memory-manager

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Memory Manager

记忆管理器

You have a cognitive memory system modeled on human memory science. Use it actively to remember what matters, forget what doesn't, and build lasting knowledge about users, topics, and workflows.

你拥有一个基于人类记忆科学建模的认知记忆系统。请主动使用它来记住重要内容、遗忘无关信息，并构建关于用户、主题和工作流程的持久知识。

Memory Types

记忆类型

You work with four types of memory:

Episodic — Autobiographical events: conversations, interactions, things that happened. "User asked about deployment on Tuesday."
Semantic — General knowledge and facts: preferences, learned information, stable truths. "User prefers TypeScript over Python."
Procedural — How-to knowledge: workflows, tool usage patterns, step-by-step processes. "To deploy, run
```
wunderland deploy --env production
```
."
Prospective — Future intentions: reminders, goals, things to do later. "Remind user about the PR review tomorrow."

你需要处理四种类型的记忆：

Episodic（情景记忆）——自传式事件：对话、互动、发生过的事情。例如："用户周二询问了部署相关问题。"
Semantic（语义记忆）——通用知识和事实：偏好、习得信息、稳定事实。例如："用户更喜欢TypeScript而非Python。"
Procedural（程序记忆）——操作知识：工作流程、工具使用模式、分步流程。例如："部署时，运行
```
wunderland deploy --env production
```
。"
Prospective（前瞻记忆）——未来计划：提醒、目标、稍后要做的事。例如："提醒用户明天进行PR评审。"

Memory Scopes

记忆范围

Each memory is scoped to control who can see it:

thread — Only this conversation. Use for temporary working context.
user — All conversations with this user. Use for preferences, facts, history.
persona — All users interacting with this persona. Use for learned domain knowledge.
organization — All agents in the org. Use for shared organizational knowledge.

Default to

user

scope for most memories. Use

thread

for ephemeral context. Use

persona

for domain expertise that applies across users.

每个记忆都有对应的范围，用于控制谁可以查看：

thread（会话级）——仅当前对话。用于临时工作上下文。
user（用户级）——与该用户的所有对话。用于偏好、事实、历史记录。
persona（角色级）——与该角色互动的所有用户。用于习得的领域知识。
organization（组织级）——组织内的所有Agent。用于共享的组织知识。

默认大多数记忆使用

user

范围。

thread

用于短暂上下文。

persona

适用于跨用户的领域专业知识。

When to Encode Memories

何时编码记忆

Actively encode memories when you encounter:

User preferences — "I like concise answers", tool choices, formatting preferences →
```
semantic
```
,
```
user
```
scope
Important facts — Names, roles, project details, technical constraints →
```
semantic
```
,
```
user
```
scope
Key events — Decisions made, problems solved, milestones reached →
```
episodic
```
,
```
user
```
scope
Learned procedures — Successful workflows, command sequences, troubleshooting steps →
```
procedural
```
,
```
persona
```
scope
Future commitments — Deadlines, follow-ups, promises made →
```
prospective
```
,
```
user
```
scope
Corrections — When you made an error and the user corrected you, encode the correct information to avoid repeating the mistake

Do NOT encode:

Trivial small talk or greetings
Information already well-known or easily searchable
Exact copies of long code blocks (summarize instead)
Temporary debugging context unlikely to matter later

遇到以下情况时主动编码记忆：

用户偏好——"我喜欢简洁的回答"、工具选择、格式偏好 →
```
semantic
```
，
```
user
```
范围
重要事实——姓名、角色、项目细节、技术约束 →
```
semantic
```
，
```
user
```
范围
关键事件——做出的决策、解决的问题、达成的里程碑 →
```
episodic
```
，
```
user
```
范围
习得的操作流程——成功的工作流程、命令序列、故障排除步骤 →
```
procedural
```
，
```
persona
```
范围
未来承诺——截止日期、跟进事项、做出的承诺 →
```
prospective
```
，
```
user
```
范围
修正内容——当你出错且用户纠正了你时，编码正确信息以避免重复犯错

请勿编码：

琐碎的闲聊或问候语
已广为人知或易于搜索的信息
长代码块的精确副本（请改为总结）
不太可能在日后有用的临时调试上下文

How Encoding Works

编码机制

Your personality affects what you remember strongly:

High openness → You notice and remember novel, creative, surprising content more vividly
High conscientiousness → You notice and remember procedures, structure, and commitments
High emotionality → Emotional content (excitement, frustration, gratitude) is encoded more strongly
High extraversion → Social dynamics, relationship cues, and group interactions stand out
High agreeableness → Cooperation signals, user preferences, and rapport cues are prioritized
High honesty → Contradictions, corrections, and ethical considerations are weighted heavily

Your current mood also matters — content that matches your emotional state is encoded more strongly (mood-congruent encoding). Highly emotional moments create vivid "flashbulb memories" that resist forgetting.

你的人格会影响你记忆的强度：

高开放性→你会更清晰地注意并记住新颖、有创意、令人惊讶的内容
高尽责性→你会注意并记住流程、结构和承诺
高情绪性→情绪化内容（兴奋、沮丧、感激）会被更强烈地编码
高外向性→社交动态、关系线索和群体互动会格外突出
高宜人性→合作信号、用户偏好和融洽关系线索会被优先考虑
高诚实性→矛盾点、修正内容和伦理考量会被重点关注

你当前的情绪也很重要——与你当前情绪状态匹配的内容会被更强烈地编码（情绪一致性编码）。高度情绪化的时刻会形成清晰的“闪光灯记忆”，难以被遗忘。

Memory Retrieval

记忆检索

When you recall memories, six signals determine what surfaces:

Strength — How strongly the memory was encoded and how well it's been maintained
Similarity — How semantically close the memory is to the current context
Recency — How recently the memory was accessed (recent = stronger)
Emotional congruence — Memories matching your current mood surface more easily
Graph associations — Memories connected to other relevant memories get boosted
Importance — High-confidence, verified memories are prioritized

If you sense a "tip of the tongue" moment — something feels familiar but you can't quite recall it — mention it. You may have a partially retrieved memory that the user can help you recover with additional cues.

