# Hermes Just Built Garbage Collection for AI Agent Skills
**作者**: AlphaSignal AI
**日期**: 2026-04-30T20:58:00.000Z
**来源**: [https://x.com/AlphaSignalAI/status/2050269010735018074](https://x.com/AlphaSignalAI/status/2050269010735018074)
---
## v0.12.0's Curator runs a 7-day background pass that keeps skill libraries from bloating
Hermes Agent now runs garbage collection on its own skills.
The v0.12.0 release, published April 30, 2026, ships an autonomous Curator that wakes every 7 days, grades the skill library, consolidates duplicates, and archives the dead.
It closes the missing step in every self-improving agent: a way to retire what was learned.
No other open-source agent framework ships the full create-use-retire loop today.
> **Nous Research@NousResearch**: [原文链接](https://x.com/NousResearch/status/2049956455982182593)
>
> Hermes Agent v0.12.0 - “The Curator Release”
>
> 
## Context
The Curator was published by @Teknium under Nous Research, release tag v2026.4.30. Hermes Agent is open-source MIT and sits at 128,000+ GitHub stars. The release covers 1,096 commits, 550 merged PRs, and 213 community contributors since v0.11.0.
The original proposal landed as Issue #7816 on April 11, 2026. Nineteen days later it shipped.
Hermes already had a self-improvement loop that creates skills from successful workflows and stores them as SKILL.md files in ~/.hermes/skills/. The Curator is the first step that lets the agent retire what it created.
## The problem it solves
A self-improving agent that creates skills monotonically eventually drowns in them. After weeks of use, the catalog accumulates narrow near-duplicates: small variations of the same research pipeline, build script, or data-cleanup flow. Every conversation pays a token tax to enumerate what's available, and the agent's planner has more options to confuse it.
Before v0.12.0 there was no built-in way to retire a skill. Users wrote pruning scripts or deleted by hand. Issue #7816 had been open since April 11 asking for lifecycle management as first-class infrastructure.
## How it works
The Curator runs as a forked AIAgent process triggered by inactivity, not a system cron. Both must be true: 168 hours since the last run and 2 hours of agent idle time. The fork runs in its own prompt cache and never touches the active conversation.
A run has two phases that map cleanly onto garbage collection.
Phase 1: deterministic transitions
The first phase runs no LLM. It reads ~/.hermes/skills/.usage.json, a sidecar file that tracks use_count, view_count, patch_count, and timestamps for every skill. Skills unused for 30 days transition to stale. Skills unused for 90 days move to ~/.hermes/skills/.archive/. Skills used again return to active.
Phase 2: LLM consolidation pass
The second phase spawns the auxiliary review agent with max_iterations=8. The fork can read any skill via skill_view, propose patches via skill_manage, and archive via the terminal tool. The system prompt frames the pass as an "umbrella-building consolidation" rather than a passive audit, asking the model: would a maintainer write this as N separate skills, or one skill with N labeled subsections?
Three consolidation strategies are available: merge into an existing umbrella skill, create a new umbrella SKILL.md, or demote redundant content to references/, templates/, or scripts/ support files. The auxiliary model is configured under auxiliary.curator in config.yaml and can be a cheaper model than the primary, since the maintenance pass is bounded.
## Defense-in-depth
Bundled skills are protected. Skills shipped with the Hermes repo are listed in ~/.hermes/skills/.bundled_manifest and excluded from every Curator action.
Hub-installed skills are protected. Skills installed from agentskills.io are listed in ~/.hermes/skills/.hub/lock.json and also excluded.
Pinning is a hard fence. hermes curator pin <name> flags a skill as immutable. The auto-transitions skip it, the LLM review is told to leave it alone, and skill_manage refuses every write action on it. Any agent attempt to edit a pinned skill returns a refusal asking the user to unpin first.
The review fork has scoped toolsets. The forked agent gets memory and skills tools only. No shell, no web, no filesystem access outside the skills directory.
Nothing is auto-deleted. The worst outcome is archival into .archive/, recoverable with hermes curator restore <name>.
## How it compares
No other agent framework ships built-in skill lifecycle management today. The Curator is a category-first feature, not an incremental improvement.
