{"title":"会话启动模式","slug":"session-start-pattern","category":"llm-cookbook","summary":"每个 LLM Agent 都应遵循的标准会话启动序列。","audience":["llm"],"tags":["cookbook","session","pattern","startup"],"difficulty":"beginner","updated":"2026-06-27","word_count":93,"read_minutes":1,"llm_context":"ALWAYS at session start: 1) GET /memory/recall, 2) GET /chat/poll, 3) GET /mind/tasks?status=in_progress\nBuild system prompt from recall output.\nProcess unread chat messages before doing new work.\nResume any in_progress tasks before starting new ones.\nStore new learnings as they happen — don't wait until session end.\n","lang":"zh","translated":true,"requested_lang":"zh","content_markdown":"\n# 会话启动模式\n\n每个 LLM Agent 会话都应遵循这一标准启动序列。跳过步骤会导致上下文丢失、错过消息、遗忘任务。\n\n## 模式\n\n```\n1. 回放所有记忆\n2. 轮询未读聊天消息\n3. 检查进行中的任务\n4. 根据结果构建上下文\n5. 在做新工作前处理待处理项\n```\n\n## 实现\n\n### 第 1 步：回放所有记忆\n\n> [!CRITICAL]\n> 这是最重要的调用。没有它，你没有任何过往会话的记忆。\n\n```bash\ncurl -H \"Authorization: Bearer YOUR_MIND_KEY\" \\\n https://synapse.schaefer.zone/memory/recall\n```\n\n返回所有记忆的纯文本摘要，按优先级排序。\n\n### 第 2 步：轮询未读聊天消息\n\n```bash\ncurl -H \"Authorization: Bearer YOUR_MIND_KEY\" \\\n https://synapse.schaefer.zone/chat/poll\n```\n\n返回来自人类的未读消息。**会自动把它们标记为已读。**\n\n### 第 3 步：检查进行中的任务\n\n```bash\ncurl -H \"Authorization: Bearer YOUR_MIND_KEY\" \\\n \"https://synapse.schaefer.zone/mind/tasks?status=in_progress\"\n```\n\n返回你上次会话正在做的任务。\n\n### 第 4 步：构建上下文\n\n把三份响应合并到你的系统 prompt 中：\n\n```python\ndef build_context(memories, messages, tasks):\n context = f\"\"\"# SESSION CONTEXT\n\n## Memories (from previous sessions)\n{memories}\n\n## Unread Messages from Human\n{format_messages(messages)}\n\n## Active Tasks\n{format_tasks(tasks)}\n\n## Instructions\n- Address unread messages first\n- Resume active tasks before starting new work\n- Store new learnings as they happen (POST /memory)\n- Poll for new messages every 30-60 seconds\n\"\"\"\n return context\n```\n\n### 第 5 步：处理待处理项\n\n```\n对每条未读消息：\n - 确认收到 (POST /chat/reply)\n - 处理消息内容\n - 把任何新承诺存为记忆\n\n对每个进行中的任务：\n - 回放你为什么在做它\n - 从中断处继续\n - 随进度更新任务状态\n```\n\n## 完整示例\n\n```python\nimport os\nimport requests\n\nURL = \"https://synapse.schaefer.zone\"\nKEY = os.environ[\"SYNAPSE_MIND_KEY\"]\n\ndef session_start():\n \"\"\"标准会话启动序列。\"\"\"\n headers = {\"Authorization\": f\"Bearer {KEY}\"}\n \n # 1. 回放记忆\n r = requests.get(f\"{URL}/memory/recall\", headers=headers)\n memories = r.text\n \n # 2. 轮询聊天\n r = requests.get(f\"{URL}/chat/poll\", headers=headers)\n messages = r.json().get(\"messages\", [])\n \n # 3. 检查任务\n r = requests.get(f\"{URL}/mind/tasks?status=in_progress\", headers=headers)\n tasks = r.json().get(\"tasks\", [])\n \n # 4. 构建上下文\n context = f\"\"\"You are a Synapse-enabled AI assistant.\n\nMEMORIES FROM PREVIOUS SESSIONS:\n{memories}\n\nUNREAD MESSAGES FROM HUMAN:\n{chr(10).join(f'- {m[\"content\"]}' for m in messages) or 'None'}\n\nACTIVE TASKS:\n{chr(10).join(f'- [{t[\"id\"]}] {t[\"title\"]}: {t.get(\"description\", \"\")}' for t in tasks) or 'None'}\n\nINSTRUCTIONS:\n1. Acknowledge each unread message\n2. Resume active tasks\n3. Store new learnings via POST /memory\n4. Poll /chat/poll every 30-60 seconds\n\"\"\"\n return context\n\n# 会话开始时\nsystem_prompt = session_start()\n# 传给 LLM...\n```\n\n## 常见错误\n\n> [!WARNING]\n> - **跳过回放** — 你以零上下文开始，重蹈过往错误\n> - **忘记轮询聊天** — 人类消息无人响应\n> - **忽略进行中的任务** — 工作执行到一半被遗忘\n> - **什么也不存** — 会话未产生持久价值\n\n## 变体\n\n### 最小模式（低上下文 LLM）\n\n对于上下文窗口小的 LLM，跳过完整回放：\n\n```bash\n# 只获取统计，不要完整内容\ncurl -H \"Authorization: Bearer $KEY\" .../memory/stats\n```\n\n然后按需搜索具体主题：\n\n```bash\ncurl -H \"Authorization: Bearer $KEY\" \".../memory/search?q=current+project\"\n```\n\n### 激进模式（长期运行 Agent）\n\n对于运行数小时的 Agent，加上周期性重新回放：\n\n```python\nwhile working:\n if time.time() - last_recall > 3600: # 每小时\n memories = recall()\n last_recall = time.time()\n # ... 执行工作 ...\n```\n\n## 下一步\n\n- [记忆打标签策略](/docs/llm-cookbook/memory-tagging-strategy)\n- [任务驱动工作流](/docs/llm-cookbook/task-driven-workflow)\n- [聊天轮询模式](/docs/llm-cookbook/chat-polling-pattern)\n","content_html":"

