Description

LLM 增强记忆系统 — 语义理解、智能摘要、敏感信息保护。三层架构自动管理 agent 记忆。

README (SKILL.md)

Auto Memory - LLM 增强记忆系统 v3.0.0

Name: Auto Memory
Author: dtldhjh

LLM 驱动的智能记忆系统，让 agent 拥有完整的持久记忆能力。

v3.0.0 核心特性

特性	说明
🤖 LLM 分析	语义理解对话内容，智能分类
🔒 敏感信息保护	自动检测并脱敏 API Key、Token、密码等
📊 结构化提取	偏好/决策/待办/问题/学习五维分类
📝 智能摘要	自动生成对话摘要
🔄 智能去重	跨次运行去重，不再重复累积

三层记忆架构

用户对话
    ↓ (extract-memory.sh v3 + LLM)
┌─────────────────────────┐
│  memory/YYYY-MM-DD.md   │  日常日志
│  - LLM 智能分析         │
│  - 五维分类提取         │
│  - 敏感信息脱敏         │
└───────────┬─────────────┘
            ↓ (update-long-memory.sh)
┌─────────────────────────┐
│      MEMORY.md          │  长期记忆
│  - 用户偏好             │
│  - 重要决策             │
│  - 项目进展             │
└───────────┬─────────────┘
            ↓ (openclaw memory index)
┌─────────────────────────┐
│     向量数据库          │  可检索
│  - openclaw memory search │
└─────────────────────────┘

安装

# 创建目录
mkdir -p ~/.openclaw/scripts
mkdir -p ~/.openclaw/workspace/memory/archive
mkdir -p ~/.openclaw/workspace/.learnings/shared

# 下载脚本
# extract-memory.sh (v3.0 LLM 增强版)
# update-long-memory.sh

# 设置权限
chmod +x ~/.openclaw/scripts/extract-memory.sh
chmod +x ~/.openclaw/scripts/update-long-memory.sh

配置 HEARTBEAT.md

## 1. 自动记忆更新
\`\`\`bash
~/.openclaw/scripts/extract-memory.sh main
\`\`\`

## 2. 长期记忆同步
\`\`\`bash
~/.openclaw/scripts/update-long-memory.sh main
\`\`\`

LLM 分析输出

## LLM 分析 (10:34)

**摘要**: 本次对话完成了记忆系统的三层架构改造

### 🎯 偏好
- OpenCode 模型偏好设置
- 服务开发配置：端口 10001-10100

### 📋 决策
- 采用三层互补架构管理记忆

### 📌 待办
- ✅ 优化脚本
- ✅ 创建更新脚本

### ❌ 问题
- MEMORY.md 未及时更新

### 💡 学习
- 去重机制需要跨次检查

敏感信息保护

类型	模式	替换为
API Key	`api_key=xxx`	`[REDACTED_KEY]`
Token	`token=xxx`	`[REDACTED_TOKEN]`
密码	`password=xxx`	`[REDACTED_PASSWORD]`
手机号	`13812345678`	`[REDACTED_PHONE]`
邮箱	`[email protected]`	`[REDACTED_EMAIL]`
内网 IP	`192.168.x.x`	`[REDACTED_IP]`

脱敏流程：

检测敏感模式
发送给 LLM 前脱敏
LLM 返回结果再次检查

示例输出

🧹 检查过期文件...
📚 加载学习经验...
📄 分析 session: e0262284-xxx.jsonl
   🤖 使用 LLM 分析对话...
✅ LLM 分析完成
   📝 摘要: 本次对话完成了记忆系统的三层架构改造...
🔄 更新向量索引...
✅ 索引已更新

更新日志

v3.0.0 (2026-03-22)

🤖 LLM 语义分析（使用当前模型）
🔒 敏感信息自动脱敏
📊 五维分类提取
📝 智能摘要生成
🔄 智能去重

v2.0.0 (2026-03-22)

三层记忆架构
自动同步到 MEMORY.md

v1.0.0 (2026-03-12)

初始版本

Usage Guidance

This skill contains shell+Python scripts that will be placed in your home (~/.openclaw) and operate on your agent session files and local OpenClaw config. It promises automatic redaction but the code does not perform redaction before calling an LLM endpoint: smart-summarize.sh reads ~/.openclaw/openclaw.json and, if baseUrl and apiKey exist, sends memory content to baseUrl with that apiKey. That could leak sensitive content or your API key if baseUrl is untrusted. Before installing: (1) review and/or run the scripts manually instead of blindly placing them in ~/.openclaw; (2) ensure your ~/.openclaw/openclaw.json provider.baseUrl is a trusted host or remove the apiKey from that config; (3) if you need redaction, add/verify regex-based replacement in the code paths that call external services; (4) consider running the skill in a sandbox or with a non-main agent ID to limit access to sensitive sessions; (5) ask the author for the missing update-long-memory.sh and for explicit, auditable redaction logic. If you cannot verify these points, treat the skill as risky and avoid granting it write/execution access in your primary environment.

