玛丽，玛丽，真倔强，你的智能体活多长？
像 Grok Bot 这样的新产品以及其他 元框架（meta-harness） 都要求我们创建智能体。它们应该活多久？
当你设计一个日历智能体时，它的会话应该运行多久？它应该在你于一家公司任职的整整五年里一直存活，还是每天都重置？
越来越强大的模型诱使我们构建永不关闭的永久会话。但长时间运行的会话会从内到外腐烂。
随着对话轮次不断累积，注意力会退化。现代模型能轻松在一份长文档中找出单个事实。但研究表明，即便在前沿模型中，智能体记忆也像人类记忆一样：它会随着增长而退化。1
临时指令会变成永久的幽灵。三月时告诉你的机器人“我这周感冒了，取消晨间会议”，到了十一月它仍在回避早晨时段。
长会话还会破坏安全性。一个对你的收件箱和日历持有多年读写权限的智能体就是一扇敞开的门。一封恶意邮件或一份恶意日历邀请就能污染对话，在数月之后悄然劫持你的日程。2
制胜的模式很简单：给你的日常助手 24 小时的生命，并把各项任务委派给范围狭窄的专家。
每日重置契合了人类实际的工作方式。白天，你的助手会记住即时上下文：“我要迟到十五分钟”，或者“留出两到三个空闲时段用于准备”。到了午夜，当前对话被清空，于是第二天从全新状态开始。
当有工作需要完成时，每日协调器会把任务交给一个单一用途的助手：日历智能体负责安排日程，邮件智能体负责起草回复，或者新闻智能体负责搜索网络。每个助手只存活三十秒，仅配备它所需的特定工具，完成任务后便消失。3
对于每日协调器而言，系统提示词充当调度器的角色：
You are Tomasz's daily coordinator.
Your session lives for 24 hours.
Workflow:
- Morning: Load preferences from `preferences.md`
& today's calendar.
- Intraday: Do not execute directly. Delegate
to sub-agents (`calendar_bot`, `email_bot`).
- Night: At midnight, save durable learnings
to `preferences.md` and terminate.
对于日历助手而言，系统提示词是一个无状态的执行器：
You are an ephemeral calendar specialist.
Process this single request, call the tool,
report the result, & terminate.
Rules:
- Time zone: America/Los_Angeles.
- Duration: 30 minutes.
- Hours: 9:00 AM – 6:00 PM.
- Always check availability first.
Never double-book.
- If full, propose 2 nearest openings & stop.
Output: Return event title, time, & attendees,
then exit.
在当日会话于午夜清空之前，会运行一次快速的整合过程。一个离线摘要器会回顾这一天，把持久的偏好（“Tomasz 偏好三十分钟的会议”）保存到磁盘上的永久笔记中，并丢弃其余每日的闲聊内容。4
Grok Bot 或你的聊天助手会自动执行这种睡眠周期吗？
目前还不会。大多数机器人会让会话线程永远保持打开，直到你点击“+ New Chat”，或者上下文压缩悄悄抹掉你的规则。
这首童谣问的是花园：银铃与鸟蛤壳，排成一行。那些持续存在的事物，按照某人选定的顺序排列。对智能体而言，答案是一样的：丢弃对话；把规则保存在文件里，让你的智能体和它的花园保持健康。
-
Amirali Ebrahimzadeh 和 Seyyed Muhammad Salili，《并非所有针都能被找到：事实分布与提示如何塑造长上下文大语言模型中的推理》，arXiv:2601.02023，2026 年 1 月；Kelly Hong 等，《上下文腐化：输入 token 增加如何影响大语言模型性能》，Chroma Research，2025 年。↩︎
-
《大语言模型智能体中的潜伏记忆投毒》，arXiv:2605.15338，2026 年 5 月。展示了有状态 AI 助手中持续性的跨会话记忆投毒攻击。↩︎
-
Shiyang Chen，《治理衰减：上下文压缩如何悄然抹除长时程大语言模型智能体中的安全约束》，arXiv:2606.22528，2026 年 6 月。展示了压缩在 30–59% 的回合中丢弃了常设规则。↩︎
-
Anthropic，《梦境：记忆巩固》，研究预览
dreaming-2026-04-21，2026 年 4 月；以及 Letta v2 有状态智能体框架（2026 年）。↩︎
Mary, Mary, quite contrary, how long do your agents live?
