Claude Memory Pro

进化版统一记忆系统，融合Claude Code的5层记忆架构+Token预算管理+智能成本追踪。当用户要求记忆跨会话信息、智能管理Token预算、优化AI调用成本、构建个人知识库、跨项目学习进化时使用。

Security Scan

The skill is an instruction-only memory manager that is internally consistent with its stated purpose: it describes a local file-based 5-layer memory architecture and token-budget logic and does not request extra credentials, installs, or external endpoints.

Purpose & Capability

Name/description (a cross-session memory and token-budget system) matches the instructions: all runtime steps involve creating/reading/writing local memory files and applying token-budget heuristics. No unrelated credentials, binaries, or external services are requested.

Instruction Scope

SKILL.md instructs the agent to create and read/write files under a memory/ directory (e.g., MEMORY.md, hot/current_turn.md, typed/*). This is appropriate for a local memory manager, but the instructions do not specify storage location, access controls, encryption, or retention policies—so users should be aware that personal or sensitive data could be written to disk unless the runtime enforces safeguards.

Install Mechanism

No install spec and no code files — instruction-only. Nothing is downloaded or written by an installer, which minimizes supply-chain risk.

Credentials

No environment variables, credentials, or external endpoints are requested. The declared requirements are proportional to the described functionality.

Persistence & Privilege

always=false and no modification of other skills or global agent config. The skill's persistence is limited to instructions that write to a local memory/ directory (expected for this purpose).

Scan Findings in Context

[regex-scan-none] expected: The static scanner had no findings because this is an instruction-only skill with no code files to analyze; that is expected for this type of skill.

Assessment

This skill appears coherent and does what it says (a local, file-based memory and token-budget system). Before enabling it: (1) confirm where the memory/ directory will be created and whether the runtime isolates or restricts it; (2) avoid storing secrets or sensitive data in plain text files created by the skill, or require encryption/access controls; (3) verify retention and deletion policies (how long memories are kept and how to purge them); (4) confirm your agent platform won't autonomously transmit these memory files to external endpoints (SKILL.md contains no network calls, but the agent runtime could still send data); and (5) test in an isolated environment first. If you need help specifying encryption, access control, or retention rules for the memory directory, request those additions before installing.

进化版统一记忆系统 - Claude Code核心技能提炼与升级

核心能力

1. 5层记忆架构 - 热记忆 → 会话记忆 → 长期记忆 → 实体银行 → 反思层
2. Token预算管理 - 智能追踪、边际收益检测、自动决策
3. 成本追踪优化 - 多维度统计、实时显示、优化建议
4. 跨会话学习 - 从历史交互中持续进化

记忆类型（4类）

类型	用途	示例
user	用户偏好、背景	语言习惯、工作风格
feedback	工作指导、纠正	代码规范、回复偏好
project	项目上下文、决策	技术选型、截止日期
reference	外部系统指针	API文档、数据库地址

5层记忆架构

memory/
├── hot/                     # Layer 1: 热记忆（当前对话）
│   └── current_turn.md
├── session/                 # Layer 2: 会话记忆
│   └── session_YYYY-MM-DD.md
├── typed/                   # Layer 3: 长期记忆
│   ├── user/
│   ├── feedback/
│   ├── project/
│   └── reference/
├── bank/                    # Layer 4: 实体银行
│   ├── entities/
│   └── procedures/
└── reflections/             # Layer 5: 反思层

Token预算管理

预算阈值配置

const BUDGET_CONFIG = {
  completionThreshold: 0.9,    // 90% 触发完成检查
  diminishingThreshold: 500,   // 边际收益递减阈值
  maxContinuations: 10,        // 最大继续次数
  warningThreshold: 0.75       // 75% 警告阈值
}

决策流程

Token使用检查
      ↓
< 75% → 正常继续
      ↓
< 90% → 继续 + nudge
      ↓
边际递减？ → 停止（边际收益递减）
      ↓
之前有继续？ → 停止（正常完成）

边际收益检测

function isDiminishing(tracker): boolean {
  return tracker.continuationCount >= 3 &&
         delta < 500 &&
         lastDelta < 500
}

成本追踪

成本指标

类型	说明	优化价值
input_tokens	输入tokens	高
output_tokens	输出tokens	中
cache_read	缓存命中	低成本
cache_creation	缓存创建	一次性成本
web_search	搜索请求	按需优化

优化策略

1. 复用缓存 - 相同上下文用cache_read
2. 压缩上下文 - 减少input_tokens
3. 精简输出 - 减少output_tokens
4. 批量操作 - 减少API调用

记忆保存流程

对话前

1. 读取MEMORY.md获取索引
2. 按需加载typed/相关记忆

对话中

• 热记忆写入hot/current_turn.md
• 实时成本监控

对话后

1. 评估值得保留的信息
2. 选择正确的记忆层
3. 更新MEMORY.md索引

决策树

用户给反馈/纠正
├─ 关于用户本身 → typed/user/
├─ 关于工作方式 → typed/feedback/
├─ 关于项目 → typed/project/
└─ 关于外部系统 → typed/reference/

了解用户背景 → typed/user/
项目决策/变化 → typed/project/
外部系统位置 → typed/reference/
复杂档案 → bank/entities/
可复用流程 → bank/procedures/
自我反思 → reflections/

不保存内容

• 代码模式（可从代码推导）
• Git历史（git log权威）
• 临时任务状态（用tasks）
• 已文档化的内容（CLAUDE.md）
• 未经证实的推断
• 整段对话记录
• 大段代码

快速开始

初始化

mkdir -p memory/{hot,session,typed/{user,feedback,project,reference},bank/{entities,procedures},reflections}

创建记忆

---
name: 用户回复偏好
type: feedback
created: 2026-04-02
---

用户不希望在每次回复末尾总结。

**Why:** 用户说"我能直接读懂diff"
**How to apply:** 保持简洁，不添加总结性语句

维护检查

• 每月检查typed/过时记忆
• 每季度清理reflections/
• 验证实体银行信息准确性
• 检查Token预算使用效率
• 评估成本优化效果

与其他技能对比

特性	基础版	Pro版
记忆层数	3层	5层
Token管理	无	智能预算
成本追踪	无	多维度统计
跨会话学习	基础	反思进化
优化建议	无	自动生成