Self Improving Agent SkillSAFE

From the repository's own README, as read at the audited commit. Badges and raw HTML are left out.

clawhub install self-improving-agent-skill

用户:帮我修复这个 bug...
Agent: [修复完成]
→ 自动触发:提取经验 → 抽象模式 → 提议更新

用户:自我进化
Agent: 正在分析本次会话经验...
- 提取 3 个新模式
- 提议更新 2 个技能
- 请确认是否应用更改

[检测到错误]
→ 捕获错误上下文
→ 识别根本原因
→ 提议修正方案(需用户确认)
→ 应用修正并添加标记

{workspace}/memory/self-improving/
├── semantic/
│   └── patterns.json          # 抽象模式和规则
├── episodic/
│   └── YYYY/
│       └── YYYY-MM-DD-{task}.json  # 具体经验
├── working/
│   ├── current_session.json   # 当前会话
│   ├── last_error.json        # 最近错误
│   └── session_end.json       # 会话结束标记
└── index.json                 # 全局索引和指标

任务完成 → 提取经验 → 抽象模式 → 提议更改 → ★用户确认★ → 更新技能 → 存入记忆 → 收集反馈

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: self-improving-agent-skill
version: 0.2.0
description: 基于对经验的持续学习,不断优化 Agent 能力。适用于完成重要任务后、出现错误时、会话结束时,或用户输入“自我进化”“总结经验”“从经验中学习”等指令时触发。
---

# Self-Improving Agent

> "An AI agent that learns from every interaction, accumulating patterns and insights to continuously improve its own capabilities." — Based on 2025 lifelong learning research

## Overview

This is a **universal self-improvement system** that learns from ALL task experiences. It implements a complete feedback loop:

- **Multi-Memory Architecture**: Semantic (patterns/rules) + Episodic (experiences) + Working (session context)
- **Self-Correction**: Detects and fixes guidance errors
- **Self-Validation**: Periodically verifies skill accuracy
- **Evolution Markers**: Traceable changes with source attribution
- **Confidence Tracking**: Measures pattern reliability over time
- **User Confirmation Gate**: All skill file modifications require explicit user approval before applying
- **Human-in-the-Loop**: Collects feedback to validate improvements

## Research-Based Design

| Research | Key Insight | Application |
|----------|-------------|-------------|
| [SimpleMem](https://arxiv.org/html/2601.02553v1) | Efficient lifelong memory | Pattern accumulation system |
| [Multi-Memory Survey](https://dl.acm.org/doi/10.1145/3748302) | Semantic + Episodic memory | World knowledge + experiences |
| [Lifelong Learning](https://arxiv.org/html/2501.07278v1) | Continuous task stream learning | Learn from every task |
| [Evo-Memory](https://shothota.medium.com/evo-memory-deepminds-new-benchmark) | Test-time lifelong learning | Real-time adaptation |

## The Self-Improvement Loop

```
┌──────────────────────────────────────────────────────────────┐
│                  UNIVERSAL SELF-IMPROVEMENT                   │
├──────────────────────────────────────────────────────────────┤
│                                                              │
│  Task Event → Extract Experience → Abstract Pattern → Update │
│       │               │                 │              │     │
│       ▼               ▼                 ▼              ▼     │
│  ┌────────────────────────────────────────────────────────┐  │
│  │              MULTI-MEMORY SYSTEM                       │  │
│  ├────────────────────────────────────────────────────────┤  │
│  │ Semantic Memory  │ Episodic Memory  │ Working Memory   │  │
│  │ (Patterns/Rules) │ (Experiences)    │ (Current)        │  │
│  │ memory/self-improving/semantic/ │ memory/self-improving/episodic/ │ memory/self-improving/working/  │  │
│  └────────────────────────────────────────────────────────┘  │
│                                                              │
│  ┌────────────────────────────────────────────────────────┐  │
│  │              FEEDBACK LOOP                             │  │
│  │ User Feedback → Confidence Update → Pattern Adapt      │  │
│  └────────────────────────────────────────────────────────┘  │
│                                                              │
└──────────────────────────────────────────────────────────────┘
```

## When This Activates

### Automatic Triggers

| Event | Action |
|-------|--------|
| Any significant task completes | Extract patterns, propose skill updates (requires user confirmation) |
| An error or failure occurs | Capture error context, trigger self-correction (requires user confirmation before applying fixes) |
| Session ends | Consolidate working memory into long-term memory |

### Manual Triggers

- User says "自我进化", "self-improve", "从经验中学习"
- User says "分析今天的经验", "总结教训", "总结经验"
- User asks to improve a specific skill or workflow

## Memory Storage

### Workspace Discovery

Before accessing any memory files, the agent MUST first determine the workspace root path:

1. **Check environment** — Use the workspace path provided by the IDE/environment context
2. **Verify structure** — Confirm the workspace root by checking for project markers (e.g., `.git/`, `package.json`, `pom.xml`, etc.)
3. **All paths below are relative to the workspace root** — e.g., `{workspace}/memory/self-improving/`

### Relationship with Agent Memory

The Self-Improving Agent's memory lives **inside** the Agent's `memory/` directory as a dedicated subdirectory. This design ensures:

- **No confusion**: Agent's own memory (`MEMORY.md`, `memory/YYYY-MM-DD.md`) and Self-Improving Agent's memory (`memory/self-improving/`) are clearly separated by directory structure
- **Discoverability**: The Agent can browse `memory/` and naturally find self-improving insights
- **Supplement, not replace**: Self-Improving Agent can **append** high-confidence patterns to Agent's memory files (with user confirmation), enriching the Agent's knowledge

```
{workspace}/
├── MEMORY.md                          # Agent core memory (Self-Improving Agent can append)
├── memory/
│   ├── YYYY-MM-DD.md                  # Agent daily memory (Self-Improving Agent can append)
│   └── self-improving/                # Self-Improving Agent dedicated memory space
│       ├── semantic/
│       │   └── patterns.json          # Abstract patterns and rules
│       ├── episodic/
│       │   └── YYYY/
│       │       └── YYYY-MM-DD-{task}.json  # Specific experiences
│       ├── working/
│       │   ├── current_session.json   # Active session data
│       │   ├── last_error.json        # Error context for self-correction
│       │   └── session_end.json       # Session end marker for consolidation
│       └── index.json                 # Memory index and metrics
```

### Memory Interaction Rules

| Action | Target | Condition |
|--------|--------|-----------|
| Read | `MEMORY.md` | Always — to understand Agent's accumulated knowledge |
| Read | `memory/YYYY-MM-DD.md` | Always — to understand today's context |
| Append to | `MEMORY.md` | Only high-confidence patterns (>= 0.9), requires user confirmation |
| Append to | `memory/YYYY-MM-DD.md` | Session summary and key learnings, requires user confirmation |
| Full CRUD | `memory/self

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (1)

<key>k5Ey9KFlkqpj+SDkUw+5ED9lTA3En/qUi0zdrydUCH3kMWTE3Eh65NXnFCaxlY2omY2JHnlEoK7Li7oOEvM7eG5VPdcO/sFlMfoCRdnLYdepJ+uLzYwOWR8W4yQVve/clxVFTVRL4DFleKInGdpAxIbHZT2yi4ADAMENls1N1XSLojRuqXePXDeAT/4Mv4TTx0s

Gates applied: no_behavioural_pass.