Tiered Recall

分层回忆系统 - 解决上下文长度限制，保持项目延续性。每次新session自动加载核心记忆+最近日志+活跃项目，支持手动深度回忆。索引含10字内摘要，方便区分同名条目。

Security Scan

The skill's code generally matches a local 'tiered memory/recall' purpose, but there are surprising/incoherent pieces (a hardcoded Windows System32 path, a missing script referenced in docs, and brittle assumptions) that warrant review before installing.

Purpose & Capability

Name, description, and most scripts (build-index.py, load.py) align with a local workspace memory/indexing and recall feature: they scan memory/*.md, build an index, detect active projects, and load snippets for sessions. The SKILL.md and README describe the same behavior and token budgeting.

Instruction Scope

Most runtime instructions stay within the stated purpose (reading MEMORY.md, memory/*.md, .tiered-recall files and project key files). However scripts/check_index.py reads a hardcoded absolute path 'C:/Windows/System32/UsersAdministrator.openclawworkspace/.tiered-recall/index.json' outside the workspace — this is inconsistent with the described workspace-local behavior and could unexpectedly access system or other users' files if that path exists. The SKILL.md also references a script update-projects.py that is not present in the file manifest (documentation/code mismatch).

Install Mechanism

There is no install spec (instruction-only install), so nothing is downloaded or installed by the platform. The risk is limited to the included code files being executed locally; they do not fetch remote code or write nonstandard system-wide binaries.

Credentials

The skill declares no environment variables or credentials (good). The included scripts read local workspace files and, in load.py, will load and preview key files (first 500 chars) — reasonable for a recall tool but worth noting: any sensitive content in your workspace could be read and returned by the skill if invoked. No network exfiltration is present in the code, but the agent could still transmit loaded content to external services depending on agent behavior (not shown here).

Persistence & Privilege

The skill is not always-enabled and allows normal model invocation (defaults). It does not request system-wide persistence or modify other skills' configurations according to the provided files.

What to consider before installing

This skill appears to implement a local 'tiered recall' system and mostly behaves as described, but review and fix two issues before trusting it with real data: 1) scripts/check_index.py contains a hardcoded Windows System32 path (C:/Windows/System32/UsersAdministrator.openclawworkspace/.tiered-recall/index.json) that is inconsistent with the workspace-local model and could read unexpected files — remove or adapt that path to your workspace or delete the script. 2) The documentation mentions update-projects.py but that file is missing; expect runtime errors or incomplete features. Also: the scripts read and print local files (MEMORY.md, recent logs, project key files and previews). Although these scripts don't make network calls themselves, any sensitive data in those files could be exposed by the agent if it forwards content externally. Recommended actions: run the code in a sandbox or review/modify the scripts (remove/check hardcoded paths, add input validation, handle missing keys safely) before installing; test on a non-sensitive workspace; and prefer to keep any secrets out of the memory/ memory/ directory if you enable automated loading. My confidence is medium — the odd hardcoded path and missing file are clear red flags but do not prove malicious intent on their own.

分层回忆系统 - 解决大模型上下文长度限制，保持项目延续性

---

🎯 核心问题

大模型上下文有限（约20万token），复杂项目可能跨多天、多窗口进行。每次新session开始时，如何快速恢复上下文，保持工作延续性？

常见痛点：

• 新开窗口，之前的项目背景丢失
• 跨天任务，第二天不记得昨天做了什么
• 多项目并行，切换时混乱
• 手动回顾太慢，浪费时间

---

🚀 解决方案：分层回忆

自动加载策略（每次新session）

层级	内容	Token预算	加载条件
🔴 L0 核心	MEMORY.md	~4k	始终加载
🟠 L1 近期	最近2天日志	~10k	始终加载
🟡 L2 项目	活跃项目文件	~5k	自动检测
🟢 L3 索引	记忆索引	~1k	始终加载
总计		~20k

总预算：~20k token，约占总上下文的10%

---

📂 文件结构

workspace/
├── MEMORY.md              # L0 核心记忆（长期）
├── memory/
│   ├── 2026-03-25.md      # 每日日志
│   ├── 2026-03-24.md
│   ├── 2026-03-23.md
│   └── ...
├── .tiered-recall/
│   ├── index.json         # 记忆索引
│   ├── projects.json      # 活跃项目清单
│   └── state.json         # 加载状态
└── skills/
    └── tiered-recall/
        └── SKILL.md       # 本技能

