OpenClaw记忆管理系统v2

OpenClaw 记忆管理系统，三层架构（NOW.md + 每日日志 + 知识库），支持 heartbeat 写入、夜间反思、CRUD 验证、主动遗忘。Based on Ray Wang's 30-day production experience.

Security Scan

The skill's description matches a file-backed memory system, but its runtime instructions direct broad reading of session messages and automatic CRUD writes under ~/.openclaw without declaring required tools or clarifying what 'sessions' are — this raises privacy and persistence concerns the user should review before enabling autonomous runs.

Purpose & Capability

Name/description and SKILL.md consistently describe a three-layer local file-based memory system. File I/O, indexing, and nightly consolidation are coherent with the claimed purpose.

Instruction Scope

Instructions explicitly direct the agent to scan "active session messages", extract information, and perform automated CRUD updates to files in ~/.openclaw/workspace (NOW.md, daily logs, INDEX.md, and category files). "Session messages" is not defined (could be chat transcripts, system logs, or other sensitive content). The skill also references helper tools/commands (memlog.sh, Write NOW.md, printf >> file) but provides no implementation or explicit permission model. These behaviors expand scope to reading and persistently modifying potentially sensitive conversation data.

Install Mechanism

Instruction-only skill with no install spec or external downloads. No code is written to disk by the registry; risk from install mechanism is low.

Credentials

The skill requests no environment variables, credentials, or config paths in metadata. However, it expects access to the user's home directory (~/.openclaw/workspace) and to application 'sessions' implicitly; the metadata does not declare or describe these required access scopes.

Persistence & Privilege

The skill instructs repeated autonomous behaviors (heartbeat every 30–60 min, nightly CRUD updates, weekly GC) and persistent writes/archives under ~/.openclaw. While 'always' is false, the default model-invocation (autonomous runs allowed) plus persistent file writes increases blast radius if the agent is permitted to run unattended. The SKILL.md lacks clear safeguards, consent/confirmation steps, or a defined sandbox.

What to consider before installing

This skill is coherent with a file-backed memory system, but it will read conversation/session data and automatically modify files under ~/.openclaw/workspace. Before installing or enabling autonomous invocation: - Understand what "session messages" means in your environment — does it include private chats, API messages, or system logs? If unsure, do not allow autonomous runs. - Expect the skill to create and edit files in ~/.openclaw/workspace (NOW.md, daily logs, INDEX.md, category files). Back up important data and review created files manually. - The SKILL.md references helper commands (memlog.sh, Write NOW.md) but provides no code; verify any scripts or tools it expects before use. - If you need stricter control, disable autonomous invocation for this skill or run it in a sandboxed account/VM and inspect its outputs. - Consider adding access controls (restrict which conversation streams it can read), or require explicit user confirmation before any DELETE/ARCHIVE/UPDATE operations. If you want, I can list exact prompts or policies to add so the skill asks for explicit confirmation before modifying files, or draft a minimal-safe wrapper script that performs writes only after human approval.

OpenClaw Memory System v1.0

基于 Ray Wang 的生产实践经验，30天迭代的记忆管理系统。

核心原则

文件 = 事实来源

• Context Window 是工作台，文件才是仓库
• 不写进文件的东西 = 你从来不知道的东西

三层架构

1. 短期层：NOW.md

• 用途：Agent 的"工作台"，当前状态、优先级、阻塞项
• 特点：覆写式（每次 heartbeat 更新）
• 位置：~/.openclaw/workspace/NOW.md

2. 中期层：memory/YYYY-MM-DD.md

• 用途：事件流水，记录当天发生的一切
• 特点：追加式，永不覆写
• 位置：~/.openclaw/workspace/memory/

3. 长期层：INDEX.md → 结构化知识库

• 用途：提炼后的可复用知识
• 目录：lessons/、decisions/、people/、projects/、preferences/
• 位置：~/.openclaw/workspace/memory/INDEX.md