Capability Analysis

Type: OpenClaw Skill Name: dtldhjh-auto-memory Version: 3.0.0 The 'auto-memory' skill is a legitimate utility designed to provide long-term memory capabilities for OpenClaw agents by processing session logs into structured Markdown files and vector indices. The scripts (extract-memory.sh, smart-summarize.sh, etc.) perform file operations within the expected ~/.openclaw directory and include explicit logic to detect and redact sensitive information like API keys and passwords before storage or LLM processing. While smart-summarize.sh accesses the user's openclaw.json to retrieve API credentials, it only uses them to communicate with the user's own configured LLM provider for the purpose of generating summaries, with no evidence of data exfiltration or malicious intent.

Capability Assessment

ℹ Purpose & Capability

The scripts' actions (extracting sessions, building daily memory, indexing with openclaw) match the stated purpose of an LLM-backed memory system. However SKILL.md refers to an update-long-memory.sh script that is not present in the package, and the skill writes to shared locations (~/.openclaw/workspace/.learnings/shared) which enlarges its scope to cross-agent sharing — this is plausible for the feature but worth noting.

⚠ Instruction Scope

SKILL.md promises '敏感信息脱敏' (detect & redact before sending to LLM) but the actual code does not implement redaction before LLM calls. extract-memory.sh/Python appends session content directly to memory files; smart-summarize.sh may send recent memory content directly to an external base_url with the stored apiKey. The scripts also read local config (~/.openclaw/openclaw.json), session files (~/.openclaw/agents/.../sessions), and write shared artifacts — behaviors that go beyond the claimed 'safe' processing when redaction is not enforced.

ℹ Install Mechanism

No automated install spec (instruction-only). SKILL.md instructs the user to download scripts into ~/.openclaw/scripts and set executable permissions; that will write files into the user's home but is not automatically performed by the registry. No remote download URLs are embedded in an install step, which lowers install risk, but the act of placing executable scripts under the user's account is required and should be done with caution.

⚠ Credentials

The package declares no required env vars, but smart-summarize.sh reads ~/.openclaw/openclaw.json to obtain provider.baseUrl and provider.apiKey and will perform HTTP requests using that apiKey. This is not declared and contradicts the SKILL.md claim of redaction: unredacted memory content may be transmitted to whichever baseUrl is configured (potentially exfiltrating both content and your API key to a malicious endpoint). The scripts also write to shared cross-agent paths, increasing exposure.

ℹ Persistence & Privilege

always is false and the skill is user-invocable (normal). Installing requires adding executable scripts under ~/.openclaw/scripts and creating workspace directories; that gives the skill ongoing presence in your home directory and can be scheduled externally (HEARTBEAT.md). This persistence combined with the environment access described above increases risk but does not itself indicate malicious intent.

Version History

v3.0.0

Auto Memory 3.0.0 — LLM Enhanced Memory System - Added LLM-driven semantic analysis for intelligent dialog classification. - Introduced automatic sensitive information protection (API keys, tokens, passwords, etc.). - Five-dimension structured extraction: preferences, decisions, todos, issues, and learnings. - Enhanced smart summarization and cross-session deduplication. - Implements a three-layer memory architecture for full, persistent agent memory management.

v1.3.0

Auto Memory v1.3.0 introduces smarter memory management and regular refinement: - 新增智能摘要功能，自动提取关键词摘要 - 实现增量索引，仅对发生变更的文件重建索引 - 增加定期提炼机制，每周日汇总并提炼长期记忆

Metadata

Slug dtldhjh-auto-memory

Version 3.0.0

License MIT-0

All-time Installs 0

Active Installs 0

Total Versions 2

Frequently Asked Questions

What is Auto Memory?

LLM 增强记忆系统 — 语义理解、智能摘要、敏感信息保护。三层架构自动管理 agent 记忆。 It is an AI Agent Skill for Claude Code / OpenClaw, with 277 downloads so far.

How do I install Auto Memory?

Run "/install dtldhjh-auto-memory" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is Auto Memory free?

Yes, Auto Memory is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does Auto Memory support?

Auto Memory is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created Auto Memory?

It is built and maintained by dtldhjh (@dtldhjh); the current version is v3.0.0.

More Skills

Auto Memory