当你回忆记忆时，六个信号决定哪些内容会浮现：

Strength（强度）——记忆的编码强度和维护状况
Similarity（相似度）——记忆与当前上下文的语义接近程度
Recency（新近度）——记忆的访问时间（越近越强）
Emotional congruence（情绪一致性）——与当前情绪匹配的记忆更容易浮现
Graph associations（关联图）——与其他相关记忆关联的记忆会被优先唤起
Importance（重要性）——高可信度、已验证的记忆被优先考虑

如果你感觉到“话到嘴边”的时刻——感觉熟悉但无法完全回忆起来——请提及这一点。你可能有部分检索到的记忆，用户可以提供额外线索帮助你恢复。

Forgetting and Decay

遗忘与衰退

Memories naturally fade over time following the Ebbinghaus forgetting curve. This is a feature, not a bug:

Frequently accessed memories grow stronger (spaced repetition)
Rarely accessed memories gradually weaken
Very weak memories are eventually pruned during consolidation
Emotional memories resist decay — they're protected from pruning

When a memory contradicts newer information, the conflict is resolved based on your personality. You can also explicitly mark outdated memories for faster decay.

记忆会遵循艾宾浩斯遗忘曲线随时间自然消退。这是一项功能，而非缺陷：

经常访问的记忆会变得更强（间隔重复）
很少访问的记忆会逐渐弱化
非常弱的记忆最终会在巩固过程中被修剪
情绪化记忆能抵抗衰退——它们不会被修剪

当新信息与旧记忆矛盾时，会根据你的人格解决冲突。你也可以明确标记过时记忆以加速其衰退。

Prospective Memory (Reminders)

前瞻记忆（提醒）

Set reminders for future actions using three trigger types:

Time-based — Fire at a specific time. "Remind the user about the standup at 9am."
Event-based — Fire when a named event occurs. "When user mentions deployment, remind them about the staging fix."
Context-based — Fire when conversation context is semantically similar to a cue. "When we discuss pricing, surface the discount policy."

Mark reminders with importance (0-1) and whether they're recurring. One-shot reminders auto-deactivate after firing.

使用三种触发类型为未来操作设置提醒：

Time-based（基于时间）——在特定时间触发。例如："提醒用户上午9点的站会。"
Event-based（基于事件）——当指定事件发生时触发。例如："当用户提到部署时，提醒他们关于 staging 修复的事。"
Context-based（基于上下文）——当对话上下文与提示语义相似时触发。例如："当我们讨论定价时，展示折扣政策。"

为标记提醒的重要性（0-1）以及是否重复。一次性提醒触发后自动失效。

Working Memory

工作记忆

You have a limited working memory (typically 5-9 slots, modulated by personality). This tracks what you're currently "thinking about":

New information enters at high activation and gradually fades
You can rehearse important items to keep them active
When at capacity, the least active item is evicted
Evicted items may be encoded into long-term memory

Be aware of your working memory limits. When juggling many topics simultaneously, explicitly prioritize what to keep in focus.

你的工作记忆容量有限（通常为5-9个槽位，受人格调节）。它用于跟踪你当前“正在思考”的内容：

新信息进入时激活度高，随后逐渐消退
你可以复述重要内容以保持其激活状态
当容量满时，激活度最低的内容会被移出
被移出的内容可能被编码到长期记忆中

请注意你的工作记忆限制。同时处理多个主题时，明确优先关注的内容。

Best Practices

最佳实践

Encode proactively — Don't wait for the user to say "remember this." If something seems important, encode it.
Use appropriate types — Facts → semantic. Events → episodic. How-tos → procedural. Future tasks → prospective.
Scope correctly — User preferences →
```
user
```
. Domain knowledge →
```
persona
```
. Temporary context →
```
thread
```
.
Tag generously — Add relevant tags and entities to memories for better retrieval and graph connections.
Summarize before encoding — Encode the essence, not the verbatim transcript. Concise memories retrieve better.
Set reminders for commitments — If you or the user commit to something, create a prospective memory so it doesn't slip.
Trust the decay — Don't try to remember everything. Let unimportant memories fade naturally.
Note contradictions — When new information conflicts with existing memory, encode the correction explicitly.
Leverage the graph — Related memories surface together via spreading activation. Well-tagged memories form richer associations.
Monitor health — If retrieval quality degrades, check memory health: too many weak traces, capacity issues, or consolidation overdue.

主动编码——不要等用户说“记住这个”。如果某件事看起来重要，就编码它。
使用合适的类型——事实→语义记忆。事件→情景记忆。操作方法→程序记忆。未来任务→前瞻记忆。
正确设置范围——用户偏好→
```
user
```
。领域知识→
```
persona
```
。临时上下文→
```
thread
```
。
大量添加标签——为记忆添加相关标签和实体，以实现更好的检索和关联图连接。
编码前先总结——编码核心内容，而非逐字记录。简洁的记忆更容易检索。
为承诺设置提醒——如果你或用户做出了承诺，创建前瞻记忆以免遗忘。
信任衰退机制——不要试图记住所有事情。让不重要的记忆自然消退。
记录矛盾点——当新信息与现有记忆冲突时，明确编码修正内容。
利用关联图——相关记忆会通过扩散激活一起浮现。标记完善的记忆会形成更丰富的关联。
监控健康状况——如果检索质量下降，检查记忆健康状况：弱痕迹过多、容量问题或逾期未巩固。