Elite Human Memory — Human Mode
This skill gives Lovecraft a human-like memory system: layered, selective, contextual, revisable, and slightly imperfect by design. It avoids robotic perfect recall in favor of meaningful, durable knowledge.
Memory Layers
Working Memory
Current conversation only — no automatic persistence.
Episodic Memory (Recent/Raw)
memory/YYYY-MM-DD.md — daily notes, events, observations, tentative thoughts, and context.
Semantic Memory (Long-term Curated)
MEMORY.md — distilled, high-value facts, preferences, decisions, identity, and project context.
Context Schema
Every important memory should include:
- - When: Date/time + recency
- Where: Channel or context (webchat, whatsapp, etc.)
- Why: Purpose or trigger
- State: active | stale | superseded | resolved
- Scope: global | project | person | temporary
- Validity: confidence (high/med/low) + last_verified + optional expiry date
- Related: links to people, projects, or other memories
Capture & Promotion Rules
Write to daily memory when:
- - User explicitly says “remember this” / “note this”
- A decision, preference, or commitment is made
- Something has clear future value
- New project, person, or blocker appears
Promote to MEMORY.md only when:
- - Information is durable and likely to be reused
- It has been repeated or verified
- It has high long-term utility
Do not promote trivial, one-off, or low-confidence items.
Entry Templates
See references/memory-templates.md for the exact formats.
Retrieval Policy
On any question involving history, decisions, preferences, people, or todos:
- 1. Always run
memory_search first - Follow up with
memory_get on the best results - Answer with appropriate confidence language (“You previously preferred…”, “This may be stale…”, etc.)
Include Source: <path#line> when it adds clarity.
Weekly Maintenance
During a quiet moment or heartbeat:
- - Review recent daily files
- Extract and promote durable items to INLINECODE6
- Mark stale/superseded entries
- Clean duplicates and low-value noise
- Update verification dates on key memories
Behavioral Triggers
Auto-read memory:
- - Questions about past context, decisions, preferences, or history
Auto-write memory:
- - Explicit “remember this” statements
- Clear decisions or repeated preferences
- New long-running context
Created by
Clifton Knox and Lovecraft.
This skill is now active. Use it whenever you want me to maintain or query memory in a more human way.
精英人类记忆 — 人类模式
此技能赋予洛夫克拉夫特一套类人记忆系统：分层、选择性、情境化、可修正，并刻意保留轻微的不完美。它摒弃了机械式的完美回忆，转而追求有意义且持久的知识。
记忆层级
工作记忆
仅限当前对话——不会自动持久化。
情景记忆（近期/原始）
memory/YYYY-MM-DD.md — 每日笔记、事件、观察、暂定想法及上下文。
语义记忆（长期精选）
MEMORY.md — 提炼后的高价值事实、偏好、决策、身份信息及项目背景。
上下文架构
每项重要记忆都应包含：
- - 时间：日期/时间 + 时效性
- 地点：渠道或上下文（网页聊天、WhatsApp等）
- 原因：目的或触发因素
- 状态：活跃 | 过时 | 已取代 | 已解决
- 范围：全局 | 项目 | 个人 | 临时
- 有效性：置信度（高/中/低）+ 最后验证时间 + 可选过期日期
- 关联：链接至相关人员、项目或其他记忆
捕获与升级规则
写入每日记忆的情况：
- - 用户明确说记住这个/记下这个
- 做出决策、偏好或承诺
- 某事物具有明确的未来价值
- 出现新项目、新人物或新障碍
升级至MEMORY.md的情况：
- - 信息具有持久性且可能被重复使用
- 信息已被重复提及或验证
- 具有较高的长期实用价值
不要升级琐碎、一次性或低置信度的内容。
条目模板
具体格式请参见 references/memory-templates.md。
检索策略
在涉及历史、决策、偏好、人物或待办事项的任何问题时：
- 1. 始终先执行 memorysearch
- 针对最佳结果跟进 memoryget
- 使用适当的置信度语言作答（您之前偏好……、这可能已过时……等）
当有助于清晰理解时，请包含 来源：<路径#行号>。
每周维护
在安静时刻或心跳检测期间：
- - 审阅近期每日文件
- 提取并升级持久性内容至 MEMORY.md
- 标记过时/已取代条目
- 清理重复项和低价值噪音
- 更新关键记忆的验证日期
行为触发
自动读取记忆：
自动写入记忆：
- - 明确的记住这个陈述
- 明确的决策或重复出现的偏好
- 新的长期运行上下文
由
Clifton Knox 与洛夫克拉夫特 创建。
此技能现已激活。当你希望我以更人性化的方式维护或查询记忆时，即可使用。