---

🔄 自动加载流程

Step 1: 检测触发条件

自动触发：

• 新session开始
• 用户说"继续之前的项目"
• 用户提到项目名称

Step 2: 加载L0核心记忆

# MEMORY.md 加载
- 用户基本信息
- 重要偏好
- 长期决策
- 关键项目记录

Step 3: 加载L1近期日志

# 最近2天日志
- memory/2026-03-25.md
- memory/2026-03-24.md

自动提取：

• 昨天做了什么
• 今天待办事项
• 未完成任务

Step 4: 加载L2活跃项目

// .tiered-recall/projects.json
{
  "active": [
    {
      "name": "合成天选打工人",
      "path": "games/merge-worker/",
      "lastActive": "2026-03-24",
      "keyFiles": ["index.html", "README.md"]
    },
    {
      "name": "搞钱特战队",
      "path": "products/AI-Guide/",
      "lastActive": "2026-03-24",
      "keyFiles": ["chapter1/"]
    }
  ]
}

Step 5: 加载L3记忆索引

// .tiered-recall/index.json
{
  "topics": {
    "游戏开发": ["memory/2026-03-23.md:line100-200", "memory/2026-03-24.md:line50-150"],
    "搞钱特战队": ["memory/2026-03-24.md:line200-400"],
    "OpenClaw变现": ["memory/2026-03-24.md:line400-600"]
  },
  "lastUpdated": "2026-03-25T09:00:00"
}

---

🎮 手动深度回忆

当默认加载不够时，用户可以要求深度回忆：

指令语法

指令	作用	示例
继续回忆	加载更多相关记忆	"继续回忆游戏项目"
回忆 [项目名]	加载该项目全部记忆	"回忆搞钱特战队"
回忆 [天数]	加载指定天数日志	"回忆最近7天"
回忆 [日期]	加载指定日期日志	"回忆3月20日"
回忆 [关键词]	按关键词搜索记忆	"回忆抖音小游戏"

执行流程

用户: "回忆搞钱特战队"
  ↓
1. 搜索 index.json 找到相关条目
2. 加载相关日期日志的相关段落
3. 加载项目文件
4. 生成项目上下文摘要
5. 输出给用户确认

---

📊 Token 预算控制

默认加载（~20k token）

L0 MEMORY.md        ████████░░  4k
L1 最近2天日志      ████████████████████████░░  10k
L2 活跃项目         ██████████░░  5k
L3 记忆索引         ██░░  1k
                    ─────────────────────
                    总计 ~20k token

深度回忆时（动态调整）

用户要求: "回忆搞钱特战队全部记忆"
  ↓
预估: 约50k token
  ↓
策略: 
1. 先加载摘要（~10k）
2. 询问用户是否继续
3. 分批加载，每批~20k

---

🛠️ 脚本命令

生成记忆索引

python skills/tiered-recall/scripts/build-index.py

功能：

• 扫描 memory/ 目录所有日志
• 提取关键词和主题
• 生成 index.json

更新活跃项目

python skills/tiered-recall/scripts/update-projects.py

功能：

• 扫描最近日志中提到的项目
• 更新 projects.json

完整加载

python skills/tiered-recall/scripts/load.py [--deep] [--project NAME]

参数：

• --deep 深度加载，突破默认token限制
• --project NAME 只加载指定项目

---

📝 使用示例

场景1：新session自动加载

[新session开始]
AI: 
📖 加载核心记忆... ✓ (MEMORY.md)
📅 加载近期日志... ✓ (2026-03-25, 2026-03-24)
📂 加载活跃项目... ✓ (合成天选打工人, 搞钱特战队)
📑 加载记忆索引... ✓

=== 今日上下文 ===
1. 昨天完成：搞钱特战队商业模式确定
2. 今天待办：制作引流资料包
3. 活跃项目：合成天选打工人、搞钱特战队
================

早安！今天要继续哪个项目？

场景2：手动深度回忆

用户: 回忆合成天选打工人全部记忆

AI:
🔍 搜索中...
  找到相关记录: 2026-03-23, 2026-03-24
  
📖 加载2026-03-23相关段落...
  - 游戏版本迭代 v2.6 → v2.24
  - 修复7个Bug
  - 11人团队评估通过
  
📖 加载2026-03-24相关段落...
  - 准备软著材料
  - 抖音小游戏适配讨论

📂 加载项目文件...
  - games/merge-worker/index.html (108KB)
  - games/merge-worker/VERSION_HISTORY.md

