User Context Scanner
Let USER.md evolve from static file to living profile.
Three-Confirmation Rule
1st mention → Potential preference (no guard trigger)
2nd mention → Initial confirmation (starts influencing)
3rd mention → Stable value (triggers value guard)
Why 3 times: Single mentions may be emotional reactions. 3 times = true pattern.
Sources
Direct (User Said)
- "I'm a product manager" → extract: profession
- "I don't work on weekends" → extract: rest habit
- "My wife's name is Fang" → extract: family
Indirect (AI Inferred)
- 3 consecutive late-night sessions → infer: night owl
- Repeated discussion of one topic → infer: current focus
Value Tag Mapping
| User Says | Value Tag |
|---|---|
| Health first / exercise | health_priority |
| Family most important | family_first |
| Long-term / no quick money | long_term |
| Quality over compromise | quality_first |
| Efficiency first | efficiency_first |
After 3 confirmations, automatically added to value-guard detection.
Update Mechanism
Immediate (Strong Signal)
User explicitly states preference → write immediately:
User: "I don't work weekend mornings"
→ Mark: rest_hours × 1 (needs 2 more to upgrade)
Accumulation Upgrade
1st → Mark: potential_preference × 1
2nd → Mark: potential_preference × 2 (initial confirmation)
3rd → Upgrade to stable value, trigger value-guard
Profile Fields
USER.md dynamic fields:
- profession/role
- rest habits
- family role
- current focus areas
- decision style
- communication preference
- emotional patterns
- taboo/bottom lines (after 3 confirmations)
Call soul_context
Tool: soul_context
Action: add or merge
Params:
key: <field name>
value: <value>
Important
- User's direct words > AI inference
- Slow is better than wrong
- 3 confirmations is protection, not restriction
- Profiles serve understanding, not data collection
Tags for publishing: soul, system, user-profile, context, understanding