← Back to Skills Marketplace
briefness

hermes agent skill

by briefness · GitHub ↗ · v1.0.3 · MIT-0
cross-platform ✓ Security Clean
165
Downloads
1
Stars
0
Active Installs
4
Versions
Install in OpenClaw
/install hermes-agent-skill
Description
突触式多智能体调度 + 主动记忆洞察 + GEPA 技能自进化
README (SKILL.md)

Hermes Agent Skill

Hermes 协议:极致的"执行效率"与"自我进化"

功能特性

  • 🧠 突触式多智能体调度:100个Agent同时协同,通信成本降到最低,毫秒级分发
  • 🧩 主动记忆与自我建模:自动提取用户偏好习惯,SQLite FTS5 全文检索快速翻旧账
  • 🧬 GEPA 技能自进化:多次执行后自动提炼技能卡,越用越聪明
  • 极致轻量:纯Python,零依赖,启动快内存占用低
  • 🔧 工程化友好:完美集成 sessions_spawn,开箱即用
  • 🔒 隐私优先:持久化默认关闭,数据存储完全可控

安装

在 OpenClaw 中:

/install-skill https://github.com/你的仓库/hermes-agent-skill

或者手动放到 ~/.openclaw/workspace/skills/ 即可。

快速开始

1. 导入

from hermes_agent import (
    hermes,                 # 核心路由器
    hermes_workflow,        # 工作流调度
    hermes_sessions,        # sessions_spawn 集成
    hermes_insight,         # 记忆洞察数据库
    insight_extractor,      # 洞察提取器
    hermes_gepa,            # GEPA 技能进化
    hermes_skill_executor,  # 带自进化的执行器
    hermes_config           # 全局配置(控制持久化开关)
)

2. 多智能体任务分发

# spawn 子智能体之后,自动注册 Hermes 订阅
hermes_sessions.on_agent_spawn(
    session_key="session-code-agent",
    agent_id="code-review-agent",
    hermes_topics=["task:code-review"]
)

# 提交任务,自动分发给所有订阅了该类型的 Agent
task_id = hermes_sessions.submit_task_to_agents(
    task_type="code-review",
    creator="user",
    session_id="main",
    payload={"pr": "https://github.com/openclaw/openclaw/pull/123"}
)

3. 主动记忆用户洞察

# 从对话提取洞察
insights = insight_extractor.extract_from_conversation(
    "我喜欢用 Python 写脚本,更快,不喜欢重型框架",
    context="对话上下文"
)

# 存储
for ins in insights:
    hermes_insight.add_insight(ins)

# 全文检索
results = hermes_insight.search_memory("Python")

4. GEPA 技能自进化

# 开始任务,自动记录
exec_id = hermes_skill_executor.start_task(
    "video-clip",
    {"input": "input.mp4", "start": 10, "end": 20}
)

# 一步一步执行,自动记录
hermes_skill_executor.step("load-video", load_video, path)
hermes_skill_executor.step("cut-segment", cut, start, end)
result = hermes_skill_executor.step("export-video", export, output)

# 完成,自动触发提炼
hermes_skill_executor.finish_task(True, result)

# 几次之后自动生成技能卡
skill = hermes_gepa.get_skill_card("video-clip")
print(f"推荐步骤: {skill.steps}")
print(f"成功率: {skill.success_rate:.1%}")

架构

hermes.py                     # 核心路由器(突触式通信)
├─ hermes_config.py           # 全局配置(持久化开关、隐私控制)
├─ hermes_openclaw.py         # 工作流调度(任务/进度/完成)
├─ hermes_sessions_integration.py  # sessions_spawn 自动集成
├─ hermes_agent_insight.py    # 主动记忆 + FTS5 全文检索
└─ hermes_skill_evolution.py  # GEPA 技能自进化

控制参数(避免 token 浪费)

GEPA 默认参数:

  • min_success_samples = 2 - 最少 2 次成功才提炼
  • min_new_executions = 3 - 已有技能后新增 3 次才重新提炼
  • max_refines_per_task = 10 - 单个任务最多提炼 10 次
  • min_improvement = 0.05 - 成功率变化 \x3C 5% 不提炼

自定义:

from hermes_skill_evolution import GEPASkillEvolution
my_gepa = GEPASkillEvolution(
    min_success_samples=5,
    max_refines_per_task=5
)

数据存储

默认关闭,需显式开启。

# 方式一:环境变量
export HERMES_PERSISTENCE_ENABLED=true

# 方式二:运行时代码控制
from hermes_agent import hermes_config
hermes_config.set_persistence(True)

开启后数据存储位置:

  • ~/.hermes/insights.db - 洞察和记忆(SQLite FTS5)
  • ~/.hermes/skills.db - 技能卡和执行记录

首次运行自动创建,不需要手动初始化。

隐私控制参数

环境变量 默认值 说明
HERMES_PERSISTENCE_ENABLED false 是否启用持久化(默认关闭)
HERMES_INSIGHT_EXTRACTION_ENABLED true 是否提取对话洞察
HERMES_SENSITIVE_FILTER_ENABLED true 是否自动过滤敏感信息
HERMES_SESSION_LOG_LEVEL summary fallback 日志级别:off/summary/full
HERMES_INSIGHTS_DB ~/.hermes/insights.db 洞察 DB 路径
HERMES_SKILLS_DB ~/.hermes/skills.db 技能 DB 路径