New products like Grok Bot & other meta-harnesses ask us to create agents. How long should they live?
When you design a calendar agent, how long should its session run? Should it stay alive for your entire five-year tenure at a company, or reset every day?
Ever more powerful models tempt us to build perpetual sessions that never close. But long-running sessions rot from the inside out.
As conversational turns pile up, attention degrades. Modern models easily spot a single fact in a long document. But research shows even across frontier models, agent memory is like human memory : it degrades as it grows.1
Temporary commands turn into permanent ghosts. Tell your bot in March, “I have a cold this week, cancel morning meetings,” & by November it is still avoiding morning slots.
Long sessions also break security. An agent holding multi-year read & write access to your inbox & calendar is an open door. One malicious email or calendar invite can poison the conversation, quietly hijacking your schedule months down the road.2
The winning pattern is simple : give your daily assistant a 24-hour life, & delegate individual tasks to narrow specialists.
A daily reset matches how humans actually work. During the day, your assistant remembers immediate context : “I’m running fifteen minutes late,” or “keep two to three free for prep.” At midnight, the active conversation wipes clean so tomorrow starts fresh.
When work needs doing, the daily coordinator hands the job to a single-purpose helper : a calendar agent to schedule, an email agent to draft a reply, or a news agent to search the web. Each helper lives for thirty seconds with only the specific tools it needs, does the job, & disappears.3
For the daily coordinator, the system prompt acts as a dispatcher :
You are Tomasz's daily coordinator.
Your session lives for 24 hours.
Workflow:
- Morning: Load preferences from `preferences.md`
& today's calendar.
- Intraday: Do not execute directly. Delegate
to sub-agents (`calendar_bot`, `email_bot`).
- Night: At midnight, save durable learnings
to `preferences.md` and terminate.
For the calendar helper, the system prompt is a stateless executor :
You are an ephemeral calendar specialist.
Process this single request, call the tool,
report the result, & terminate.
Rules:
- Time zone: America/Los_Angeles.
- Duration: 30 minutes.
- Hours: 9:00 AM – 6:00 PM.
- Always check availability first.
Never double-book.
- If full, propose 2 nearest openings & stop.
Output: Return event title, time, & attendees,
then exit.
Before the day’s session wipes at midnight, a quick consolidation pass runs. An offline summarizer reviews the day, saves lasting preferences (“Tomasz prefers thirty-minute meetings”) into a permanent note on disk, & throws away the rest of the daily chatter.4
Does Grok Bot or your chat assistant perform this sleep cycle automatically?
Not today. Most bots leave threads open forever until you click “+ New Chat” or context compaction silently erases your rules.
The nursery rhyme asks about a garden : silver bells & cockle shells, all in a row. Things that persist, in an order someone chose. The answer for an agent is the same : throw away the conversation ; keep the rules in a file to keep your agent & its garden healthy.
-
Amirali Ebrahimzadeh and Seyyed Muhammad Salili, “Not All Needles Are Found: How Fact Distribution and Prompting Shape Inference in Long-Context LLMs,” arXiv:2601.02023, January 2026; Kelly Hong et al., “Context rot: How increasing input tokens impacts LLM performance,” Chroma Research, 2025. ↩︎
-
“Sleeper Memory Poisoning in LLM Agents,” arXiv:2605.15338, May 2026. Demonstrates persistent cross-session memory poisoning attacks in stateful AI assistants. ↩︎
-
Shiyang Chen, “Governance Decay: How Context Compaction Silently Erases Safety Constraints in Long-Horizon LLM Agents,” arXiv:2606.22528, June 2026. Demonstrates that compaction drops standing rules in 30–59% of episodes. ↩︎
-
Anthropic, “Dreams: Memory Consolidation,” research preview
dreaming-2026-04-21, April 2026; and the Letta v2 stateful agent framework (2026). ↩︎