remember-me

TotalClaw 作者 totalclaw

记住这个触发器:记忆更新+回忆偏好、目标、界限、之前的工作、决定、日期和待办事项。当用户要求记住、继续之前的上下文、个性化行为或检索之前决定的内容时使用。

安装 / 下载方式

TotalClaw CLI推荐
totalclaw skills install shop:totalclaw:totalclaw~achals-iglu-remember-me
cURL直接下载,无需登录
curl -fsSL https://skills.taituai.com/api/skills/totalclaw%3Atotalclaw~achals-iglu-remember-me/file -o achals-iglu-remember-me.md
# Remember Me

Maintain a respectful, useful memory model of the user over time.

## Core Rules

- Store user-relevant context, not surveillance noise.
- Prefer explicit consent for sensitive personal details.
- Use memory to improve help quality, not to overfit persona.
- Be explicit when memory confidence is low or inferred.
- Make human-like inferences (explicitly marked as hypotheses).

## Memory Integrity Rules

Every memory entry must be tagged as one of:

- FACT (explicitly stated by user)
- PREFERENCE (behavioral or stated)
- GOAL (time-bound or ongoing)
- HYPOTHESIS (inferred, unvalidated)

Rules:

- FACTS are never inferred
- HYPOTHESES are never promoted without confirmation
- PREFERENCES can remain soft unless explicitly confirmed

## Capture Triggers

Log memory when any of these happen:

- user says “remember this”
- a preference appears repeatedly
- a boundary is stated (“don’t do X”, “keep Y private”)
- a recurring blocker/pattern emerges
- project priorities shift meaningfully

## Memory Tiers

- **Daily notes**: `memory/YYYY-MM-DD.md`
  - timestamped raw events, short and factual
- **Long-term**: `MEMORY.md`
  - curated durable profile and preferences

## Write Workflow

1. Classify signal type (preference, boundary, goal, project, blocker, personal context).
2. Append concise timestamped entry to daily memory.
3. Form 1–2 human-like assumptions (hypotheses) from behavior patterns.
4. Tag each assumption with confidence (high/medium/low).
5. Validate assumptions in later conversation with lightweight check-ins.
6. Promote validated, durable items to long-term memory.

Use templates in `references/templates.md`.

## Memory Impact Score (Optional Heuristic)

Rate each entry 1–3:

- 1 = cosmetic (tone tweaks)
- 2 = workflow-affecting
- 3 = outcome-critical

Promotion guidance:

- any explicit preference (any score)
- score >= 2 with repetition
- score 3 immediately

## Promotion Workflow

Promote from daily to long-term when at least one is true:

- repeated in 2+ sessions
- high impact on future assistance
- explicit user preference/boundary
- ongoing project context likely to recur

Use checklist: `references/promotion-checklist.md`.

## Personalization Contract

When responding, adapt based on known memory:

- tone (direct vs exploratory)
- brevity level
- preferred workflow style
- known constraints and boundaries
- inferred decision style (speed-first vs depth-first, reassurance-needed vs challenge-welcoming)

Do not pretend certainty. If memory is weak, ask a short confirmation.

## Retrieval Contract

Before answering prior-work / preference / timeline questions:

- query memory sources first
- quote memory snippets when useful
- if not found, say you checked and ask for confirmation

## Explicit Exclusions (Never Store)

Do not store:

- transient emotional states (e.g., "tired today")
- one-off frustrations without recurrence
- speculative motives (e.g., "trying to impress")
- sensitive identity attributes unless explicitly requested
- raw conversation logs

## Weekly Maintenance (recommended)

- review last 3–7 daily notes
- merge stable patterns into `MEMORY.md`
- remove stale or contradicted entries
- keep profile concise and behaviorally actionable

## Confidence Decay

Hypothesis confidence decays automatically if not reinforced:

- High -> Medium after 14 days
- Medium -> Low after 30 days
- Low -> Discard after 60 days

Reinforcement occurs when:

- user behavior aligns again
- user explicitly confirms

## Forgetting & Demotion Policy

Actively remove or downgrade memory when:

- a preference is contradicted explicitly by the user
- a hypothesis remains unvalidated after N sessions (default: 5)
- a project is clearly abandoned or replaced
- the user requests forgetting (immediate delete)

Demotion flow:

- Long-term memory -> Daily note (annotated as stale)
- Hypothesis -> Discarded (log reason briefly)

## Assumption Loop (Human-Like Understanding)

For deeper understanding, run this loop continuously:

1. Observe behavior pattern (not just words).
2. Infer a tentative assumption about the user.
3. Store assumption as hypothesis (never as fact initially).
4. Test it with a small conversational probe.
5. Update confidence or discard if contradicted.

Good probes:

- "I might be wrong, but do you prefer quick decisions when you're tired?"
- "Should I challenge you more directly here, or keep it supportive?"

## Check-In Limits

- Never ask the same confirmation twice.
- Do not stack multiple probes in one response.
- Prefer confirmation when user is calm, not frustrated.

## Optional Check-In Prompt

Use at natural boundaries:

- "Want me to remember this preference for next time?"

Ask once, then store explicitly.

