主要命题这方面，不改模型权重，只改“怎么用模型”

Hermes Agent 的自进化能力。本质上不是模型层面的学习——不涉及微调、强化学习或参数更新——而是一次精妙的 Prompt Engineering + 文件持久化 的工程化实践。 它的主要命题是：

简单来说，Hermes 的“进化”体现在三个维度的自动积累：

三者共同构成一个闭环：执行 → 审查 → 沉淀 → 注入 → 再执行。

四维持久记忆程序

架构设计

Hermes 的记忆程序由 tools/memory_tool.py中的MemoryStore类实现。维护两个平行的存储：

# ~/.hermes/memories/
├── MEMORY.md # Agent 的个人笔记
└── USER.md # 使用者特点

两个文件使用 作为条目分隔符，每个条目是一段自然语言文本。说到例如，

User prefers concise code style without redundant comments
Project uses Python,package manager uv
This API endpoint often times out,recommend adding retry logic

冻结快照机制

This is core design decision. During .load_from_disk,we capture a frozen snapshot used for system prompt injection:

def load_from_disk:
self.memory_entries = self._read_file
self.user_entries = self._read_file
# Capture frozen snapshot for system prompt injection
self._system_prompt_snapshot = {
"memory": self._render_block,"user": self._render_block,}
}

• At session start,system prompt loads current MEMORY.md and USER.md as snapshot.
• A session sees this snapshot even if files update during runtime.
• The next session reflects updated file contents.

This protects against prefix cache invalidation and ensures model context consistency.

读写安全 & 容量管理

The write operations use atomic file replacement ensuring concurrency safety:

@staticmethod
def _write_file:
fd。tmp_path = tempfile.mkstemp,suffix=".tmp",prefix=".mem_")
with os.fdopen as f:
f.write
f.flush
os.fsync)
os.replace) # atomic on same filesystem
}
The read side needs no locks because os.replace guarantees eir old or new complete file visibility.
Safety scans filter malicious patterns . Memory files have character limits to prevent unbounded growth.

记忆管理器 

The orchestrator handles built-in and optional external providers via a unified interface: {build_system_prompt,prefetch_all,sync_all,handle_tool_call}. It enforces single external provider policy to keep consistency.

技能自动创造程序 

技能生命周期 


User writes Skill → installs into Agent → executes → ends .

A Hermes skill: Agent executes task → background review → decides wher valuable → automatically creates/updates Skill.






Skill directory supports rich attachments:
├── SKILL.md
├── references/
│ └── api-spec.md
├── templates/
│ └── output-format.json
└── scripts/
└── validate.py