目录结构

workspace/
├── NOW.md                  # 短期：状态仪表盘（覆写式）
├── MEMORY.md               # 指向 INDEX.md 的指针
├── AGENTS.md               # Agent 操作手册
├── HEARTBEAT.md            # Heartbeat 执行流程
│
└── memory/
    ├── INDEX.md            # 知识导航（启动时必读）
    ├── YYYY-MM-DD.md       # 每日日志（追加式）
    │
    ├── decisions/          # 战略决策记录
    ├── lessons/            # 可复用教训
    ├── people/             # 人物画像
    ├── projects/           # 项目状态
    ├── preferences/        # 用户偏好
    ├── reflections/        # 每日自省
    ├── actions/            # 任务卡片
    │
    └── .archive/           # 冷存储（不索引）

Frontmatter 规范

---
title: "主题"
date: 2026-03-24
category: lessons|person|decision
priority: 🔴|🟡|⚪
status: active|superseded|conflict
last_verified: 2026-03-24
tags: []
---

状态流转

active ──→ superseded    (被更新版本取代)
active ──→ conflict      (发现矛盾信息，待人工裁决)
conflict ──→ active      (人工裁决后恢复)

生命周期

日间（Heartbeat 每 30-60 分钟）

1. 扫描活跃 session 消息
2. 提取重要信息写入日志
3. 路由到对应知识文件（轻量去重）
4. 刷新 NOW.md

23:30 日志同步

• 扫描全天所有 session 补漏

23:45 夜间反思

1. 读取上下文（今日日志 + NOW.md + 相关知识文件）
2. 写反思（计划 vs 实际 + 教训 + 明天改变）
3. 清理日志（去重）
4. CRUD 回写知识库 ← 核心
5. 过时扫描（>30 天标记 [⚠️ STALE]）

每周日 00:00 GC 归档

• 日志 >30 天 + 无引用 → .archive/
• 反思 >30 天 → 归档
• 核心知识（decisions/、lessons 🔴、people/）永不归档

温度模型

温度	状态	处理
T > 0.7	🔥 Hot	保持活跃索引
0.3 < T ≤ 0.7	🌤️ Warm	保留但检索降权
T ≤ 0.3	🧊 Cold	移至 .archive/

Temperature = w_age × age_score + w_ref × ref_score + w_pri × priority_score

写入工具

工具	用途	场景
memlog.sh "Title" "Body"	日记追加	事件、完成、决策
Write NOW.md	覆写	仅 NOW.md
printf >> file	追加	lessons/ people/ 等

写入禁忌

• ❌ 硬编码时间戳（用脚本自动取）
• ❌ 用 Edit 修改 memory/ 文件
• ❌ 用 Write 覆写已有 memory/ 文件
• ❌ 写无实质内容的噪音
• ❌ 不读就写知识文件

检索策略

L1: 扫 INDEX.md → 定位目标文件

• 适用：知道要找什么类别

L2: 直接读取目标文件

• 适用：已知具体文件路径

L3: QMD 语义搜索

• 适用：不确定信息在哪个文件

CRUD 验证

写入知识文件时必须：

1. READ - 读取目标文件当前内容
2. COMPARE - 比较新知识与已有内容
   • 已有覆盖新知识 → NOOP (跳过)
   • 新知识更新 → UPDATE (旧版标记 superseded)
   • 新知识矛盾 → CONFLICT (两版保留 + 标记)
   • 全新的知识 → ADD (追加)
3. METADATA - 更新 last_verified 日期

记忆幻觉防护

类型	描述	例子
编造	写入从未发生过的事情	Agent 声称"昨天做了 X"
错误	写入不准确的信息	把 A 的观点归因给 B
冲突	同一事实的两个矛盾版本并存	说"用方案 A"又说"用方案 B"
遗漏	重要信息没有被记录	关键决策没写进文件

快速上手（最小配置）

workspace/
├── NOW.md           # 状态板
├── AGENTS.md        # Agent 操作手册
└── memory/
    └── YYYY-MM-DD.md  # 每日日志

Obsidian 集成

将 memory/ 目录设为 Obsidian vault：

• 启用核心插件：File explorer、Graph view、Backlinks、Daily notes
• 设置 useMarkdownLinks: false，启用 [[wiki-link]]
• 开启"显示隐藏文件"浏览归档

设计灵感

• 人脑记忆整合（睡眠时整合）
• 艾宾浩斯遗忘曲线
• Stanford Generative Agents
• Mem0 / MemGPT / HaluMem