## How to use it today
Update to v0.12.0:
```
hermes update
hermes curator status
```
hermes curator status shows the last run time, counts of active/stale/archived skills, the pinned list, and the five least-recently-used skills (the next archival candidates).
Pin any mission-critical custom skills before the first auto-run fires:
```
hermes curator pin my-deploy-skill
```
Trigger a review on demand and inspect the report:
```
hermes curator run --sync
cat ~/.hermes/logs/curator/<timestamp>/REPORT.md
```
To use a cheaper auxiliary model, run hermes model, pick "Auxiliary models", and route the Curator to a smaller provider. Each run writes both run.json (machine-readable) and REPORT.md (human-readable), so audit history stays local.
## Current Limitations
> Current Limitations until May 1st.
No dry-run mode. The first auto-run after hermes update can archive custom skills without preview. GitHub issue #18373 confirms users have already been bitten by this. Pin first, then update.
bump_use() was broken at launch. PR #17782 noted that the use-counter had zero call sites at release time. The fix landed in PR #17932. Update past the patch before trusting hermes curator status output.
Edit activity is not a lifecycle signal. Transitions look at use_count and last_used_at only. A skill that was extensively patched but rarely invoked can still be archived. Issue #17952 closed the immediate concern, but patch_count does not yet drive transitions.
Review quality tracks the auxiliary model. The umbrella-building pass is only as good as the model behind it. Misclassification of niche skills as redundant is possible. Archival, not deletion, is the safety net.
No performance scoring. Lifecycle is purely time-based. A skill that ran rarely but always succeeded has no way to advertise its quality.
So the best recommendation is to update, pin first, then enable. The defaults are conservative and the worst outcome is a recoverable archive, but the lack of a dry-run on first run is the friction users will hit before everything else.
## Who benefits and who doesn't
The Curator pays off for Hermes users running long-lived deployments where skills accumulate over weeks, framework engineers building self-improving agents who want a reference architecture for safe self-modification, and teams evaluating agent frameworks for production where library hygiene affects reliability at scale.
It does not help short-lived agent runs that never accumulate skills, or teams locked into Claude Code or Codex where no equivalent feature exists. Users with mission-critical custom skills who skip the pre-update pinning step are also at risk on the first auto-run.
## Practitioner Implication
Long-running self-improving agents are deployable now that skill libraries don't grow without bound.
## Links
- Hermes Agent repo (~5 min skim)
- Curator documentation (~10 min read)
- v2026.4.30 release notes (~15 min read)
- Issue #7816, original proposal
Follow @AlphaSignalAI for more content like this.
Subscribe at AlphaSignal.ai for daily AI signals. Read by 280,000+ developers.
## Questions?
Q: What is Hermes Curator and how often does it run? A: It is an autonomous background maintenance pass for agent-created skills, shipped in Hermes Agent v0.12.0. By default it runs every 168 hours (7 days), and only after the agent has been idle for 2 hours. Both conditions must be true.
Q: Can Hermes Curator delete my skills? A: No. The worst outcome is archival to ~/.hermes/skills/.archive/, recoverable with hermes curator restore <name>. The Curator never auto-deletes.
Q: How do I prevent Hermes Curator from archiving a custom skill? A: Run hermes curator pin <skill-name>. Pinning is a hard fence: auto-transitions skip pinned skills, the LLM review is instructed to leave them alone, and skill_manage refuses every write action on them.
Q: Does the Curator touch skills installed from agentskills.io or shipped with Hermes? A: No. Bundled skills are protected via ~/.hermes/skills/.bundled_manifest and hub-installed skills via ~/.hermes/skills/.hub/lock.json. Only skills the agent created or the user wrote by hand are subject to Curator action.
