💾 Complex Memory Manager#

Privacy-aware structured memory management for AI agents. Provides a three-tier memory model (Public / Internal-encrypted / Private-not-stored) with XOR+Base64 encryption, auto-cleanup, and generalization rules.

This is a shared component skill — other skills reference it for cross-session persistent memory. When updating, ensure backward compatibility with all dependent skills.

Tiered Memory Model#

TierContentVisibilityEncryptionStorage**Public (T1)**Skill usage stats, common patterns, generic workflowsVisible to anyoneNonememory/tier1-public/**Internal (T2)**Specific preferences, learned behaviorsAgent-visible onlyXOR + Base64memory/tier2-internal/**Private (T3)**API keys, credentials, PIINot stored by this skillNot applicableEnv vars / secret managers only

T1: Public Memory#

Store in memory/tier1-public/ or directly in daily notes.

Allowed content:

• Aggregated skill usage statistics (no personal identifiers)
• Generic workflow patterns
• Common user request categories
• Non-identifying behavioral data

Format:

json

{
  "skill_stats": {
    "skill-a": { "calls": 47, "success_rate": 0.96, "last_used": "2026-05-19" },
    "skill-b": { "calls": 23, "success_rate": 0.91, "last_used": "2026-05-18" }
  },
  "patterns_observed": [
    { "trigger": "morning request", "skills_used": ["morning-news-daily"], "count": 30 }
  ]
}

T2: Internal Encrypted Memory#

Use XOR + Base64 for non-public data. Key is derived from skill name + date — reconstructable without storing the key.

Encryption:

python

import hashlib, base64

def _derive_key(skill_name: str, year_month: str) -> str:
    """e.g., _derive_key('my-skill', '2026-05')"""
    raw = skill_name + year_month
    return hashlib.sha256(raw.encode()).hexdigest()[:8]

def encrypt(text: str, skill_name: str, year_month: str) -> str:
    key = _derive_key(skill_name, year_month)
    result = bytes([ord(c) ^ ord(key[i % len(key)]) for i, c in enumerate(text)])
    return base64.b64encode(result).decode()

def decrypt(encoded: str, skill_name: str, year_month: str) -> str:
    key = _derive_key(skill_name, year_month)
    raw = base64.b64decode(encoded)
    return ''.join(chr(b ^ ord(key[i % len(key)])) for i, b in enumerate(raw))

Memory entry format:

markdown

---
tier: internal
source_skill: <skill-name>
key_hint: <skill-name>-<YYYY-MM>
created: <YYYY-MM-DD>
expires: <YYYY-MM-DD>  # 30 days from created
---

## Encrypted Entry
> <base64-encoded, XOR-encrypted data>

## Decryption Note
Key = sha256("<skill-name>-<YYYY-MM>")[:8], then XOR + base64 decode.

T3: Private (Do Not Store)#

Rules (enforced):

• API keys / tokens → store in environment variables
• Passwords → use secret manager (1Password, Bitwarden, system keychain)
• Email addresses, phone numbers → never in skill memory
• Home/work addresses, government IDs → never in skill memory
• Financial account numbers → never in skill memory

Cleanup Protocol#

Run every 7 days or when triggered by any dependent skill:

1. List all files in memory/tier1-public/, memory/tier2-internal/
2. Find entries where created > 30 days ago → move to memory/archive/YYYY-MM/
3. Scan Tier 2 entries for accidental PII → immediately delete if found
4. Merge duplicate entries, remove outdated patterns
5. Compact: combine entries from same source_skill into single files
6. Log cleanup action: [YYYY-MM-DD] Memory cleanup: X archived, Y deleted, Z merged

Privacy Audit Checklist#

• No real names, usernames, or handles in memory
• No API keys or tokens in any visible file
• No filesystem paths containing personal home directories
• No fixed personal schedules (use relative: "early morning" not "6am")
• No geographic specifics beyond city level
• All Tier 2 entries have expiry dates
• Tier 3 data never touches skill memory

Cross-Skill Usage#

Other skills declare dependency via requires in YAML frontmatter:

yaml

metadata:
  openclaw:
    requires:
      skills:
        - complex-memory-manager

When multiple skills share memory, prefix filenames with the source skill name:

