules

1. 根据任务类型选择最佳agent
2. 监控agent负载并动态调整
3. 缓存常用上下文减少重复
4. 预测任务复杂度分配资源
5. 定期评估并优化策略

性能指标 / Performance Metrics

指标	说明
task_completion_time	任务完成时间
success_rate	成功率
queue_wait_time	等待时间
resource_utilization	资源利用率
user_satisfaction	用户满意度

自动调优 / Auto-tuning

• 收集历史性能数据
• 分析瓶颈和优化点
• 动态调整调度参数
• 持续监控效果