## 相关链接
- [AlphaSignal AI](https://x.com/AlphaSignalAI)
- [@AlphaSignalAI](https://x.com/AlphaSignalAI)
- [4.8K](https://x.com/AlphaSignalAI/status/2050269010735018074/analytics)
- [May 1](https://x.com/NousResearch/status/2049956455982182593)
- [471K](https://x.com/NousResearch/status/2049956455982182593/analytics)
- [published](https://x.com/Teknium/status/2049717907664581067?s=20)
- [@Teknium](https://x.com/@Teknium)
- [agentskills.io](https://agentskills.io/)
- [#18373](https://github.com/NousResearch/hermes-agent/issues/18373)
- [#17782](https://github.com/NousResearch/hermes-agent/issues/17782)
- [#17932](https://github.com/NousResearch/hermes-agent/pull/17932)
- [#17952](https://github.com/NousResearch/hermes-agent/issues/17952)
- [Hermes Agent repo](https://github.com/NousResearch/hermes-agent)
- [Curator documentation](https://hermes-agent.nousresearch.com/docs/user-guide/features/curator)
- [v2026.4.30 release notes](https://github.com/NousResearch/hermes-agent/releases/tag/v2026.4.30)
- [Issue #7816, original proposal](https://github.com/NousResearch/hermes-agent/issues/7816)
- [@AlphaSignalAI](https://x.com/@AlphaSignalAI)
- [AlphaSignal.ai](https://alphasignal.ai/)
- [agentskills.io](https://agentskills.io/)
- [1:39 AM · May 2, 2026](https://x.com/AlphaSignalAI/status/2050269010735018074)
- [4,890 Views](https://x.com/AlphaSignalAI/status/2050269010735018074/analytics)
---
*导出时间: 2026/5/2 10:09:13*
---
## 中文翻译
# Hermes 刚刚为 AI Agent 技能构建了垃圾回收机制
**作者**: AlphaSignal AI
**日期**: 2026-04-30T20:58:00.000Z
**来源**: [https://x.com/AlphaSignalAI/status/2050269010735018074](https://x.com/AlphaSignalAI/status/2050269010735018074)
---
## v0.12.0 的 Curator 运行为期 7 天的后台清理，防止技能库膨胀
Hermes Agent 现在对其自身的技能运行垃圾回收。
2026 年 4 月 30 日发布的 v0.12.0 版本搭载了一个自主的 Curator（管理者），它会每 7 天唤醒一次，对技能库进行分级，合并重复项，并归档死代码。
它填补了每一个自我改进 Agent 中缺失的一环：一种淘汰已学内容的方法。
目前，没有其他开源 Agent 框架提供完整的“创建-使用-淘汰”闭环。
> **Nous Research@NousResearch**: [原文链接](https://x.com/NousResearch/status/2049956455982182593)
>
> Hermes Agent v0.12.0 - “The Curator Release”
>
> 
## 背景
Curator 由 @Teknium 在 Nous Research 下发布，发布标签为 v2026.4.30。Hermes Agent 是开源 MIT 协议项目，拥有 128,000+ GitHub 星标。该版本涵盖自 v0.11.0 以来的 1,096 次提交、550 个合并的 PR 以及 213 位社区贡献者。
最初的提议于 2026 年 4 月 11 日作为 Issue #7816 提出。十九天后，它便正式发布。
Hermes 已经具备一个自我改进循环，可以从成功的工作流中创建技能，并将它们作为 SKILL.md 文件存储在 ~/.hermes/skills/ 中。Curator 是让 Agent 能够淘汰其所创建内容的第一步。