=== 项目摘要 ===
游戏名称：天选打工人
当前版本：v2.24
状态：已完成，可上线
下一步：软著申请、抖音上线
================

需要继续回忆更多细节吗？

场景3：跨天任务延续

用户: 继续昨天的引流资料包制作

AI:
🔍 查找昨天记录...
  2026-03-24: "制作免费引流资料包"
  
📖 加载相关上下文...
  - 第一章教程已完成（17000字）
  - 封面v8已完成
  - 小红书账号已设置
  
📂 加载工作文件...
  - products/AI-Guide/chapter1/

好的，继续制作引流资料包。昨天第一章已完成，今天要做的是：
1. 整理成PDF
2. 添加封面
3. 准备下载链接

开始吗？

---

⚙️ 配置选项

.tiered-recall/config.json

{
  "defaultLayers": ["L0", "L1", "L2", "L3"],
  "recentDays": 2,
  "maxTokensPerLayer": {
    "L0": 4000,
    "L1": 10000,
    "L2": 5000,
    "L3": 1000
  },
  "deepRecallBudget": 50000,
  "autoLoadOnNewSession": true
}

---

🔧 技术实现

记忆索引生成算法

def build_index(memory_dir):
    """扫描日志，提取主题和关键词"""
    index = {"topics": {}, "lastUpdated": datetime.now().isoformat()}
    
    for file in sorted(memory_dir.glob("*.md"), reverse=True):
        content = file.read_text(encoding="utf-8")
        
        # 提取标题和章节
        sections = extract_sections(content)
        
        # 识别主题
        for section in sections:
            topic = classify_topic(section["title"])
            if topic:
                index["topics"].setdefault(topic, []).append({
                    "file": file.name,
                    "lines": section["lines"],
                    "summary": summarize(section["content"])
                })
    
    return index

活跃项目检测

def detect_active_projects(memory_dir, days=7):
    """从最近日志中检测活跃项目"""
    projects = {}
    
    for file in get_recent_files(memory_dir, days):
        content = file.read_text(encoding="utf-8")
        
        # 匹配项目关键词
        for match in PROJECT_PATTERNS:
            if match["pattern"] in content:
                project = match["name"]
                projects[project] = {
                    "lastMentioned": file.stem,
                    "path": match.get("path"),
                    "keyFiles": match.get("keyFiles", [])
                }
    
    return projects

---

🎯 最佳实践

1. 保持MEMORY.md精简

• ✅ 只放长期重要的信息
• ❌ 不要放每日琐事
• 目标：< 200行

2. 每日日志结构化

# 2026-03-25 日志

## 📌 今日重点
- [x] 任务A
- [ ] 任务B

## 🎮 项目：合成天选打工人
进度：v2.24完成
下一步：软著申请

## 💰 项目：搞钱特战队
进度：商业模式确定
下一步：制作资料包

3. 项目关键词规范

在日志中使用统一的项目名称：

• ✅ "合成天选打工人"、"搞钱特战队"
• ❌ "那个游戏"、"赚钱的项目"

4. 定期清理索引

# 每周运行一次
python skills/tiered-recall/scripts/build-index.py --clean

---

📈 效果对比

场景	无分层回忆	有分层回忆
新session启动	手动回顾5-10分钟	自动加载，即刻恢复
跨天任务	"我们昨天做什么来着"	"继续昨天的X任务"
多项目切换	混乱、遗忘	自动加载项目上下文
Token消耗	随机、不稳定	可控、~20k预算

---

🤝 与其他技能协作

技能	协作方式
self-improving-enhancement	共享记忆结构，增量更新
team-collab	团队会议记录自动索引
proactive-agent	心跳检查时更新索引

---

📝 Changelog

v1.0.0 (2026-03-25)

• ✨ 初始版本
• 🔄 支持分层自动加载
• 🔍 支持手动深度回忆
• 📊 Token预算控制
• 🗂️ 记忆索引生成

---

💬 Feedback

• Issues: GitHub Issues
• Rate: clawhub star tiered-recall
• Update: clawhub sync tiered-recall

---

Made with 🧠 by davidme6