Memory Pro

> 精选层·主动精馏：主动从原始会话和笔记中提取知识，沉淀到长期记忆。。触发词：skill, 优化。

Security Scan

The skill's instructions ask the agent to write files and run git commits (and reference exec/read/write tools) but the metadata omits required binaries/config and gives no limits on where or how it will write—this mismatch is concerning.

Purpose & Capability

The skill claims to be a memory/refinement layer (writing distilled knowledge into long-term notes). That purpose legitimately requires filesystem write access and a VCS like git, but the skill metadata declares no required binaries or config paths (e.g., no 'git' requirement, no target directory). The absence of these declarations is an incoherence between stated purpose and requested/declared capabilities.

Instruction Scope

SKILL.md explicitly instructs creating/updating notes in notes/<...>.md and running 'git add' and 'git commit'. It also mentions using existing tools (read/write/exec). While writing and committing notes aligns with the skill purpose, the instructions do not constrain target paths, do not limit what may be written, and reference 'exec' which could enable arbitrary command execution if the agent uses it. The instructions also don't state whether commits are local or pushed (remote operations would use remote credentials).

Install Mechanism

This is an instruction-only skill with no install spec or code to download, which minimizes install-time risk.

Credentials

No environment variables or credentials are declared, which is reasonable for simple local note-writing. However, because the skill uses git commands, it will implicitly interact with the user's git configuration and possibly remote credentials (if push is later performed). The metadata should have declared that git and filesystem access are required so users can evaluate credential exposure.

Persistence & Privilege

always:false (good), but the skill performs persistent actions (creating/updating files and committing them). Because the agent can invoke skills autonomously, the combination of automatic triggers and write/exec capabilities could result in unexpected persistent changes to the user's repository or filesystem unless constrained. The skill does not declare path or sandboxing limits.

What to consider before installing

This skill is coherent in purpose (saving distilled session notes) but the SKILL.md instructs the agent to create/modify files and run git commits while the metadata declares no required binaries or paths. Before installing: (1) confirm where 'notes/' will be created and that it will be confined to a safe directory or sandbox; (2) ensure you have git installed and understand whether commits will be pushed to a remote (pushes could use remote credentials); (3) ask the skill author to declare required binaries (git) and explicitly state/limit the filesystem paths the skill may write to; (4) consider disabling autonomous invocation or requiring user confirmation for write/commit actions until you trust behavior; and (5) test the skill in a disposable workspace or repo to observe exactly what it writes and commits.

SKILL.md — memory-pro (CURATED · 主动精馏层)

精选层·主动精馏：主动从原始会话和笔记中提取知识，沉淀到长期记忆。

与 memory-setup（初始化）配合，memory-setup 负责"读取"，memory-pro 负责"写入"和"提炼"。

触发条件

满足以下任一条件时激活：

1. 会话结束：一段完整会话结束后，主动总结并沉淀
2. 重要发现：在研究/开发过程中发现新知识或重要结论
3. 决策时刻：用户做出重要决策，需要记录理由
4. 错误纠正：任务失败或犯错后，记录教训
5. 用户要求：用户说"记下来"、"保存这个"、"这是个教训"

精馏流程

第1步：识别精馏价值

判断当前内容是否值得沉淀：

• 是否是重复会用到的知识？
• 是否影响未来决策？
• 是否是重要的用户偏好或约定？
• 是否是教训或纠正？

不值得精馏：一次性任务输出、临时草稿、明显过期内容

第2步：选择目标位置

内容类型	目标位置	文件命名
用户偏好、风格、约定	notes/people/ 或 USER.md	USER.md
项目进展、里程碑	notes/projects/<name>.md	项目名.md
知识积累、方法论	notes/knowledge/<topic>.md	主题名.md
教训、错误纠正	notes/lessons/YYYY-MM-DD-<lesson>.md	日期-教训
重要决策及理由	notes/decisions/YYYY-MM-DD-<decision>.md	日期-决策
人物关系	notes/people/<name>.md	人名.md
跨领域洞察	notes/ontology/	知识图谱节点

第3步：写入或更新文件

新增文件：创建 notes/<category>/YYYY-MM-DD-<title>.md

# <标题>

## 背景
[什么时候、为什么这件事重要]

## 核心内容
[提炼出的关键知识/结论/决策]

## 启示/后续
[对未来工作的指导意义]

---
来源：session_YYYY-MM-DD

更新现有文件：追加到相关文件的对应 section

第4步：触发 COLD STORE（git-notes）

重要内容写入后，执行归档：

git add <file>
git commit -m "[cold-store] <type>: <简短描述>"

精馏质量标准

好的精馏

• 一句话能说清楚核心
• 包含"为什么重要"
• 有明确的适用场景
• 标注了来源和日期

不好的精馏

• 照抄会话记录，不做提炼
• 过于笼统，没有具体细节
• 没有标注来源，难以追溯

主动精馏的触发词

用户可能不会主动说"精馏一下"，但以下表述都是精馏信号：

• "这个以后还会用到"
• "我之前好像说过这个..."
• "这次学到了..."
• "记住这个偏好..."
• "下次遇到这种情况应该..."

与 self-improving-agent 的关系

memory-pro 是精馏的"执行层"，self-improving-agent 的 multi-memory 架构（semantic + episodic + working）是技术支撑。 memory-pro 利用现有工具（read/write/exec）完成精馏动作。

ELITE-LONGTERM 前置条件

当某条 knowledge/lesson/decisions 被多次引用、或被验证为高价值洞察时，触发 elite-longterm SKILL.md，将其提升到"精英长期记忆"层。