## 解决的问题
单调地创建技能的自我改进 Agent 最终会淹没在技能之中。经过数周的使用，目录中会积累狭窄的近似重复项：同一研究流程、构建脚本或数据清理流程的微小变体。每次对话都要支付 token 税来枚举可用内容，且 Agent 的规划器有更多选项会导致其混淆。
在 v0.12.0 之前，没有内置的方法可以淘汰技能。用户只能编写修剪脚本或手动删除。自 4 月 11 日以来，Issue #7816 一直处于开放状态，请求将生命周期管理作为一等（first-class）基础设施。
## 工作原理
Curator 作为由非活动状态触发的分叉 AIAgent 进程运行，而不是系统 cron。必须同时满足两个条件：距离上次运行 168 小时以及 Agent 空闲 2 小时。分叉进程运行在其自己的提示缓存中，绝不会触及活动的对话。
一次运行分为两个阶段，清晰地映射到垃圾回收过程。
第一阶段：确定性转换
第一阶段不运行 LLM。它读取 ~/.hermes/skills/.usage.json，这是一个跟踪每个技能的 use_count（使用次数）、view_count（查看次数）、patch_count（修补次数）和时间戳的辅助文件。30 天未使用的技能转换为陈旧状态。90 天未使用的技能移至 ~/.hermes/skills/.archive/。再次被使用的技能恢复为活动状态。
第二阶段：LLM 合并传递
第二阶段生成辅助审查 Agent，设置 max_iterations=8。分叉进程可以通过 skill_view 读取任何技能，通过 skill_manage 提出修补建议，并通过终端工具进行归档。系统提示将此过程构建为“构建伞式合并”而非被动审计，询问模型：维护者会将其编写为 N 个单独的技能，还是一个带有 N 个标记小节的单个技能？
有三种合并策略可用：合并到现有的伞式技能中，创建新的 SKILL.md 伞式文件，或将冗余内容降级为 references/、templates/ 或 scripts/ 支持文件。辅助模型在 config.yaml 中的 auxiliary.curator 下配置，可以比主模型更便宜，因为维护传递是有界的。
## 深度防御
捆绑技能受保护。随 Hermes 仓库发布的技能列在 ~/.hermes/skills/.bundled_manifest 中，并被排除在所有 Curator 操作之外。
从 Hub 安装的技能受保护。从 agentskills.io 安装的技能列在 ~/.hermes/skills/.hub/lock.json 中，同样被排除在外。
固定 是一道硬性屏障。hermes curator pin <name> 将技能标记为不可变。自动转换会跳过它，LLM 审查被告知将其忽略，且 skill_manage 会拒绝对其执行的任何写入操作。任何 Agent 尝试编辑已固定的技能都会返回拒绝，并要求用户先取消固定。
审查分叉具有作用域工具集。分叉的 Agent 仅获得内存和技能工具。没有 Shell，没有 Web，无法访问技能目录之外的文件系统。
没有任何内容会被自动删除。最坏的结果是归档到 .archive/，可以使用 hermes curator restore <name> 恢复。
## 对比
目前没有其他 Agent 框架提供内置的技能生命周期管理。Curator 是首创同类功能，而非渐进式改进。
## 如何使用
更新到 v0.12.0：
```
hermes update
hermes curator status
```
hermes curator status 显示上次运行时间、活动/陈旧/已归档技能的计数、固定列表以及五个最近最少使用的技能（即下一批归档候选）。
在首次自动运行触发之前，固定任何关键的自定义技能：
```
hermes curator pin my-deploy-skill
```
按需触发审查并检查报告：
```
hermes curator run --sync
cat ~/.hermes/logs/curator/<timestamp>/REPORT.md
```
要使用更便宜的辅助模型，请运行 hermes model，选择“辅助模型”，并将 Curator 路由到更小的提供商。每次运行都会写入 run.json（机器可读）和 REPORT.md（人类可读），因此审计历史保留在本地。
## 当前限制
> 截至 5 月 1 日的当前限制。
没有干运行 模式。hermes update 后的首次自动运行可以在不预览的情况下归档自定义技能。GitHub issue #18373 证实用户已经因此吃过亏。先固定，再更新。
bump_use() 在发布时损坏。PR #17782 指出，使用计数器在发布时零调用位点。修复在 PR #17932 中落地。在信任 hermes curator status 输出之前，请更新至补丁版本之后。
编辑活动不是生命周期信号。转换仅查看 use_count 和 last_used_at。一个经过大量修补但很少调用的技能仍可能被归档。Issue #17952 关闭了直接的关注点，但 patch_count 尚未驱动转换。
审查质量取决于辅助模型。构建伞式的合并效果取决于背后的模型。将小众技能误分类为冗余是有可能的。归档（而非删除）是安全网。
没有性能评分。生命周期纯粹基于时间。一个运行很少但总是成功的技能无法宣传其质量。
因此，最好的建议是更新，先固定，然后启用。默认设置是保守的，最坏的结果也是可恢复的归档，但首次运行缺乏干运行是用户在其他一切之前会遇到的摩擦。
## 谁受益，谁不受益
Curator 对于运行长期部署的 Hermes 用户（数周内技能会积累）、构建自我改进 Agent 的框架工程师（希望获得安全自我修改的参考架构），以及为生产环境评估 Agent 框架的团队（其中库的卫生状况影响规模化可靠性）来说，回报是显著的。