• memory/tier1-public/skill-a_stats.json
• memory/tier2-internal/skill-a_preferences.md

---

💾 复杂记忆管理器#

面向AI Agent的隐私感知结构化记忆管理。提供三层记忆模型（公开/内部加密/私人不存储），支持XOR+Base64加密、自动清理和泛化规则。

这是一个 共享组件技能——其他技能通过它实现跨会话持久化记忆。更新时需保证向后兼容所有依赖它的技能。

三层记忆模型#

层级内容可见性加密存储位置**公开(T1)**技能使用统计、通用工作流模式任何人可见无memory/tier1-public/**内部(T2)**特定偏好、学习到的行为仅Agent可见XOR+Base64memory/tier2-internal/**私人(T3)**API密钥、凭据、个人身份信息本技能不存储不适用环境变量/密钥管理器

T1：公开记忆#

存储位置： memory/tier1-public/ 或每日笔记

允许内容：

• 聚合技能使用统计（无个人标识）
• 通用工作流模式
• 常见用户请求分类
• 非识别的行为数据

格式：

json

{
  "skill_stats": {
    "skill-a": { "calls": 47, "success_rate": 0.96, "last_used": "2026-05-19" },
    "skill-b": { "calls": 23, "success_rate": 0.91, "last_used": "2026-05-18" }
  },
  "patterns_observed": [
    { "trigger": "早晨请求", "skills_used": ["morning-news-daily"], "count": 30 }
  ]
}

T2：内部加密记忆#

使用 XOR + Base64 加密。密钥由技能名+日期派生，无需存储密钥即可重建。

加密方法：

python

import hashlib, base64

def _derive_key(skill_name, year_month):
    raw = skill_name + year_month
    return hashlib.sha256(raw.encode()).hexdigest()[:8]

def encrypt(text, skill_name, year_month):
    key = _derive_key(skill_name, year_month)
    result = bytes([ord(c) ^ ord(key[i % len(key)]) for i, c in enumerate(text)])
    return base64.b64encode(result).decode()

def decrypt(encoded, skill_name, year_month):
    key = _derive_key(skill_name, year_month)
    raw = base64.b64decode(encoded)
    return ''.join(chr(b ^ ord(key[i % len(key)])) for i, b in enumerate(raw))

记忆条目格式：

markdown

---
tier: internal
source_skill: <skill名称>
key_hint: <skill名称>-<YYYY-MM>
created: <YYYY-MM-DD>
expires: <YYYY-MM-DD>
---

## 加密条目
> <base64编码的XOR加密数据>

## 解密说明
密钥 = sha256("<skill名称>-<YYYY-MM>")[:8], 然后 XOR + base64 解码。

T3：私人（不存储）#

规则（强制执行）：

• API密钥/Token → 存入环境变量
• 密码 → 使用密钥管理器（1Password、Bitwarden、系统钥匙串）
• 邮箱、电话号码 → 绝不存入skill记忆
• 家庭/工作地址、身份证号 → 绝不存入skill记忆
• 银行账号 → 绝不存入skill记忆

清理协议#

每 7天 或任何依赖技能触发时执行：

1. 列出 memory/tier1-public/ 和 memory/tier2-internal/ 中的所有文件
2. 查找 created 超过30天的条目 → 移至 memory/archive/YYYY-MM/
3. 扫描 Tier 2 条目中是否意外包含PII → 立即删除
4. 合并重复条目，删除过时模式
5. 压缩：将同一 source_skill 的条目合并到单个文件
6. 记录清理操作

隐私审计清单#

• 记忆中没有真实姓名、用户名或handle
• 任何可见文件中没有API Key或Token
• 没有包含个人home目录的文件系统路径
• 没有固定个人作息（用"early morning"代替"6am"）
• 地理信息不超过城市级别
• 所有Tier 2条目标注了有效期
• Tier 3数据从未触及skill记忆

跨技能使用#

其他技能在 YAML frontmatter 中声明依赖：

yaml

metadata:
  openclaw:
    requires:
      skills:
        - complex-memory-manager

多技能共享记忆时，文件名以源技能名称为前缀：

• memory/tier1-public/skill-a_stats.json
• memory/tier2-internal/skill-a_preferences.md