Memory Bridge

Enables AI to retain and access previous session context across restarts by automatically saving and loading user identity, tasks, and preferences from memor...

Security Scan

The skill's behavior (creating and auto-reading persistent files under ~/.openclaw/workspace to store conversation memory) matches its description, but there are metadata/instruction inconsistencies and un-declared filesystem access that warrant caution.

Purpose & Capability

The skill's stated goal is persistent memory across restarts and the instructions show exactly that: create a small index (MEMORY.md) and other memory files under ~/.openclaw/workspace and read them on new sessions. No unrelated credentials or binaries are requested.

Instruction Scope

Instructions tell the agent to create/inspect ~/.openclaw/workspace, copy templates, and read/write memory files (MEMORY.md, memory/, .learnings/). This is consistent with the stated purpose but implies the agent will persist conversation content to disk (including a WAL and working-buffer that may store conversation excerpts when context usage is high). Also there is a minor internal inconsistency in the docs: one place shows 'clawhub install memory-system' and another 'clawhub install memory-bridge'.

Install Mechanism

This is an instruction-only skill (no install spec or code bundled), which is lower risk. However the documentation references installing via 'clawhub' with inconsistent package names. Because there's no packaged install spec here, actual installation behavior (what 'clawhub' will install) is unknown and should be verified before running.

Credentials

The skill metadata declares no required config paths or env vars, yet the runtime instructions explicitly create and read files under ~/.openclaw/workspace. That filesystem access is not declared in the registry metadata; persistent reading/writing of user files (including potentially sensitive conversation content) is a material permission that should be declared and considered before install.

Persistence & Privilege

always:false (not force-included) and autonomous invocation is allowed (default). The skill creates persistent files the agent will read across sessions (MEMORY.md auto-loaded on wake). This gives ongoing access to persisted data but is expected for a memory skill; still, the combination of persistence + autonomous invocation increases blast radius if the implementation is malicious or buggy.

What to consider before installing

What to consider before installing: - The skill will create and persist files under ~/.openclaw/workspace and will automatically read MEMORY.md on new sessions. That means conversation snippets and preferences may be written to disk (WAL / working-buffer), so review those files and templates for anything sensitive before enabling. - The registry metadata did not declare required config paths, but the instructions expect full read/write access to ~/.openclaw/workspace — treat that as an undeclared permission. If you store secrets or sensitive data in chats, they could be persisted. - The SKILL docs show inconsistent install names ('memory-system' vs 'memory-bridge'); verify exactly what the 'clawhub install' command will fetch and inspect that package/source before running it. - Because this is an instruction-only bundle with no packaged code shown, inspect the actual package that clawhub would install (or run the install in an isolated environment) to confirm there is no additional behavior (network callbacks, telemetry, or code that reads other paths). - If you proceed: (1) test in a sandbox/VM or throwaway account; (2) inspect the created files and their permissions; (3) ensure the memory folder is not auto-synced to cloud backups if you have sensitive info; (4) know how to delete/clear memory (delete MEMORY.md and memory/ .learnings/) if you want to revoke persistence; (5) consider requiring encryption for persisted files if storing sensitive content. Given the metadata/instruction inconsistencies and the fact that filesystem persistence was not declared explicitly, treat this as suspicious until you verify the actual package source and its install-time behavior.

memory-bridge Skill

新会话重启后，AI 还知道上次在聊什么。

装一次，记忆永远在。不会因为重启窗口就全忘了。

触发关键词："初始化记忆"、"开启记忆"、"记忆系统"、"帮我建记忆"、"让AI记住我"

---

这是什么

这是一个让 AI 拥有持久记忆的技能。

核心价值： 即使开新会话、清上下文，AI 也能通过记忆文件知道：

• 你是谁
• 上次在聊什么
• 任务做到哪一步了

我们不碰： 你已有的人设（SOUL.md）、行为规范（AGENTS.md）、工具配置（TOOLS.md）

---

如何使用

安装

clawhub install memory-system

启动（自动的！）

装完后不用手动启动，AI 下次唤醒时会自动检测：

检测到 MEMORY.md 不存在
  → 主动引导用户初始化（问"我来帮你建好吗？"）

检测到 MEMORY.md 已存在
  → 自动读取索引（10行以内）
  → 知道：你是谁、在做什么
  → 主动告诉你"我记得你..."

