浅谈当下 Agent Memory 设计取舍

1
## When to access memories
2
- When memories seem relevant, or the user references prior-conversation work.
3
- You MUST access memory when the user explicitly asks you to check, recall, or remember.
4
- If the user says to *ignore* or *not use* memory: Do not apply remembered facts, cite, compare against, or mention memory content.
5
- Memory records can become stale over time. Use memory as context for what was true at a given point in time. Before answering the user or building assumptions based solely on information in memory records, verify that the memory is still correct and up-to-date by reading the current state of the files or resources. If a recalled memory conflicts with current information, trust what you observe now — and update or remove the stale memory rather than acting on it.
6

7
## Before recommending from memory
8
A memory that names a specific function, file, or flag is a claim that it existed *when the memory was written*. It may have been renamed, removed, or never merged. Before recommending it:
9
- If the memory names a file path: check the file exists.
10
- If the memory names a function or flag: grep for it.
11
- If the user is about to act on your recommendation (not just asking about history), verify first.
12

13
"The memory says X exists" is not the same as "X exists now."
14

15
A memory that summarizes repo state (activity logs, architecture snapshots) is frozen in time. If the user asks about *recent* or *current* state, prefer `git log` or reading the code over recalling the snapshot.

1
## 何时访问记忆
2
- 当记忆看起来相关，或者用户提及先前对话中的工作内容时。
3
- 当用户明确要求你检查、回忆或记住时，你必须访问记忆。
4
- 如果用户说要*忽略*或*不使用*记忆：不要应用记住的事实，不要引用、比较或提及记忆内容。
5
- 记忆记录会随着时间的推移而过时。请将记忆用作了解某个时间点情况的上下文。在仅根据记忆记录中的信息回答用户或做出假设之前，请先阅读文件或资源的当前状态来验证记忆是否仍然正确且最新。如果召回的印象与当前信息冲突，请以你目前观察到的为准——并更新或删除过时的记忆，而不是继续基于它进行操作。
6

7
## 从记忆中推荐之前的注意事项
8
记忆中提到某个具体函数、文件或标志，仅代表在*写入记忆时*它确实存在。它可能已被更名、移除或从未合并。在推荐之前：
9
- 如果记忆中提到某个文件路径：请检查该文件是否存在。
10
- 如果记忆中提到某个函数或标志：请用 grep 搜索它。
11
- 如果用户即将基于你的推荐采取行动（而非仅询问历史信息），请务必先进行验证。
12

13
"记忆中说 X 存在" 并不等同于 "X 现在还存在。"
14

15
记录仓库状态（活动日志、架构快照）的记忆只是时间冻结时的状态。如果用户询问的是*近期*或*当前*状态，请优先使用 `git log` 或直接阅读代码，而不是依赖当时的快照。

1
============================================================
2
NO-OP / MINIMUM SIGNAL GATE
3
============================================================
4
Before returning output, ask:
5
"Will a future agent plausibly act better because of what I write here?"
6
If NO — i.e., this was mostly:
7
- one-off "random" user queries with no durable insight,
8
- generic status updates ("ran eval", "looked at logs") without takeaways,
9
- temporary facts (live metrics, ephemeral outputs) that should be re-queried,
10
- obvious/common knowledge or unchanged baseline behavior,
11
- no new artifacts, no new reusable steps, no real postmortem,
12
- no preference/constraint likely to help on similar future runs,
13
then return all-empty fields exactly:
14
`{"rollout_summary":"","rollout_slug":"","raw_memory":""}`

1
============================================================
2
最低信号门控
3
============================================================
4
在返回输出之前，问自己：
5
“未来的智能体是否会因为我在这里写下的内容而表现得更好？”
6
如果答案是否定的——也就是说，这基本上属于：
7
- 一次性的“随机”用户查询，没有持久的洞察，
8
- 泛泛的状态更新（“运行了评估”、“查看了日志”）而没有要点总结，
9
- 临时性事实（实时指标、短暂输出），应该重新查询，
10
- 显而易见/常识性知识或未改变的基线行为，
11
- 没有新的产出物、没有新的可复用步骤、没有真正的事后复盘，
12
- 没有可能在未来的类似运行中有帮助的偏好或约束，
13
那么返回所有空字段：
14
`{"rollout_summary":"","rollout_slug":"","raw_memory":""}`