敏感信息过滤:自动过滤 API Key、密码、Token、私钥、证书、邮箱等。

依赖

  • Python 3.8+
  • 不需要第三方包,SQLite 内置

性能

测试数据(100 Agent,1000 条发布,49500 次投递):

  • 平均单消息处理:14 微秒 → 真·毫秒级分发

License

MIT

Usage Guidance
This skill appears to be what it claims: a pure-Python multi-agent router, local insight DB (SQLite/FTS5), and a local skill-evolution component. Key things to check before installing: 1) Confirm the skill source (SKILL.md mentions a GitHub URL placeholder). 2) Persistence is off by default — only enable HERMES_PERSISTENCE_ENABLED if you accept local disk storage at ~/.hermes/*.db. 3) The sessions integration calls sessions_send to deliver fallback records; check what sessions_send does in your OpenClaw environment (where it sends logs) and set HERMES_SESSION_LOG_LEVEL to 'summary' or 'off' if you want to avoid sending payloads. 4) Keep HERMES_SENSITIVE_FILTER_ENABLED=true if you want automatic filtering of tokens/emails/etc. If you need higher assurance, review the sessions_send implementation used by your platform and inspect the on-disk DB files after enabling persistence.
Capability Analysis
Type: OpenClaw Skill Name: hermes-agent-skill Version: 1.0.3 The Hermes Agent skill bundle is a well-structured framework for multi-agent coordination and memory management. It implements a pub/sub router for agent communication (hermes.py) and a memory system that extracts user preferences into a local SQLite database (hermes_agent_insight.py). Notably, the skill includes proactive privacy measures, such as a sensitive information filter that masks API keys and tokens using regex, and it defaults to disabling disk persistence (hermes_config.py). No evidence of data exfiltration, malicious execution, or unauthorized persistence was found.
Capability Tags
cryptorequires-walletrequires-sensitive-credentials
Capability Assessment
Purpose & Capability
Name/description (multi-agent scheduling, memory, GEPA evolution) match the provided Python modules: router, workflow scheduler, sessions integration, insight DB, and evolution engine. Declared zero required binaries/credentials is proportionate because the implementation uses only the Python stdlib and SQLite.
Instruction Scope
SKILL.md and code keep persistence disabled by default and require an explicit env var or API call to enable disk writes. The sessions integration forwards lightweight records to sessions_send; by default log level 'summary' avoids sending payloads, but if configured to 'full' the handler will include payloads. Users should verify what sessions_send does in their environment before enabling full logging or persistence.
Install Mechanism
No install spec is included in registry metadata; the package is delivered as code files in the skill bundle (no network download at install time). SKILL.md suggests installing from a GitHub repo but that's informational. No URL-based downloads or opaque installers are part of the skill bundle.
Credentials
The skill does not require secrets or third‑party credentials. It documents optional HERMES_* environment variables that control local persistence, DB paths, extraction and filtering behavior — all relevant to the skill's functionality. No unrelated env vars or credentials are requested.
Persistence & Privilege
Persistence writes are local (~/.hermes/*.db) and are explicitly disabled by default; enabling requires setting HERMES_PERSISTENCE_ENABLED=true or an explicit API call. The skill does not request permanent platform-wide privileges (always is false) and does not modify other skills' configs.
How to Use
  1. Make sure OpenClaw is installed (local or Docker)
  2. Run the install command in chat: /install hermes-agent-skill
  3. After installation, invoke the skill by name or use /hermes-agent-skill
  4. Provide required inputs per the skill's parameter spec and get structured output
Version History
v1.0.3
- Introduced a new `hermes_config.py` module for global configuration and privacy controls. - Added support for fully disabling persistence by default; data storage is now opt-in and user-controllable. - Enhanced privacy: defaults, docs, and structure updated to clarify sensitive data handling and storage options. - Updated documentation to include new architecture file, usage, and environment variable configuration examples. - Minor version bump to 1.0.3.
v1.0.2
- Bump version from 1.0.1 to 1.0.2 in documentation. - No code or feature changes; documentation version updated for consistency.
v1.0.1
- Bumped version to 1.0.1. - No other changes; documentation and functionality remain the same.
v1.0.0
Initial release of Hermes Agent skill—multi-agent scheduling, active memory, and self-evolving task skills. - Introduces synapse-style multi-agent scheduling for ultra-fast, low-cost communication. - Adds active memory extraction and SQLite FTS5-based instant search. - Implements GEPA-based self-evolving skills that improve with repeated use. - Provides lightweight, zero-dependency design with rapid startup and low memory usage. - Seamlessly integrates with OpenClaw sessions_spawn for out-of-the-box workflows.
Metadata
Slug hermes-agent-skill
Version 1.0.3
License MIT-0
All-time Installs 0
Active Installs 0
Total Versions 4
Frequently Asked Questions

What is hermes agent skill?

突触式多智能体调度 + 主动记忆洞察 + GEPA 技能自进化. It is an AI Agent Skill for Claude Code / OpenClaw, with 165 downloads so far.

How do I install hermes agent skill?

Run "/install hermes-agent-skill" in the OpenClaw or Claude Code chat to install it in one step — no extra setup required.

Is hermes agent skill free?

Yes, hermes agent skill is completely free, licensed under MIT-0. You can download, install and use it at no cost.

Which platforms does hermes agent skill support?

hermes agent skill is cross-platform and runs anywhere OpenClaw / Claude Code is available (cross-platform).

Who created hermes agent skill?

It is built and maintained by briefness (@briefness); the current version is v1.0.3.

💬 Comments