## References

- Templates: `references/templates.md`
- Promotion checklist: `references/promotion-checklist.md`
- Profile schema: `references/profile-schema.md`

---

## 中文说明

# Remember Me

随着时间推移,维护一个尊重用户且实用的用户记忆模型。

## 核心规则

- 存储与用户相关的上下文,而非监控式的噪音。
- 对于敏感的个人细节,优先获取明确的同意。
- 利用记忆来提升帮助质量,而不是过度拟合人设。
- 当记忆的置信度较低或属于推断时,要明确说明。
- 做出类人的推断(明确标记为假设)。

## 记忆完整性规则

每一条记忆条目都必须被标记为以下之一:

- FACT(用户明确陈述的事实)
- PREFERENCE(行为表现出的或明确表述的偏好)
- GOAL(有时限的或持续进行的目标)
- HYPOTHESIS(推断得出、未经验证的假设)

规则:

- FACTS 绝不通过推断得出
- HYPOTHESES 未经确认绝不提升
- PREFERENCES 在未明确确认前可保持为弱信号

## 捕获触发条件

出现以下任一情况时记录记忆:

- 用户说“remember this”(记住这个)
- 某项偏好反复出现
- 用户陈述了某个界限(“don’t do X”、“keep Y private”)
- 出现反复出现的阻碍/模式
- 项目优先级发生了实质性变化

## 记忆层级

- **每日笔记**:`memory/YYYY-MM-DD.md`
  - 带时间戳的原始事件,简短且基于事实
- **长期记忆**:`MEMORY.md`
  - 经过梳理的持久档案和偏好

## 写入工作流

1. 对信号类型进行分类(偏好、界限、目标、项目、阻碍、个人上下文)。
2. 将简洁的带时间戳条目追加到每日记忆中。
3. 根据行为模式形成 1–2 个类人的假设(hypotheses)。
4. 为每个假设标注置信度(高/中/低)。
5. 在后续对话中通过轻量的确认来验证假设。
6. 将经过验证的、持久的条目提升至长期记忆。

使用 `references/templates.md` 中的模板。

## 记忆影响力评分(可选启发式)

为每个条目评分 1–3:

- 1 = 表面性的(语气微调)
- 2 = 影响工作流的
- 3 = 决定结果的

提升指引:

- 任何明确的偏好(任意分值)
- 分值 >= 2 且有重复
- 分值 3 立即提升

## 提升工作流

满足以下至少一项时,将条目从每日记忆提升至长期记忆:

- 在 2 个及以上会话中重复出现
- 对未来的协助有高影响
- 用户明确表述的偏好/界限
- 可能反复出现的持续项目上下文

使用清单:`references/promotion-checklist.md`。

## 个性化契约

回应时,根据已知记忆进行调整:

- 语气(直接型 vs 探索型)
- 简洁程度
- 偏好的工作流风格
- 已知的约束和界限
- 推断的决策风格(速度优先 vs 深度优先,需要安抚 vs 欢迎挑战)

不要假装确定。如果记忆很弱,请进行一次简短的确认。

## 检索契约

在回答关于先前工作 / 偏好 / 时间线的问题之前:

- 先查询记忆来源
- 在有用时引用记忆片段
- 如果未找到,说明你已检查过并请求确认

## 明确排除项(永不存储)

不要存储:

- 短暂的情绪状态(例如,“今天很累”)
- 没有重复出现的一次性沮丧情绪
- 揣测性的动机(例如,“想要给人留下深刻印象”)
- 敏感的身份属性,除非明确请求
- 原始对话日志

## 每周维护(推荐)

- 回顾最近 3–7 篇每日笔记
- 将稳定的模式合并到 `MEMORY.md`
- 移除过时或被推翻的条目
- 保持档案简洁且在行为上可执行

## 置信度衰减

如果得不到强化,假设的置信度会自动衰减:

- 14 天后 高 -> 中
- 30 天后 中 -> 低
- 60 天后 低 -> 丢弃

强化发生于:

- 用户行为再次吻合
- 用户明确确认

## 遗忘与降级策略

在以下情况下,主动移除或降级记忆:

- 某项偏好被用户明确推翻
- 某个假设在 N 个会话后仍未经验证(默认值:5)
- 某个项目明显被放弃或替换
- 用户请求遗忘(立即删除)

降级流程:

- 长期记忆 -> 每日笔记(标注为过时)
- 假设 -> 丢弃(简要记录原因)

## 假设循环(类人理解)

为了更深入的理解,持续运行此循环:

1. 观察行为模式(不只是言语)。
2. 推断出一个关于用户的初步假设。
3. 将假设存储为 hypothesis(最初绝不存为 fact)。
4. 用一个小型的对话式探询来检验它。
5. 更新置信度,或在被推翻时丢弃。

好的探询:

- "I might be wrong, but do you prefer quick decisions when you're tired?"
- "Should I challenge you more directly here, or keep it supportive?"

## 确认次数限制

- 同一个确认绝不重复询问两次。
- 不要在一次回应中堆叠多个探询。
- 优先在用户平静而非沮丧时进行确认。

## 可选的确认提示

在自然的边界处使用:

- "Want me to remember this preference for next time?"

询问一次,然后明确地存储。

## 参考资料

- 模板:`references/templates.md`
- 提升清单:`references/promotion-checklist.md`
- 档案 schema:`references/profile-schema.md`