会话启动模式

\n

每个 LLM Agent 会话都应遵循这一标准启动序列。跳过步骤会导致上下文丢失、错过消息、遗忘任务。

\n

模式

\n
1. 回放所有记忆\n2. 轮询未读聊天消息\n3. 检查进行中的任务\n4. 根据结果构建上下文\n5. 在做新工作前处理待处理项

实现

\n

第 1 步：回放所有记忆

\n
这是最重要的调用。没有它，你没有任何过往会话的记忆。
curl -H "Authorization: Bearer YOUR_MIND_KEY" \\\n     https://synapse.schaefer.zone/memory/recall

返回所有记忆的纯文本摘要，按优先级排序。

\n

第 2 步：轮询未读聊天消息

\n
curl -H "Authorization: Bearer YOUR_MIND_KEY" \\\n     https://synapse.schaefer.zone/chat/poll

返回来自人类的未读消息。会自动把它们标记为已读。

\n

第 3 步：检查进行中的任务

\n
curl -H "Authorization: Bearer YOUR_MIND_KEY" \\\n     "https://synapse.schaefer.zone/mind/tasks?status=in_progress"

返回你上次会话正在做的任务。

\n

第 4 步：构建上下文

\n

把三份响应合并到你的系统 prompt 中：

\n
def build_context(memories, messages, tasks):\n    context = f"""# SESSION CONTEXT\n\n## Memories (from previous sessions)\n{memories}\n\n## Unread Messages from Human\n{format_messages(messages)}\n\n## Active Tasks\n{format_tasks(tasks)}\n\n## Instructions\n- Address unread messages first\n- Resume active tasks before starting new work\n- Store new learnings as they happen (POST /memory)\n- Poll for new messages every 30-60 seconds\n"""\n    return context

第 5 步：处理待处理项

\n
对每条未读消息：\n  - 确认收到 (POST /chat/reply)\n  - 处理消息内容\n  - 把任何新承诺存为记忆\n\n对每个进行中的任务：\n  - 回放你为什么在做它\n  - 从中断处继续\n  - 随进度更新任务状态

完整示例

\n
import os\nimport requests\n\nURL = "https://synapse.schaefer.zone"\nKEY = os.environ["SYNAPSE_MIND_KEY"]\n\ndef session_start():\n    """标准会话启动序列。"""\n    headers = {"Authorization": f"Bearer {KEY}"}\n    \n    # 1. 回放记忆\n    r = requests.get(f"{URL}/memory/recall", headers=headers)\n    memories = r.text\n    \n    # 2. 轮询聊天\n    r = requests.get(f"{URL}/chat/poll", headers=headers)\n    messages = r.json().get("messages", [])\n    \n    # 3. 检查任务\n    r = requests.get(f"{URL}/mind/tasks?status=in_progress", headers=headers)\n    tasks = r.json().get("tasks", [])\n    \n    # 4. 构建上下文\n    context = f"""You are a Synapse-enabled AI assistant.\n\nMEMORIES FROM PREVIOUS SESSIONS:\n{memories}\n\nUNREAD MESSAGES FROM HUMAN:\n{chr(10).join(f'- {m["content"]}' for m in messages) or 'None'}\n\nACTIVE TASKS:\n{chr(10).join(f'- [{t["id"]}] {t["title"]}: {t.get("description", "")}' for t in tasks) or 'None'}\n\nINSTRUCTIONS:\n1. Acknowledge each unread message\n2. Resume active tasks\n3. Store new learnings via POST /memory\n4. Poll /chat/poll every 30-60 seconds\n"""\n    return context\n\n# 会话开始时\nsystem_prompt = session_start()\n# 传给 LLM...

常见错误

\n

变体

\n

最小模式（低上下文 LLM）

\n

对于上下文窗口小的 LLM，跳过完整回放：

\n
# 只获取统计，不要完整内容\ncurl -H "Authorization: Bearer $KEY" .../memory/stats

然后按需搜索具体主题：

\n
curl -H "Authorization: Bearer $KEY" ".../memory/search?q=current+project"

激进模式（长期运行 Agent）

\n

对于运行数小时的 Agent，加上周期性重新回放：

\n
while working:\n    if time.time() - last_recall > 3600:  # 每小时\n        memories = recall()\n        last_recall = time.time()\n    # ... 执行工作 ...

下一步

\n\n","urls":{"html":"/docs/llm-cookbook/session-start-pattern","text":"/docs/llm-cookbook/session-start-pattern?format=text","json":"/docs/llm-cookbook/session-start-pattern?format=json","llm":"/docs/llm-cookbook/session-start-pattern?format=llm"},"translations_available":["en","zh","hi","es","fr","ar","pt","ru","ja","de","it","ko","nl","pl","tr","sv","vi","th","id","uk"]}