OpenClaw 终极记忆系统

春晚太难看了，我一边看春晚，一边和 antigravity、cursor 合作，花了四五个小时，设计出了这一套 OpenClaw 终极记忆系统。看着还不错，分享给大家，帮助大家解决 OpenClaw 的失忆问题，也顺便祝大家新年快乐！

方案介绍

参考了哪些方案？

在设计这套记忆系统之前，我们调研了以下方案：

1.2 各方案的优缺点

1.3 我们的方案特点

经过三轮对比分析，我们设计出这套「终极记忆系统」：

核心优势：

特点	说明
幂等性保护	Session ID 去重，避免重复记录
硬性限制	MEMORY.md 上限 80 行 / 5KB，防止膨胀
信号过滤	跳过不足 2 条消息的会话，减少噪音
四维验证	进入长期记忆需满足四个条件
自动备份	修改前自动备份，安全性
精简架构	仅 2 个 Cron Job，易于维护

2. 如何配置

2.1 前置要求

• OpenClaw 已安装并运行
• 已配置 qmd 作为记忆后端
• 工作目录 ~/.openclaw/workspace/

2.2 配置步骤

步骤 1：创建目录结构

mkdir -p ~/.openclaw/workspace/memory/weekly
mkdir -p ~/.openclaw/workspace/memory/archive

步骤 2：创建 MEMORY.md 模板

# Long-Term Memory

> Only write info here that you'd make mistakes without. Event logs stay in daily files.
> Hard limit: 80 lines / 5KB. Must compress before adding when over limit.

## User Preferences

## Active Projects

## Key Decisions

## Important Contacts

步骤 3：更新 AGENTS.md

在 ## Memory 部分添加：

### 🔍 Memory Retrieval (MANDATORY)

When you need to recall past events, **search first — never read all files**:
1. `qmd query "<question>"` — hybrid search with reranking
2. `qmd get <file>:<line> -l 20` — pull only the snippet you need
3. Only fall back to reading files directly if qmd returns nothing

### ✍️ Memory Writing - Don't Wait for Cron

When you make a decision, the user says "I prefer X", or a key task completes → **immediately append to `memory/YYYY-MM-DD.md`**.
Cron captures sessions automatically every few hours, but it's just the safety net. Write it down in the moment.

步骤 4：添加 Cron Job

Job 1: memory-sync（记忆同步）

openclaw cron add \
  --name "memory-sync" \
  --cron "0 10,14,18,22 * * *" \
  --tz "Asia/Shanghai" \
  --session isolated \
  --wake now \
  --delivery none \
  --message "MEMORY SYNC — You are the memory capture agent. Run silently, no notifications.

1. Use sessions_list to get sessions with activity in the last 4 hours
2. Skip isolated sessions
3. For each session, use sessions_history to read conversation content
4. Skip sessions with user_message_count < 2
5. If no valid sessions remain, do nothing — reply ANNOUNCE_SKIP and stop here.
6. Read today's memory/YYYY-MM-DD.md (create if it doesn't exist)
7. Idempotency check: if a session_id's first 8 characters already appear in the file, skip that session
8. For unrecorded sessions, extract: user's key requests, assistant's conclusions/decisions, important action results. Compress each session to 3-10 summary items.
9. Append to the daily file in this format: ## HH:MM session:FIRST8 | N messages
10. Run qmd update && qmd embed
11. Reply ANNOUNCE_SKIP when done."

Job 2: memory-tidy（记忆整理）

openclaw cron add \
  --name "memory-tidy" \
  --cron "0 3 * * *" \
  --tz "Asia/Shanghai" \
  --session isolated \
  --wake now \
  --deliver \
  --message "MEMORY TIDY — You are the memory maintenance agent. You are explicitly authorized to read and modify MEMORY.md in this isolated session.

[Phase 1: Compress]
1. List all date-named files (YYYY-MM-DD.md) in memory/
2. Identify files older than 7 days. If none, skip this phase.
3. Group by natural week, generate memory/weekly/YYYY-MM-DD.md (named after Monday)
4. Extract [Decisions] [Discoveries] [Preferences] [Tasks], tag each with (src: YYYY-MM-DD)
5. Idempotent: if ### YYYY-MM-DD section already exists in weekly file, skip it

[Phase 2: Distill]
6. Read daily files from the last 7 days + current MEMORY.md
7. Identify info worth keeping long-term. All four criteria must be met: (a) agent would make a concrete mistake without it (b) applies to many future conversations (c) self-contained and understandable (d) not duplicated in existing MEMORY.md
8. Reverse check: before writing, ask yourself — what specific error would occur without this? If you can't answer, don't write it.
9. Backup: mkdir -p memory/archive && cp MEMORY.md memory/archive/MEMORY.md.bak-\$(date +%F)
10. Update MEMORY.md. Hard limit: 80 lines. If over, compress/merge existing entries first.

[Phase 3: Archive]
11. Move daily files that have been compressed into weekly summaries to memory/archive/YYYY/

[Wrap-up]
12. Run qmd update && qmd embed
13. If changes were made, send a brief summary. If no changes: reply 🧠 memory-tidy done, no changes."

2.3 文件结构

~/.openclaw/workspace/
├── MEMORY.md              # 长期记忆（80 行上限）
├── AGENTS.md              # 已更新记忆规则
└── memory/
    ├── YYYY-MM-DD.md     # 每日日志
    ├── weekly/           # 周度压缩摘要
    └── archive/          # 归档旧日志

3. 工作流程

┌─────────────────────────────────────┐
│     memory-sync (10/14/18/22)        │
│  • 扫描过去 4 小时的会话            │
│  • 跳过 <2 条消息的会话             │
│  • Session ID 去重                  │
│  • 写入 memory/YYYY-MM-DD.md        │
│  • qmd update && qmd embed         │
└─────────────────────────────────────┘
                │
                ▼
┌─────────────────────────────────────┐
│      memory-tidy (每日 3 AM)         │
│  Phase 1: 压缩 → 周度摘要           │
│  Phase 2: 蒸馏 → 更新 MEMORY.md     │
│  Phase 3: 归档 → 移动旧文件        │
│  qmd update && qmd embed           │
└─────────────────────────────────────┘

4. 验证

配置完成后：

1. 运行 openclaw cron list 确认 Job 已注册
2. 发起几轮对话，等待下一个 sync 窗口（10/14/18/22 点）
3. 检查 memory/2026-02-16.md 是否生成
4. 测试 qmd query "关键词" 搜索功能
5. 一周后检查 memory/weekly/ 是否有压缩摘要