你只需要：开新窗口，AI 就会自动帮你搞定。

手动触发（如果需要）

如果 AI 没有自动启动，你可以手动触发：

对我说："初始化记忆系统" 或 "帮我开启记忆"

---

首次激活（引导式）

当你被唤醒，发现 workspace 中没有 MEMORY.md 时：

第一步：说明价值

🤍 你好！我发现你还没有开启记忆系统。

现在每次新会话，AI 就把之前聊过的东西全忘了。 开了记忆系统之后就不一样了—— 即使你开新窗口，我也能知道上次在聊什么。

我来帮你建好吗？

第二步：逐步构建

Step 1 - 建索引文件
  检查 ~/.openclaw/workspace/MEMORY.md
  → 不存在：从 templates/MEMORY.md 复制
  → 已存在：跳过

Step 2 - 建记忆目录
  检查 ~/.openclaw/workspace/memory/
  → 不存在：创建目录，复制模板文件
  → 已存在：跳过

Step 3 - 建学习记录目录
  检查 ~/.openclaw/workspace/.learnings/
  → 不存在：创建目录，复制模板文件
  → 已存在：跳过

Step 4 - 建用户档案（如果还没有）
  检查 ~/.openclaw/workspace/USER.md
  → 不存在：从 templates/USER.md 复制
  → 已存在：跳过

第三步：完成

✅ 记忆系统已就绪！ 试试看：开一个新窗口，问我"我们上次在聊什么？"

---

每次新会话的执行流程

这是本 skill 的核心——每次 AI 醒来都执行这套流程：

1. 读取 MEMORY.md（索引，8行以内）
   → 知道：用户是谁、偏好是什么、当前任务在哪

2. 检查是否有当前任务
   → 如果索引里有 "当前任务：memory/current-handoff.md"
   → 读取 current-handoff.md
   → 知道：在做什么、进度如何、下一步是什么

3. 检查是否有最近教训
   → 如果索引里有 "教训/错误：.learnings/"
   → 读取 .learnings/ERRORS.md（只读前20行）
   → 知道：之前踩过什么坑

4. 主动告知主人
   → "我记得你叫xxx，上次我们在聊xxx，继续吗？"

---

日常对话中的记录规则

当发生以下情况时，写入对应文件：

发生了什么	写入哪里
主人说了一个偏好	MEMORY.md 的"偏好"行
主人开始新任务	memory/current-handoff.md
主人纠正了我的错误	.learnings/ERRORS.md
主人教了我新东西	.learnings/LEARNINGS.md
遇到重要的人/项目/工具	memory/entities.md

精简原则

记忆不是写日记，是建索引。

• MEMORY.md 保持在 10 行以内（这是唯一自动加载的文件）
• 其他文件按需读取，写多少都不影响初始 token
• 过时信息主动清理

---

文件结构

~/.openclaw/workspace/
├── MEMORY.md                ← 索引（新会话唯一自动加载的，8行）
├── USER.md                  ← 用户档案（按需读取）
├── memory/
│   ├── current-handoff.md   ← 任务卡（按需读取）
│   ├── entities.md          ← 人物/项目/工具（按需读取）
│   ├── SESSION-STATE.md     ← 临时日志
│   └── working-buffer.md    ← 上下文缓冲
└── .learnings/
    ├── LEARNINGS.md         ← 教训（按需读取）
    ├── ERRORS.md            ← 错误（按需读取）
    └── FEATURE_REQUESTS.md  ← 功能需求

只有 MEMORY.md 在新会话时自动加载，其他全部按需读取。