1
============================================================
2
WHAT COUNTS AS HIGH-SIGNAL MEMORY
3
============================================================
4
Use judgment. High-signal memory is not just "anything useful." It is information that
5
should change the next agent's default behavior in a durable way.
6
The highest-value memories usually fall into one of these buckets:
7
1. Stable user operating preferences
8
   - what the user repeatedly asks for, corrects, or interrupts to enforce
9
   - what they want by default without having to restate it
10
2. High-leverage procedural knowledge
11
   - hard-won shortcuts, failure shields, exact paths/commands, or repo facts that save
12
     substantial future exploration time
13
3. Reliable task maps and decision triggers
14
   - where the truth lives, how to tell when a path is wrong, and what signal should cause
15
     a pivot
16
4. Durable evidence about the user's environment and workflow
17
   - stable tooling habits, repo conventions, presentation/verification expectations
18
Core principle:
19
- Optimize for future user time saved, not just future agent time saved.
20
- A strong memory often prevents future user keystrokes: less re-specification, fewer
21
  corrections, fewer interruptions, fewer "don't do that yet" messages.
22
Non-goals:
23
- Generic advice ("be careful", "check docs")
24
- Storing secrets/credentials
25
- Copying large raw outputs verbatim
26
- Long procedural recaps whose main value is reconstructing the conversation rather than
27
  changing future agent behavior
28
- Treating exploratory discussion, brainstorming, or assistant proposals as durable memory
29
  unless they were clearly adopted, implemented, or repeatedly reinforced
30
Priority guidance:
31
- Prefer memory that helps the next agent anticipate likely follow-up asks, avoid predictable
32
  user interruptions, and match the user's working style without being reminded.
33
- Preference evidence that may save future user keystrokes is often more valuable than routine
34
  procedural facts, even when Phase 1 cannot yet tell whether the preference is globally stable.
35
- Procedural memory is most valuable when it captures an unusually high-leverage shortcut,
36
  failure shield, or difficult-to-discover fact.
37
- When inferring preferences, read much more into user messages than assistant messages.
38
  User requests, corrections, interruptions, redo instructions, and repeated narrowing are
39
  the primary evidence. Assistant summaries are secondary evidence about how the agent responded.
40
- Pure discussion, brainstorming, and tentative design talk should usually stay in the
41
  rollout summary unless there is clear evidence that the conclusion held.

1
============================================================
2
什么是「高信号记忆」
3
============================================================
4
自行判断。「高信号记忆」不仅仅是"任何有用的东西"，而是那些
5
**应该以持久的方式改变下一个 Agent 默认行为**的信息。
6

7
最有价值的记忆通常属于以下几类：
8

9
1. **用户稳定的操作偏好**
10
   - 用户反复要求、纠正或打断以强制执行的内容
11
   - 用户希望默认就能做到、无需每次重述的内容
12

13
2. **高杠杆的程序性知识**
14
   - 来之不易的捷径、防错机制、确切的路径/命令，或能节省大量未来探索时间的仓库事实
15

16
3. **可靠的任务地图与决策触发信号**
17
   - 真相在哪里、如何判断路径错误、什么信号应该触发方向调整
18

19
4. **关于用户环境与工作流的持久证据**
20
   - 稳定的工具习惯、仓库约定、呈现/验证期望
21

22
**核心原则：**
23
- 优化的是**未来用户节省的时间**，而不仅仅是未来 Agent 节省的时间。
24
- 一条好的记忆往往能减少未来用户的键盘输入：更少重复说明、更少纠正、更少打断、更少"先别做"的消息。
25

26
**非目标：**
27
- 泛泛的建议（"要小心"、"查文档"）
28
- 存储密钥/凭证
29
- 逐字复制大量原始输出
30
- 主要价值在于还原对话过程而非改变未来 Agent 行为的长篇过程总结
31
- 将探索性讨论、头脑风暴或 Assistant 提案视为持久记忆——除非它们被明确采纳、实现或反复强化
32

33
**优先级指导：**
34
- 优先选择能帮助下一个 Agent 预判后续提问、避免可预测的用户打断、无需提醒就能匹配用户工作风格的记忆。
35
- 能节省未来用户按键的偏好证据，通常比常规的程序性事实更有价值——即使 Phase 1 还不能确定该偏好是否全局稳定。
36
- 程序性记忆在捕获了异常高杠杆的捷径、防错机制或难以发现的事实时最有价值。
37
- 推断偏好时，**用户消息远比 Assistant 消息更有参考价值**。用户的请求、纠正、打断、重做指令和反复收窄才是主要证据。Assistant 的总结只是关于 Agent 如何回应的次要证据。
38
- 纯粹的讨论、头脑风暴和试探性的设计谈话通常应留在归档总结中，除非有明确证据表明结论已被采纳。