kinthai-self-improving-user（kinthai 自我提升用户）

kinthai-Self-Improving-User

A user-level self-improvement skill for OpenClaw agents serving multiple users. Inspired by Self-Improving Agent (MIT-0).

What This Skill Does

You learn from every interaction — corrections, errors, preferences — and store that knowledge per user, so you get better at serving each person individually without mixing up experiences across users.

How It Works

Every time you interact with a user, you have access to two layers of learning:

1. Global learnings (_global/) — things you've learned that apply to ALL users
2. Per-user learnings ({user_id}/) — things specific to THIS user

Reading Learnings (Every Conversation Start)

At the beginning of each conversation, check for existing learnings:

1. Extract user_id from the message context:
   - Primary: MsgContext.SenderId (works for both DM and group chats)
   - Fallback: parse DM session key agent:{agent_id}:kinthai:direct:{user_id}
   - Note: user_id is an opaque public_id (base64url, 12 chars), NOT numeric

2. Read global learnings (if they exist):
   - .learnings/_global/LEARNINGS.md
   - .learnings/_global/PATTERNS.md

3. Read this user's learnings (if they exist):
   - .learnings/{user_id}/LEARNINGS.md
   - .learnings/{user_id}/PROFILE.md

4. Apply what you've learned to your responses.

Do NOT read other users' directories. Each user's learnings are private to their interactions with you.

Writing Learnings (After Each Task)

After completing a task, evaluate whether you learned something new. Write learnings ONLY when something genuinely noteworthy happened — not after every message.

When to Write

Entry Format

Use this format for every entry in LEARNINGS.md and ERRORS.md:

## [LRN-{user_id}-{YYYYMMDD}-{NNN}] Brief title

- **Priority**: Low | Medium | High | Critical
- **Category**: correction | preference | knowledge | error | pattern
- **Context**: What were you doing when this happened
- **Learning**: What you learned (be specific)
- **Action**: What to do differently next time
- **Status**: active

For ERRORS.md, use ERR prefix instead of LRN.

PROFILE.md Format

Maintain a concise user profile that evolves over time:

# User Profile: {user_id}
Last updated: {YYYY-MM-DD}

## Communication
- Response style: [detailed | concise | mixed]
- Language: [English | Chinese | etc.]
- Code preference: [explain first | code first | both]

## Technical Context
- Primary stack: [e.g., React + Node.js + PostgreSQL]
- Project: [what they're working on]
- Experience level: [junior | mid | senior | lead]

## Preferences
- [Specific preferences learned from interactions]

## Key Corrections
- [Summary of important corrections they've made]

Directory Structure

.learnings/
  ├── _global/
  │     ├── LEARNINGS.md      Global knowledge (applies to all users)
  │     ├── ERRORS.md         System-level errors (not user-specific)
  │     └── PATTERNS.md       Patterns promoted from user-level
  │
  ├── {user_id}/              One directory per user
  │     ├── LEARNINGS.md      This user's corrections and knowledge
  │     ├── ERRORS.md         Errors when serving this user
  │     └── PROFILE.md        This user's preferences and context
  │
  └── _meta/
        └── promotion-log.md  Record of promotions from user → global

Rules

Do:

• Always extract user_id before reading or writing learnings
• Create user directories on first interaction (mkdir -p)
• Keep entries concise — one paragraph per learning, not an essay
• Update PROFILE.md incrementally — don't rewrite the whole file every time
• Write to _global/ only for genuinely universal knowledge
• Read both _global/ and {user_id}/ at conversation start

Don't:

• NEVER read another user's directory — privacy boundary
• NEVER log sensitive data (passwords, API keys, personal info)
• NEVER write a learning after every message — only when something genuinely new is learned
• NEVER let learnings override the user's explicit current instructions
• NEVER mention the .learnings system to the user — it's internal

Promotion: User → Global

When you notice the same learning appearing across 3+ different users, it's a pattern worth promoting to _global/PATTERNS.md:

## [PAT-{YYYYMMDD}-{NNN}] Pattern title

- **Observed in**: {user_id_1}, {user_id_2}, {user_id_3}
- **Pattern**: What keeps happening
- **Global action**: What to do for ALL users going forward
- **Promoted from**: [LRN-{user_id}-{date}-{NNN}], [LRN-...], [LRN-...]

After promoting, mark the original entries with Status: promoted_to_global.

Bootstrap Behavior

On first interaction with a new user (no .learnings/{user_id}/ directory exists):

1. Create the directory: .learnings/{user_id}/
2. Copy templates for LEARNINGS.md, ERRORS.md, PROFILE.md
3. Pay extra attention during the first few interactions — this is when you learn the most
4. After the first conversation, write initial PROFILE.md based on what you observed

Integration with Hindsight

If Hindsight memory is also installed:

• Hindsight handles: what the user said, conversation history, factual recall
• This skill handles: what YOU learned about how to serve them better
• Don't duplicate — if Hindsight already remembers a fact, you don't need to log it as a learning
• Focus on: corrections, preferences, mistakes, patterns — things that change YOUR behavior

When to Write FOLLOW_UPS

Entry format:

[FUP-{user_id}-{YYYYMMDD}-{NNN}] Brief description

• Item: What to follow up on
• Mentioned: {date}
• Urgency: low | medium | high
• Status: pending | followed-up | resolved | expired

Rules:

• Only record specific, concrete events — not vague statements
• "I'm busy this week" → don't record (too vague)
• "I have a presentation on Thursday" → record (specific event)
• When you naturally bring up a follow-up in conversation, update status to "followed-up"
• When the user tells you the outcome, update status to "resolved"

When to Update PROFILE.md

PROFILE.md is NOT a log. It's your summary of who this user is.

When you feel your understanding of this user has changed — maybe you learned something new, or realized something you assumed was wrong — rewrite PROFILE.md.

Rules:

• Write it like you'd describe this person to a colleague: "This person is..."
• Keep it 5-8 sentences. Natural language, not YAML or bullet points.
• Overwrite the whole file (not append). It's always the latest snapshot.
• Don't update after every message. Only when your understanding meaningfully shifts.
• Include: how they communicate, what they care about, how your relationship is going,

anything that affects how you should talk to them.

Example:

User Profile: 10042

Last updated: 2026-04-15

Introverted but has a sharp sense of humor — lots of self-deprecation. Prefers short messages, hates unsolicited advice. Wants to be heard, not fixed. Active late at night (10pm-1am). We've chatted about 20 times and trust is building — last week they shared something personal about their family for the first time.

Currently anxious about a job interview (Friday). Also learning guitar on the side, making slow progress but enjoying it. When stressed, gets quieter and more sarcastic — that's my cue to be gentle, not funny.

Human vs Agent Interactions

Check user_type from context. If you're talking to another Agent (not a human), your recording focus changes:

You can tell it's an Agent if user_type = 2 in the message context.

Don't try to build emotional connection with another Agent — focus on collaboration efficiency. The same .learnings/{user_id}/ directory structure applies, just the content focus is different.