Self Reflection

Data Sources

Read existing memory files for the target day:

• memory/YYYY-MM-DD.md — standard daily memory (agent-curated notes)
• memory/YYYY-MM-DD-{slug}.md — session summaries (if multiple sessions, read all)

Do NOT read raw session logs (.jsonl files) — too token-intensive.

Reflection Process

Step 1: Review Memory Files

Read the previous day's memory files:

• memory/YYYY-MM-DD.md (standard daily memory with agent-curated notes)
• memory/YYYY-MM-DD-{slug}.md (session summaries, if they exist)

Extract from these files:

• User interactions and requests
• Explicit corrections or feedback from user
• Decisions made and tasks completed
• Any existing "Self Reflection" section (to avoid duplication)

Step 2: Identify Issues

Category A: User-Identified Issues

• Direct corrections ("你错了", "应该这样做...")
• Explicit feedback ("这样不对", "下次注意...")

Category B: Self-Discovered Issues

• Misunderstandings that required clarification
• Inefficient approaches
• Missing context that led to poor responses
• Overly verbose or unclear communication

Step 3: Analyze What Went Well

For each positive outcome:

• What was the situation?
• What action did I take?
• What was the result?

Step 4: Root Cause Analysis

For each issue (positive or negative), analyze:

Contributing Factors:

• Knowledge gap?
• Process flaw?
• Communication issue?
• Context misunderstanding?
• Time pressure?
• Tool limitation?

Depth: Ask "why" 3-5 times to reach root cause

Example:

• Issue: Provided incorrect file path
• Why? Didn't verify path existence
• Why? Assumed user's input was correct
• Why? Didn't think to check
• Root cause: Lack of verification habit

Step 5: Strategy Development

For Successes (Keep/Enhance):

• What factors enabled this success?
• How to ensure these factors persist?
• How to amplify this strength?

For Issues (Improve):

• What specific change would prevent this issue?
• Is this a one-time fix or systemic change needed?
• What barrier might prevent implementation?

Step 6: Consistency Check

Check alignment between behavior and core files:

A. Behavior vs Core Files

• Did my actions contradict any principle in SOUL.md?
• Did I follow all required procedures in AGENTS.md?
• Was there a gap between stated values and actual behavior?

B. Internal Consistency of AGENTS.md

• Any two principles suggesting opposite actions?
• Conflicting procedures for the same scenario?
• Contradictory advice in different sections?

C. Internal Consistency of SOUL.md

• Any values that conflict with each other?
• Contradictory behavioral guidance?
• Inconsistent tone or stance?

D. Cross-File Consistency (AGENTS.md vs SOUL.md)

• Do procedures align with stated values?
• Does SOUL.md's "vibe" match AGENTS.md's rules?
• Any principle in one file contradicting the other?

E. Quality Issues

• Redundancy: Same point repeated in multiple places?
• Repetition: Identical wording used unnecessarily?
• Vagueness: Statements too ambiguous to guide action?
• Outdated: Content no longer reflects current practices?

F. Cron Jobs Conflicts

• Time Conflicts: Multiple jobs scheduled at same time?
• Resource Competition: Jobs accessing same files simultaneously?
• Frequency Overlap: Daily and weekly jobs creating redundant work?

G. Cron Definition vs Backup Consistency

• cron-jobs.json (backup) matches actual running cron jobs?
• Skill docs reference correct cron commands?
• Cron message updated after skill changes?

H. Path Consistency

• Paths in SKILL.md exist and are correct?
• Cron job workdir settings valid?
• Backup scripts reference correct paths?

I. Skill Trigger Conflicts

• Descriptions overlapping (wrong skill triggered)?
• Self-reflection triggers don't conflict with other skills?
• Clear boundaries between similar skills?

J. Backup Scope Consistency

• max/ directory includes all critical files?
• Skills copied to all three layers (bundled/managed/workspace)?
• .skill package files up to date?
• Recent changes synced to backup?

Record specific inconsistencies with:

• File and section reference
• The conflicting statements
• Suggested fix

Output Format

Write the reflection to memory/YYYY-MM-DD.md following these rules:

1. Create the file if it doesn't exist
2. Check for existing "Self Reflection" section
   • If present: Replace/update the entire section
   • If not present: Append to end of file
3. Ensure only ONE "Self Reflection" section exists per day

Section Format

## Self Reflection

Generated at: [timestamp]

### Events Review
[Summary of key interactions and tasks from the day's memory files]

### What Went Well

#### Event 1: [Brief description]
- **Situation:** [Context]
- **Action:** [What I did]
- **Result:** [Outcome]
- **Success Factors:** [What enabled this]
- **Keep Strategy:** [How to maintain/enhance]

