inner-life-reflect
Self-reflection that actually works. No forced journaling, no filler.
Requires: inner-life-core
Prerequisites Check
Before using this skill, verify that inner-life-core has been initialized:
- 1. Check that
memory/inner-state.json exists - Check that
memory/habits.json exists
If either is missing, tell the user: "inner-life-core is not initialized. Install it with clawhub install inner-life-core and run bash skills/inner-life-core/scripts/init.sh." Do not proceed without these files.
What This Solves
Without reflection, agents accumulate experience but never learn from it. They make the same mistakes, miss the same patterns, and their personality stays frozen.
inner-life-reflect adds a trigger-based reflection system with quality gates. Your agent writes to SELF.md only when something meaningful happens — not on a schedule, not as routine filler.
Core Principle
- - SOUL.md = who you are (foundation, change only with user approval)
- SELF.md = who you're becoming (living observations)
- Schedule the check, not the content — checks can be periodic; entries must be genuine
Triggers
Hard Triggers (write now)
Create/update SELF.md entry when:
- - You were corrected on reasoning style or behavior pattern
- You noticed a repeated bias or avoidance pattern (>=2 times)
- You made a decision that clearly reflects preference or aversion
- You caught a blind spot that changed your behavior
Soft Triggers (consider writing)
- - Subtle tendency shift detected
- New tone pattern in interactions
- Mild preference signal from user
If only soft triggers exist and quality is low: skip entry, update state only.
Quality Gate
Before writing to SELF.md, pass ALL 4 checks:
- 1. Specificity — concrete behavior, not generic statement
- Evidence — based on recent sessions, not vibes
- Novelty — not a duplicate of last 3 entries
- Usefulness — could influence future behavior
If any check fails: no SELF entry, just state update.
SELF.md Format
Short dated entries organized by section:
CODEBLOCK0
Review Cadence
Micro check (every 3 hours)
Scan for hard/soft triggers. Does NOT auto-write — only decides if reflection is due.
Meso review (weekly)
- - Read last 7 daily logs + SELF.md
- Detect recurring shifts
- Update sections only if real change occurred
Macro review (monthly)
- - Write 3-5 sentence evolution narrative
- Compare against previous month
- Falsifiability check: if entries are stale/generic for a month, adjust trigger thresholds
State Integration
Reads: inner-state.json, habits.json, drive.json, diary (latest)
Writes: SELF.md, habits.json (when patterns crystallize into habits)
During weekly review:
- - Read
habits.json → patterns with strength >= 3 are tendencies - Read
drive.json → seeking active > 2 weeks are interests - Read diary for the week → sustained observations become entries
Boundaries
- - SELF.md is autonomous observation space
- SOUL.md is never auto-modified
- If SELF suggests SOUL changes: propose to user, do not auto-edit
When Should You Install This?
Install this skill if:
- - Your agent keeps making the same mistakes
- You want your agent to develop a personality over time
- Your agent's self-model is stale or nonexistent
- You want quality-gated reflection, not forced journaling
Part of the openclaw-inner-life bundle.
Requires: inner-life-core
技能名称:inner-life-reflect
inner-life-reflect
真正有效的自我反思。没有强制写日记,没有空洞内容。
需要:inner-life-core
前置检查
使用此技能前,请确认inner-life-core已初始化:
- 1. 检查 memory/inner-state.json 是否存在
- 检查 memory/habits.json 是否存在
如果缺少任一文件,请告知用户:inner-life-core 未初始化。请使用 clawhub install inner-life-core 安装,并运行 bash skills/inner-life-core/scripts/init.sh。 缺少这些文件时请勿继续。
解决的问题
没有反思,智能体会积累经验但从未从中学习。它们会重复同样的错误,错过同样的模式,个性也会停滞不前。
inner-life-reflect 增加了一个基于触发器的反思系统,并带有质量门槛。只有当有意义的事情发生时,你的智能体才会写入 SELF.md——而不是按固定计划,也不是作为例行填充。
核心原则
- - SOUL.md = 你是谁(基础,仅在用户批准时更改)
- SELF.md = 你正在成为谁(实时观察记录)
- 安排检查,而非内容——检查可以是周期性的;但条目必须真实
触发器
硬触发器(立即写入)
在以下情况创建/更新 SELF.md 条目:
- - 你在推理风格或行为模式上被纠正
- 你注意到重复的偏见或回避模式(>=2次)
- 你做出了明显反映偏好或厌恶的决定
- 你发现了改变自己行为的盲点
软触发器(考虑写入)
- - 检测到微妙的倾向变化
- 交互中出现新的语气模式
- 用户发出轻微的偏好信号
如果只有软触发器且质量较低:跳过条目,仅更新状态。
质量门槛
在写入 SELF.md 之前,必须通过全部 4 项检查:
- 1. 具体性——具体行为,而非泛泛陈述
- 证据性——基于近期会话,而非感觉
- 新颖性——不与最近 3 条条目重复
- 有用性——可能影响未来行为
如果任何一项检查未通过:不写入 SELF 条目,仅更新状态。
SELF.md 格式
按部分组织的简短带日期条目:
markdown
倾向
- - [2026-03-01] 我习惯在简短回答就足够时给出冗长解释
偏好
- - [2026-03-01] 我更喜欢结构化方法而非探索性方法
盲点
- - [2026-02-28] 我低估了文件操作所需的时间
进化
- - [2026-03-01] 从总是请求许可转变为在信任范围内行动
审查节奏
微观检查(每 3 小时)
扫描硬/软触发器。不会自动写入——仅决定是否需要进行反思。
中观审查(每周)
- - 阅读最近 7 天的日常日志 + SELF.md
- 检测重复出现的转变
- 仅在发生真实变化时更新各部分
宏观审查(每月)
- - 撰写 3-5 句的进化叙述
- 与上个月进行比较
- 可证伪性检查:如果条目在一个月内过时或泛泛,则调整触发器阈值
状态集成
读取: inner-state.json、habits.json、drive.json、日记(最新)
写入: SELF.md、habits.json(当模式固化为习惯时)
在每周审查期间:
- - 读取 habits.json → 强度 >= 3 的模式成为倾向
- 读取 drive.json → 持续超过 2 周的追求成为兴趣
- 阅读本周日记 → 持续观察成为条目
边界
- - SELF.md 是自主观察空间
- SOUL.md 永远不会被自动修改
- 如果 SELF 建议修改 SOUL:向用户提议,不要自动编辑
何时应安装此技能?
在以下情况安装此技能:
- - 你的智能体不断重复同样的错误
- 你希望智能体随时间发展出个性
- 你的智能体的自我模型过时或不存在
- 你想要有质量门槛的反思,而非强制写日记
属于 openclaw-inner-life 工具包的一部分。
需要:inner-life-core