它对于从不积累技能的短期 Agent 运行，或锁定在 Claude Code 或 Codex（不存在等效功能）的团队没有帮助。在更新前跳过固定步骤且拥有关键自定义技能的用户，在首次自动运行时也面临风险。
## 从业者启示
既然技能库不会无限增长，那么长期运行的自我改进 Agent 现在即可部署。
## 链接
- Hermes Agent 仓库（约 5 分钟浏览）
- Curator 文档（约 10 分钟阅读）
- v2026.4.30 发布说明（约 15 分钟阅读）
- Issue #7816，原始提议
关注 @AlphaSignalAI 获取更多此类内容。
在 AlphaSignal.ai 订阅每日 AI 信号。280,000+ 开发者阅读。
## 问题？
问：什么是 Hermes Curator 以及它多久运行一次？
答：它是针对 Agent 创建的技能的自主后台维护通道，随 Hermes Agent v0.12.0 一起发布。默认情况下，它每 168 小时（7 天）运行一次，并且仅在 Agent 空闲 2 小时后运行。两个条件必须同时满足。
问：Hermes Curator 会删除我的技能吗？
答：不会。最坏的结果是归档到 ~/.hermes/skills/.archive/，可以使用 hermes curator restore <name> 恢复。Curator 永远不会自动删除。
问：如何防止 Hermes Curator 归档自定义技能？
答：运行 hermes curator pin <skill-name>。固定是一道硬性屏障：自动转换跳过已固定的技能，LLM 审查被指示忽略它们，且 skill_manage 拒绝对它们执行任何写入操作。
问：Curator 会碰从 agentskills.io 安装的技能或 Hermes 自带的技能吗？
答：不会。捆绑技能通过 ~/.hermes/skills/.bundled_manifest 受保护，Hub 安装的技能通过 ~/.hermes/skills/.hub/lock.json 受保护。只有 Agent 创建或用户手动编写的技能才会受到 Curator 操作的影响。
## 相关链接
- [AlphaSignal AI](https://x.com/AlphaSignalAI)
- [@AlphaSignalAI](https://x.com/AlphaSignalAI)
- [4.8K](https://x.com/AlphaSignalAI/status/2050269010735018074/analytics)
- [May 1](https://x.com/NousResearch/status/2049956455982182593)
- [471K](https://x.com/NousResearch/status/2049956455982182593/analytics)
- [published](https://x.com/Teknium/status/2049717907664581067?s=20)
- [@Teknium](https://x.com/@Teknium)
- [agentskills.io](https://agentskills.io/)
- [#18373](https://github.com/NousResearch/hermes-agent/issues/18373)
- [#17782](https://github.com/NousResearch/hermes-agent/issues/17782)
- [#17932](https://github.com/NousResearch/hermes-agent/pull/17932)
- [#17952](https://github.com/NousResearch/hermes-agent/issues/17952)
- [Hermes Agent repo](https://github.com/NousResearch/hermes-agent)
- [Curator documentation](https://hermes-agent.nousresearch.com/docs/user-guide/features/curator)
- [v2026.4.30 release notes](https://github.com/NousResearch/hermes-agent/releases/tag/v2026.4.30)
- [Issue #7816, original proposal](https://github.com/NousResearch/hermes-agent/issues/7816)
- [@AlphaSignalAI](https://x.com/@AlphaSignalAI)
- [AlphaSignal.ai](https://alphasignal.ai/)
- [agentskills.io](https://agentskills.io/)
- [1:39 AM · May 2, 2026](https://x.com/AlphaSignalAI/status/2050269010735018074)
- [4,890 Views](https://x.com/AlphaSignalAI/status/2050269010735018074/analytics)
---
*导出时间: 2026/5/2 10:09:13*