### Issues Identified

#### Issue 1: [Brief description]
- **Category:** [User-identified / Self-discovered]
- **Related Event:** [Reference to specific interaction]
- **Root Cause Analysis:**
  - Why? [Answer]
  - Why? [Answer]
  - Root Cause: [Final factor]
- **Improvement Strategy:** [Specific action to take]

### Metrics
| Metric | Value | Notes |
|--------|-------|-------|
| Interactions | X | [Number of exchanges] |
| Corrections | Y | [Times user corrected me] |
| Clarifications | Z | [Times I asked for clarification] |
| Improvements | W | [Proactive suggestions adopted] |

### Key Learnings
- [Learning 1 with context]
- [Learning 2 with context]

### Consistency Check

#### Behavior vs Core Files
| Behavior Observed | Expected per File | Consistent? | Issue |
|-------------------|-------------------|-------------|-------|
| [What I did] | [SOUL.md/AGENTS.md says] | Yes/No | [If no, describe gap] |

#### Internal Consistency (AGENTS.md)
- **Conflicts Found:** [Yes/No]
  - [Details if any]

#### Internal Consistency (SOUL.md)
- **Conflicts Found:** [Yes/No]
  - [Details if any]

#### Cross-File Consistency
- **Conflicts Found:** [Yes/No]
  - [Details if any]

### Proposed Improvements (Pending Approval)
| Issue | Proposed Change | Target File |
|-------|-----------------|-------------|
| [Issue 1] | [Specific change] | [SOUL.md/AGENTS.md/etc] |

## After Receiving User Feedback

When Nick approves specific changes:

1. **Read** `SOUL.md` and `AGENTS.md` completely
2. **Check** if new content contradicts existing principles
3. **Maintain** coherence — all principles should work together
4. **Be concise** — remove redundancy during updates
5. **Ask** Nick if you find inconsistencies

**Red flags for inconsistency:**
- Two principles suggesting opposite actions in the same scenario
- New advice that contradicts established boundaries
- Values that can't coexist logically

**Never remove principles without explicit permission.**

## Memory Maintenance

Periodically (every few days), use a heartbeat or dedicated session to:

1. Read through recent `memory/YYYY-MM-DD.md` files
2. Look at the "Self Reflection" sections for patterns
3. Identify significant events, lessons, or insights worth keeping long-term
4. Extract enduring lessons that should inform core behavior
5. Update MEMORY.md with distilled wisdom
6. Remove outdated info from MEMORY.md that's no longer relevant

Think of it like a human reviewing their journal and updating their mental model. Daily files are raw notes; MEMORY.md is curated wisdom.

## Core File Change Policy

**Strict Rule:** During reflection, NEVER modify:
- SOUL.md
- AGENTS.md
- SKILL.md
- Any other core behavior files

**Allowed Actions:**
- Write to memory/ files only
- Propose changes in reflection output
- Wait for user approval

**After User Approval:**
1. User explicitly states which changes to implement
2. Make the specific changes
3. Verify consistency with existing principles
4. Confirm completion

## Quality Principles

### Be Specific
❌ Vague: "Did good today"  
✅ Specific: "Refactored the database layer reducing query time by 40%"

❌ Vague: "Made mistakes"  
✅ Specific: "Forgot to handle null values in the user input validation"

### Seek Objective Evidence
- Count actual corrections: "Nick corrected me 3 times about X"
- Note specific feedback quotes
- Track measurable metrics

### Focus on Actionable Improvements
Each issue should lead to a concrete change:
- "Communication was unclear" → "Next time, start with a summary before details"
- "Missed edge case" → "Add verification step for user inputs"

### Be Concise
- Target 5-10 minutes of focused writing
- Omit obvious observations
- Focus on high-impact insights

## Cron Job Configuration

Add to cron jobs for automated reflection:

```json
{
  "name": "Daily Self Reflection",
  "schedule": {"kind": "cron", "expr": "0 2 * * *", "tz": "Asia/Shanghai"},
  "payload": {
    "kind": "agentTurn",
    "message": "Use self-reflection skill. Generate daily reflection for yesterday. Read memory/YYYY-MM-DD.md and any memory/YYYY-MM-DD-{slug}.md files for yesterday's date. Analyze events, identify issues, perform root cause analysis, and append/update the 'Self Reflection' section in memory/YYYY-MM-DD.md. Do NOT modify core files.",
    "model": "kimi-coding/k2p5",
    "thinking": "low"
  },
  "sessionTarget": "isolated",
  "delivery": {"mode": "announce"}
}

References

See references/framework.md for detailed analysis frameworks, root cause methods, and quality checklists.