AI Agent 的記憶困境
用過 AI Agent 的人大概都有這種體驗：每次開新對話，就像對方失憶了一樣，你得重新解釋你是誰、你的環境是什麼、你上次踩過哪些坑。這不是模型不夠聰明，而是架構設計使然——大多數 Agent 框架本質上是無狀態的，會話結束，記憶歸零。
Hermes Agent 想解決的，就是這件事。
什麼是閉環學習
「閉環學習」這個詞聽起來抽象，但概念很直觀：Agent 完成一件任務之後，不只是把結果交給你，還會把這次的操作過程提煉成結構化的技能文件，存起來備用。下次遇到類似的問題，它不需要從頭推理，直接調用已有的技能。
根據 Hermes 官方基準測試，使用自創技能的 Agent 完成研究任務的速度比全新實例快 40%。這個數字值得認真對待，因為它意味著 Agent 的能力不是固定的——它會隨著使用時間累積而增長。
這讓我想到一個類比：剛入職的工程師和做了五年的工程師，面對同一個問題的處理速度差距不是因為智商，而是因為後者腦子裡有一個隨時可以調用的「經驗庫」。Hermes 想為 AI Agent 建立的，就是這個東西。
三層記憶架構
Hermes Agent 的記憶系統分三層，各司其職：
- MEMORY.md：儲存環境資訊與長期經驗摘要，跨會話持久存在
- USER.md：記錄使用者的偏好、溝通風格、習慣指令——它在幫你建立「用戶畫像」
- SQLite FTS5 全文搜尋資料庫：保留數週的對話紀錄，可以被精確檢索
第三層特別值得一提。FTS5（Full-Text Search）是 SQLite 內建的全文搜尋引擎，這代表 Hermes 不是把歷史對話當作一個大型純文字檔案塞進 context，而是建立了一個可以快速查詢的結構化資料庫。這在 token 消耗上是個關鍵優化——它用漸進披露（progressive disclosure）的方式，只把當前任務真正需要的記憶片段提取出來，而不是每次都把所有歷史記憶一股腦丟給模型。
此外，Hermes 透過 Honcho 進行用戶建模，這是一個專門為 AI 應用設計的用戶個性化框架。換句話說，它不只記得你做了什麼，還在推斷你是什麼樣的人。
技能自動生成：最關鍵的差異點
技能文件是閉環學習的核心輸出物。每當 Agent 完成一個複雜任務，它會自動產生一份結構化的技能文件，包含：
- 操作步驟的摘要
- 遇到的邊界條件與處理方式
- 驗證方法（怎麼確認任務成功）
這些技能文件本質上是 Agent 對自己工作方式的文檔化。下次遇到類似任務，它會先搜尋技能庫，找到匹配的技能後直接套用，跳過重新推理的成本。
這個設計有個微妙之處值得注意：技能具有領域特定性。幫你分析 nginx log 學到的技能，不能直接遷移到幫你寫 Python 腳本。這不是缺點，而是現實——人類專家的直覺也是領域專屬的。重點是，在你最頻繁使用的那個領域裡，它會越來越順手。
2026 年的發展速度
Hermes Agent 在兩個月內從 v0.1.0 發展到 v0.8.0，累積超過 64,200 個 GitHub Stars，並在 GitHub Stars 排名上超越了更早期的 OpenClaw。v0.8.0（4 月 8 日）合併了 209 個提案，新增 Browser Use 整合、遠端後端支援與工作樹並行化。框架還與 MiniMax 建立合作關係，其 M2.7 模型可直接使用無需額外配置。
這個迭代速度說明了兩件事：社群的真實需求在推動它，以及 API 穩定性目前仍無保證。如果你要在生產環境使用，版本鎖定是必要的。
與 OpenClaw 的比較
兩者常被拿來比較，因為定位接近，但設計哲學不同：
- Hermes：深度記憶 + 自我進化，安全性為優先（零 CVE 紀錄）
- OpenClaw：50+ 平台整合，生態廣，但存在 CVSS 8.8 的安全漏洞
OpenClaw 的廣度優勢是真實的，但 CVSS 8.8 是個不能忽視的數字——這個分數意味著漏洞可以在無需用戶互動的情況下被遠端利用。如果你的 Agent 在處理敏感資料或連接內網服務，這個風險需要認真評估。Hermes 提供了內建的遷移工具，讓從 OpenClaw 切換過來相對容易。
成本與部署現實
MIT 授權，完全免費。一行命令安裝：
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
hermes setup
hermes實際運行成本主要來自 API 呼叫。Hermes 支援 200+ 種模型，包括 Anthropic Claude、OpenAI、Google Gemini，也可以搭配本地 Ollama 實現零 API 成本運行。若掛在 VPS 上作為常駐服務，$5-10 美元的主機費加上 $15-80 美元的 API 費用是比較典型的月費範圍。
平台接入方面支援 Telegram、Discord、Slack、WhatsApp、Signal、CLI，以及 Home Assistant——這代表同一個 Agent 實例可以同時服務你的手機、電腦和智慧家居，記憶是共享的。
我的觀點：技能累積才是護城河
很多人評估 AI Agent 框架時，看的是支援多少工具、接多少平台。這些固然重要，但我認為 Hermes 最值得關注的地方在別處：它在幫你建立一個隨時間增值的個人知識庫。
工具可以複製，平台整合可以複製，但你這幾個月用 Hermes 累積下來的技能文件、用戶模型、歷史記憶——這些是你自己的，沒辦法被別的框架直接繼承。這才是它真正的護城河，也是「越用越聰明」這個說法的底層邏輯。
當然，它現在還在快速迭代期，有些粗糙是正常的。但如果你正在尋找一個值得長期投入的 Agent 框架，Hermes 的方向是對的。
你有在用 Hermes Agent 或其他 Agent 框架嗎？歡迎留言分享你的使用心得。
Hermes Agent’s Closed-Loop Learning: When AI Finally Starts Remembering You
The Memory Problem with AI Agents
Anyone who has used an AI Agent long enough will recognize this frustration: every new session, it’s amnesia again. You re-explain your environment, your preferences, the pitfalls you already discovered. The model isn’t unintelligent — it’s stateless by design. Most Agent frameworks treat each conversation as if nothing came before it.
