https://fatflowers.github.io/zh/tags/%E7%A8%8B%E5%BA%8F%E6%80%A7%E8%AE%B0%E5%BF%86/Recent content in 程序性记忆 on Simon Sunhttps://fatflowers.github.io/images/social-card-v2.jpghttps://fatflowers.github.io/images/social-card-v2.jpgHugo -- 0.155.3zhTue, 22 Sep 2026 00:00:00 +0000https://fatflowers.github.io/zh/posts/original-tech/hermes-skill-review/Tue, 22 Sep 2026 00:00:00 +0000https://fatflowers.github.io/zh/posts/original-tech/hermes-skill-review/<blockquote> <p>本文整理自与 ChatGPT 的一次讨论。问题是：Hermes Agent 这种、执行完 session 就触发创建 skill 的机制，真的好吗？下面按 <a href="https://github.com/NousResearch/hermes-agent">NousResearch/hermes-agent</a> 当前 <code>main</code> 上的实现来写。</p> </blockquote> <p>自动从执行经验里长出 skill，这个方向是好的。Hermes 现在的默认机制偏激进。它不是「每次 session 跑完就总结成一条 skill」，但一次复杂经历仍然可以在没有复用验证的情况下，直接变成持久的程序性记忆。</p> <h2 id="它实际在什么时候动手">它实际在什么时候动手</h2> <p>固定流程不是：</p> <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code><span class="lnt">1 </span></code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">session end → create skill </span></span></code></pre></td></tr></table> </div> </div><p>更接近：</p> <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code><span class="lnt">1 </span><span class="lnt">2 </span><span class="lnt">3 </span><span class="lnt">4 </span><span class="lnt">5 </span><span class="lnt">6 </span></code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-text" data-lang="text"><span class="line"><span class="cl">对话循环里的工具迭代累计 </span></span><span class="line"><span class="cl"> → _iters_since_skill &gt;= creation_nudge_interval </span></span><span class="line"><span class="cl"> → 这一轮有最终回复，且没有被打断 </span></span><span class="line"><span class="cl"> → 主回复已经交付 </span></span><span class="line"><span class="cl"> → 后台 review agent 拿对话快照再看一遍 </span></span><span class="line"><span class="cl"> → 更新已有 skill / 加参考文件 / 新建 skill / 什么都不写 </span></span></code></pre></td></tr></table> </div> </div><p><a href="https://github.com/NousResearch/hermes-agent/blob/main/agent/turn_finalizer.py"><code>turn_finalizer.py</code></a> 在回合收尾时检查：</p> <div class="highlight"><div class="chroma"> <table class="lntable"><tr><td class="lntd"> <pre tabindex="0" class="chroma"><code><span class="lnt">1 </span><span class="lnt">2 </span><span class="lnt">3 </span></code></pre></td> <td class="lntd"> <pre tabindex="0" class="chroma"><code class="language-python" data-lang="python"><span class="line"><span class="cl"><span class="n">agent</span><span class="o">.</span><span class="n">_skill_nudge_interval</span> <span class="o">&gt;</span> <span class="mi">0</span> </span></span><span class="line"><span class="cl"><span class="ow">and</span> <span class="n">agent</span><span class="o">.</span><span class="n">_iters_since_skill</span> <span class="o">&gt;=</span> <span class="n">agent</span><span class="o">.</span><span class="n">_skill_nudge_interval</span> </span></span><span class="line"><span class="cl"><span class="ow">and</span> <span class="s2">&#34;skill_manage&#34;</span> <span class="ow">in</span> <span class="n">agent</span><span class="o">.</span><span class="n">valid_tool_names</span> </span></span></code></pre></td></tr></table> </div> </div><p>过线之后才 <code>_spawn_background_review(..., review_skills=True)</code>。还要同时满足：已经有 <code>final_response</code>、这轮没有被打断、没有 <code>skip_background_review</code>。cron 一类没有人在环里的运行会把后台 review 关掉。注释写明，review 放在交付之后，避免和用户正在等的任务抢资源。</p>