Every conversation leaves a trace. Every decision echoes.

智能记忆系统：支持用户自定义记忆类别、AI自动分析触发、记忆冲突解决。 当用户需要持久记忆跨会话信息、设定记忆规则、自动记忆重要内容、或解决记忆冲突时使用此技能。 支持L1/L2/L3三层记忆分层、JSON快速索引搜索、用户自定义记忆类别、自动分析触发、冲突时以用户为准。 详细文档见 README.md

Security Scan

The skill is largely coherent for a local auto-memory tool, but it makes misleading claims about being 'install-and-run automatic' and instructs the agent to read unspecified global memory (MEMORY.md) and to add persistent auto-run rules to AGENTS.md/HEARTBEAT.md — behavior that expands its scope and privacy impact beyond what the package itself enforces.

Purpose & Capability

Name/description match the included scripts: the package is an on-disk auto-memory system that extracts and stores conversation content. The scripts, CLI and docs align with that purpose. However the SKILL.md statement '安装后全自动运行' (install → auto-run) is misleading: there is no installer or service that will autorun on install — the README instructs the user to manually modify AGENTS.md/HEARTBEAT.md to enable automatic triggers.

Instruction Scope

Runtime instructions ask the agent/admin to: read 'MEMORY.md' (a vaguely referenced 'core long-term memory' file), add rules to AGENTS.md and HEARTBEAT.md to call memory_cli.sh every session or heartbeat, and to pass the entire conversation contents into scripts. That grants the skill broad access to per-session transcripts and (if AGENTS.md is followed) makes it run every session. The instruction to read 'MEMORY.md' is ambiguous and may require the agent to access a global file outside the skill directory, increasing scope and privacy risk.

Install Mechanism

No install spec; this is instruction-and-script-based. No downloads, no external packages fetched. That minimizes supply-chain risk. All code is local and filesystem-based.

Credentials

The skill requests no environment variables, credentials, or network endpoints. All required configuration and data is stored in files under the skill workspace (config/ and memory/). This is proportionate to a local memory tool. Note: it does ask the agent to pass full conversation text into the scripts, which is sensitive but expected for a memory system.

Persistence & Privilege

The skill does not set always:true, but the SKILL.md instructs administrators to modify agent-wide AGENTS.md and HEARTBEAT.md so the scripts run every session/heartbeat. That effectively gives persistent, automatic execution if the operator follows the instructions. The package itself does not programmatically alter other skills' configs, but its documentation advocates making persistent agent-level changes that increase blast radius and privacy exposure.

What to consider before installing

This skill is a local, file-based auto-memory system and the code matches that description — it parses conversations and writes extracted facts to memory/*.md and an index.json. Before installing: 1) Understand it will ask you (via docs) to modify AGENTS.md / HEARTBEAT.md so it runs automatically each session — it will not auto-enable itself; enabling those rules will make it receive and store entire conversation transcripts. 2) Review the scripts and memory files (they are included) and test in an isolated workspace first; the tool stores potentially sensitive personal data in plaintext under the skill directory. 3) If you want manual control, do not add the AGENTS.md/HEARTBEAT.md entries and keep auto_trigger=false in config; run capture/auto_analyze manually. 4) Be cautious about the ambiguous 'MEMORY.md' reference — confirm which file would be read before granting access. 5) If you need help assessing privacy impact (what will be recorded, retention), ask for a walkthrough of where data is written, retention/cleanup behavior, and how to disable automatic capture.

Auto Memory - 智能记忆系统 v2.0

全自动记忆：不需要用户说任何触发词，安装即自动运行

---

一句话说明

安装后全自动运行：每次会话结束自动记忆，boss 不需要说任何触发词。

---

自动触发机制（核心）

会话级自动记忆（完全自动）

不需要 boss 说任何话，系统在以下时机自动触发：

1. 每次会话结束前 → 自动调用 auto_session 分析整段对话并记忆
2. 新会话开始时 → 自动加载已有记忆到上下文
3. HEARTBEAT 定时 → 每日自动整理记忆

自动捕捉的内容（无需触发词）

内容类型	自动触发条件	类别
boss 决策	任何决定性表述	boss_decision
boss 偏好	任何喜欢/不喜欢表达	boss_preference
偏好变化	发现"之前A现在B"	boss_preference_change
boss 情绪	任何情绪/心情表达	boss_emotion
boss 评价	任何看法/评价	boss_evaluation
任务进展	任务完成/变化	work_context
数字/日期	重要数字出现3次+	work_context
偏好冲突	与旧记忆矛盾	自动更新

---

接入 AGENTS.md（自动触发）

安装此 skill 后，在 AGENTS.md 中加入以下规则：

## 记忆加载（每次会话开始）
每次会话开始：
1. 读 MEMORY.md（核心长期记忆）
2. 读 ~/.openclaw/workspace/skills/auto-memory/memory/index.json（记忆索引）
3. 搜索最近相关记忆：memory_cli.sh recent 10

## 记忆自动触发（每次会话结束）
每次会话结束：
1. 整理本会话 boss 说过的所有内容
2. 调用 auto_session 整段分析：
   cd ~/.openclaw/workspace/skills/auto-memory/scripts && bash memory_cli.sh auto_session "本会话所有对话内容"
3. 检查冲突：如有新内容与旧记忆矛盾，以 boss 最新说的为准
4. 更新 memory/index.json 索引

## 冲突解决
如果 boss 说的和记忆矛盾：
→ 以 boss 最新说的为准
→ 自动更新记忆
→ 旧记忆备份到 conflict/

---

在 HEARTBEAT 中集成

在 HEARTBEAT.md 中加入：

## 记忆维护（每天）
- 运行：memory_cli.sh auto_session 整理今日对话
- 检查：memory_cli.sh stats 查看记忆统计
- 清理：删除 temporary/ 中7天前的文件

## 每周记忆整理（周一）
- 读过去7天 memory/current/*.md
- 更新 MEMORY.md 重要内容
- 运行：memory_cli.sh dedup 去重

---

CLI 命令

cd ~/.openclaw/workspace/skills/auto-memory/scripts

# 会话级自动记忆（推荐，会自动分析并记忆所有重要内容）
bash memory_cli.sh auto_session "本会话所有对话内容"

# 手动记忆
bash memory_cli.sh capture "内容" --type boss_preference --importance 8

# 搜索
bash memory_cli.sh search "关键词"

# 统计
bash memory_cli.sh stats

# 最近记忆
bash memory_cli.sh recent 10

# AI 分析（查看分析结果，不自动记忆）
bash memory_cli.sh auto_analyze "对话内容"

---

记忆类别

类别	层级	说明
boss_info	permanent	boss 基本信息
boss_decision	permanent	重要决策
boss_preference	permanent	个人偏好
boss_preference_change	permanent	偏好变化
boss_emotion	current	情绪状态
boss_negative	current	负面情绪
work_context	current	工作上下文
learning	current	学到的新知识
boss_evaluation	current	对某事的评价
error_recovery	permanent	错误与恢复

---

用户自定义类别

# 添加自定义类别
bash memory_cli.sh config add_type "投资记录" "boss 的投资持仓"

# 调整默认层级
bash memory_cli.sh config set_tier 投资记录 permanent

---

完整文档

见 README.md