hermes-agent 平替 openclaw 自进化技能

🤖 AI 自进化技能包 — 自动从复杂任务创建可复用技能，HOT/WARM/COLD 分层记忆管理，智能 Nudge 提醒与定时自检，持续让 AI 越用越聪明。追踪 NousResearch/Hermes-Agent 上游演进。

Security Scan

The package implements what it claims: a local self‑evolution toolset (memory, nudges, auto‑skill creation, upstream GitHub tracking) that reads/writes files under the user's home and fetches public GitHub content; nothing in the code covertly exfiltrates secrets or demands unrelated credentials, but it will create and modify files in your home directory and encourages pushing changes to GitHub so review before use.

Purpose & Capability

Name/description (self‑evolving skill inspired by Hermes) align with provided code: memory management, nudge system, auto skill creation, skill evaluation and an upstream GitHub tracker. The modules and CLI entrypoint implement the documented features; required resources (none declared) are coherent with the stated purpose.

Instruction Scope

Runtime instructions and scripts operate on local directories (~/self-improving, ~/.qclaw/skills/hermes-skill). The upstream tracker calls the public GitHub API and raw.githubusercontent.com to fetch upstream files (read-only). The SKILL.md and README encourage git push of local improvements — that is a user action but could cause accidental upload of locally stored memory or metadata if followed. The scripts do not appear to transmit local files to external endpoints, but they do read and write potentially sensitive local data (memory.md, task_history.json, skills directory).

Install Mechanism

There is no automatic install spec; this is instruction + code included in the skill bundle. No downloads from untrusted URLs or archive extraction are performed by the skill itself. The only network activity in the code is benign reads from the GitHub API/raw content and standard HTTP fetches.

Credentials

The skill declares no required environment variables or credentials. Code uses unauthenticated GitHub API calls (subject to rate limiting) and accesses only local file paths. The developer guidelines suggesting git push are the only place where credentials could become relevant — pushing is a manual user action and not automated by the scripts.

Persistence & Privilege

always:false and autonomous invocation are default. The skill writes state and content under ~/self-improving and into ~/.qclaw/skills/hermes-skill (creates auto-generated skills, state.json, changelog, logs, etc.). This is expected for a memory/skill generator, but it is persistent and modifies user home; review file contents and permissions before trusting with sensitive information. The skill does not modify other skills' configurations beyond its own path.

Assessment

This package appears to do what it says, but take these precautions before installing or running it: - Review and sandbox: run inside a disposable account/container first. The scripts create and modify files under ~/self-improving and ~/.qclaw/skills/hermes-skill. - Sensitive data: do not point the skill at or populate its memory files with secrets (passwords, API keys, private data). Memory files (memory.md, task_history.json, skill meta) may be written to disk and could be accidentally pushed if you follow the README's git guidance. - Git push caution: the README/SKILL.md strongly recommends pushing every evolution to GitHub. Only push after auditing what changes will be committed to avoid leaking local data. - Network behavior: the upstream tracker fetches public GitHub API/raw content (read-only). It does not appear to send your local data to external endpoints, but review the code for any local modifications you are uncomfortable with. - Permissions and cleanup: limit filesystem permissions for the created directories if needed, and remove or back up the ~/self-improving folder if you uninstall the skill. If you want higher assurance, provide the file list to a developer you trust or run the scripts step‑by‑step to inspect generated files before enabling any automated/cron behavior.

HermesSkill — AI 自进化技能包

让 AI Agent 拥有持续进化的能力——像 Hermes Agent 一样，从经验中学习、自我改进。

核心模块

模块	脚本	说明
记忆管理	memory_tool.py	HOT/WARM/COLD 三层记忆、搜索、增删
Nudge 提醒	nudge_system.py	定时自检、智能提醒、知识保存触发
技能创建	auto_skill_creator.py	复杂任务自动生成可复用 skill
技能评估	skill_evaluator.py	评分反馈、自动改进、usage 追踪
自检	self_check_cron.sh	cron 定时健康检查
上游追踪	upstream_tracker.py	追踪 Hermes Agent 最新演进

上游追踪系统

追踪目标

NousResearch/hermes-agent — GitHub 56k+ Stars，自进化 AI Agent 标杆

追踪文件映射

上游路径	本地对应
agent/memory_manager.py	memory_tool.py
agent/memory_provider.py	memory_tool.py
agent/skill_commands.py	auto_skill_creator.py
agent/skill_utils.py	auto_skill_creator.py
agent/prompt_builder.py	nudge_system.py

使用方式

# 检查上游新提交并分析融合价值
python3 scripts/evo_main.py upstream check

# 查看追踪状态
python3 scripts/evo_main.py upstream status

# 显示具体文件对比
python3 scripts/evo_main.py upstream diff agent/memory_manager.py

# 查看融合推荐
python3 scripts/evo_main.py upstream fuse

# 列出追踪的核心文件
python3 scripts/evo_main.py upstream list

工作原理

1. Fetch — 从 GitHub API 获取 nousresearch/hermes-agent/main 最新 30 条 commit
2. Detect — 找出涉及核心模块的变更（memory、skill、prompt 等）
3. Analyze — 提取新增的类/方法/模式，对比本地实现，计算融合价值
4. Record — 高价值 diff 存入 upstream-tracker/state.json + 追加到 changelog.md
5. Notify — 每日 09:00 自动检查，有高价值发现时推送摘要

融合价值评估

分数	含义
7-10/10	上游重大新能力，值得深度融合
3-6/10	有借鉴价值，选择性融合
1-2/10	边缘改进，记录即可

存储位置

• 追踪状态: ~/self-improving/upstream-tracker/state.json
• 变更记录: ~/self-improving/upstream-tracker/changelog.md

自进化统一入口

python3 scripts/evo_main.py nudge [type]          # 检查/触发 nudge
python3 scripts/evo_main.py remember <text>       # 添加记忆
python3 scripts/evo_main.py forget <text>         # 删除记忆
python3 scripts/evo_main.py search <query>        # 搜索记忆
python3 scripts/evo_main.py stats                 # 记忆统计
python3 scripts/evo_main.py skills                # 列出 auto skills
python3 scripts/evo_main.py eval <skill,rating>   # 评估 skill
python3 scripts/evo_main.py improve <skill,notes>  # 改进 skill
python3 scripts/evo_main.py upstream <cmd>        # 上游追踪 (check|status|diff|list|fuse)
python3 scripts/evo_main.py check                 # 完整自检

定时任务

• 每小时: Nudge 自检（检查知识保存、记忆整理）
• 每天 09:00: Hermes Agent 上游追踪检查

版本历史

• v1.0.0: 初始版本（HOT/WARM/COLD + nudge + auto-skill）
• v1.1.0: 新增上游追踪系统，对接 Hermes Agent 主线演进

开发规范（重要）

本技能包本身是自进化系统，任何模块有改进时必须遵守以下规则：

GitHub 同步规则

每次自身技能进化必须同步到 GitHub 仓库

改进流程：

1. 在本地 ~/.qclaw/skills/hermes-skill/ 改进代码
2. git add + git commit -m "描述"
3. git push 推送到 origin/main
4. 同步更新 SKILL.md / README.md 说明（如果功能有变化）

Changelog 追加规则

新增模块、重大重构或 API 变更 → 在 upstream-tracker/changelog.md 底部追加条目

格式：

## YYYY-MM-DD — vX.X.X
- [改进描述]
- [文件变更]

避免事项

• ❌ 只在本地改，不 push
• ❌ 更新代码但不更新文档
• ❌ 破坏已有的 upstream-tracker 状态文件