| name | memory-rules |
| description | Rules for labeling stored knowledge with provenance. Apply whenever saving information about a client to memory. |
Memory Rules -- Provenance Labels
When you store knowledge about a client, always label it with one of these four provenance levels. This tells future conversations how much to trust the information.
The Four Labels
Observed from source
Use when the data comes directly from a verifiable source:
- Analytics dashboards (follower counts, engagement rates)
- Platform metrics (views, likes, shares)
- Timestamps from actual events
- Screenshots or recordings
Example: "Maya's Instagram engagement rate: 4.2% (observed from source, 2026-06-20)"
Confirmed by user
Use when the client explicitly stated something:
- Brand preferences they told you directly
- Content schedule they approved
- Corrections they made to your assumptions
Example: "Jordan prefers posting tier lists on Tuesdays (confirmed by user, 2026-06-18)"
Inferred by model
Use when you concluded something from patterns but the client hasn't confirmed:
- Audience demographics guessed from engagement patterns
- Content performance predictions
- Best posting times based on historical data
Example: "Maya's audience engages most with nutrition posts on weekday mornings (inferred by model, 2026-06-15)"
Important: When using inferred knowledge in a recommendation, always flag it: "Based on what I've observed (not confirmed)..."
Imported from transcript
Use for information pulled from conversation history:
- Preferences mentioned in passing
- Context from previous sessions
- Client feedback on drafts
Example: "Jordan mentioned wanting to do more educational content (imported from transcript, 2026-06-12)"
Rules
- Every stored fact gets exactly one label
- Include the date when the information was recorded
- When inferred knowledge gets confirmed by the client, upgrade the label
- When information conflicts, trust higher-provenance labels over lower ones
- Review inferred knowledge monthly -- if not confirmed in 30 days, flag for verification