That is the problem Hermes Agent is built to solve.
What Closed-Loop Learning Actually Means
The term sounds abstract, but the idea is concrete: when Hermes completes a complex task, it doesn’t just hand you the result and forget. It distills the operation into a structured skill file and stores it for future use. The next time a similar problem appears, it retrieves the relevant skill instead of reasoning from scratch.
According to official benchmarks, agents using self-generated skills complete research tasks 40% faster than fresh instances. That number matters because it means the agent’s capability is not fixed — it compounds with use.
Think of it this way: the gap between a new hire and a five-year veteran isn’t raw intelligence, it’s an accumulated library of heuristics, patterns, and hard-won lessons that can be recalled instantly. Hermes is trying to build that for AI Agents.
Three-Layer Memory Architecture
Hermes structures memory across three layers:
- MEMORY.md: Environment details and long-term experience summaries, persistent across sessions
- USER.md: User preferences, communication style, habitual instructions — effectively a behavioral profile
- SQLite FTS5 full-text search database: Weeks of conversation history, queryable with precision
The third layer deserves attention. FTS5 means Hermes doesn’t stuff the entire conversation history into the context window as one massive text blob. Instead, it uses progressive disclosure — pulling only the memory fragments actually relevant to the current task. This is a meaningful token efficiency optimization, not just a feature checkbox.
User modeling is handled through Honcho, a framework designed for AI personalization. The implication: it’s not just recording what you do, it’s building an inference about who you are.
Automatic Skill Generation: The Key Differentiator
Skill files are the core output of closed-loop learning. After completing a complex task, Hermes automatically produces a structured document containing:
- A summary of the steps taken
- Edge cases encountered and how they were handled
- Verification methods to confirm success
These files are essentially the agent documenting its own working methods. On subsequent similar tasks, it searches the skill library first, applies the matching skill, and skips the full reasoning cycle.
One nuance worth understanding: skills are domain-specific. A skill developed for analyzing nginx logs won’t transfer cleanly to writing Python scripts. This isn’t a flaw — human expertise works the same way. The payoff arrives in the domains you use most frequently, where the compounding effect becomes genuinely noticeable.
Development Velocity in 2026
Hermes moved from v0.1.0 to v0.8.0 in two months, accumulating over 64,200 GitHub Stars and surpassing OpenClaw in the process. The v0.8.0 release merged 209 proposals, adding Browser Use integration, remote backend support, and worktree parallelization. A partnership with MiniMax means the M2.7 model works out of the box without additional configuration.
This velocity signals genuine user demand — and genuine API instability risk. Version pinning is advisable before any production deployment.
Hermes vs. OpenClaw
These two frameworks are frequently compared because their positioning overlaps, but their design philosophies diverge:
- Hermes: Deep memory, self-improvement, security-first (zero CVE record)
- OpenClaw: 50+ platform integrations, broader ecosystem, but carries a CVSS 8.8 vulnerability
OpenClaw’s breadth advantage is real. But CVSS 8.8 means the vulnerability can be exploited remotely without user interaction. If your Agent handles sensitive data or connects to internal infrastructure, that risk demands serious evaluation. Hermes ships a migration tool to ease the transition from OpenClaw.
Cost and Deployment
MIT license, free to use. One-line installation:
curl -fsSL https://raw.githubusercontent.com/NousResearch/hermes-agent/main/scripts/install.sh | bash
hermes setup
hermesRunning costs come from API calls. Hermes supports 200+ models including Claude, OpenAI, Gemini, and local Ollama for zero API cost. Hosted on a VPS, expect $5–10/month for compute and $15–80/month for API usage depending on workload. Platform support covers Telegram, Discord, Slack, WhatsApp, Signal, CLI, and Home Assistant — a single agent instance serving all channels with shared memory.
The Real Moat Is Accumulated Skills
Most people evaluate Agent frameworks by tool count or platform integrations. Both matter, but I think Hermes’s most compelling value is elsewhere: it’s building a personal knowledge base that appreciates in value over time.
Tools can be replicated. Platform integrations can be replicated. But the skill files, user model, and accumulated memory you build over months of use — those are yours, and no other framework can inherit them directly. That’s the actual moat, and the real mechanism behind “gets smarter the more you use it.”
The framework is still rough in places. The rapid iteration is a feature, not just a risk. If you’re looking for an Agent framework worth a genuine long-term investment, Hermes is pointing in the right direction.
Are you running Hermes Agent or another Agent framework? Share your setup in the comments.
發佈留言