Auto Memory Curation Skill
Overview
Silently analyze every message for important information and store it appropriately. Reduces manual memory management while building rich context over time.
How It Works
Trigger
- Runs on every message (silently)
- No user activation needed
Analysis Pipeline
Step 1: Filter Noise Skip these message types:
- Greetings (hi, hello, hey)
- Thanks/thanking
- Acknowledgments (ok, sure, yes, yeah)
- Questions without context
- Single word responses
- Bot commands
Step 2: Categorize For non-noise messages, categorize:
| Category | What to look for | Store in |
|---|---|---|
| Fact | New info about user, projects, preferences | MEMORY.md |
| Decision | Choices made, conclusions reached | memory/YYYY-MM-DD.md |
| Preference | Likes, dislikes, style preferences | USER.md |
| Idea | Random thoughts, inspiration, concepts | memory/topics/ideas.md |
| Learning | Lessons, insights, discoveries | memory/topics/lessons.md |
| Project | Project updates, progress, blockers | memory/projects/ |
| Goal | Goals, targets, milestones | memory/topics/goals.md |
| Error | Mistakes, corrections to avoid | memory/topics/anti-patterns.md |
| Commitment | Promises I make, tasks to do | tasks.md |
Step 3: Extract & Store
- Extract the key information
- Add timestamp reference
- Store in appropriate file
- Use append mode (never overwrite)
Guidelines
What to Capture
- New facts about Vini (name, preferences, goals)
- Project updates or decisions
- Ideas for future projects
- Learning insights
- Corrections (what doesn't work)
- Commitments I make
What to SKIP
- Passwords, secrets, API keys
- Trivial acknowledgments
- Basic confirmations
- Questions I'm asking
- Technical errors that are fixed
Quality Rules
- Be selective - Don't store everything
- Be concise - One sentence per memory
- Be contextual - Include enough info to understand later
- Be accurate - Don't paraphrase incorrectly
- Never duplicate - Check if already stored
Format
## [Category] - YYYY-MM-DD
- **[What]:** [Brief description]
*Context:* [Why it matters or relevant message]
Examples
User says:
"I prefer concise messages, no filler words"
Stored in USER.md:
Preference - 2026-03-06
- Communication: Prefers concise messages, no filler words
User says:
"Let's build a landing page for the consultancy inspired by Nexus AI"
Stored in memory/2026-03-06.md:
Decision - 2026-03-06
- Consultancy: Will use Nexus AI style for landing page
User says:
"I realized I work better with visual examples first, then theory"
Stored in MEMORY.md:
Learning - 2026-03-06
- Learning Style: Visual examples first, theory after
Testing
Periodically review stored memories to calibrate:
- Am I capturing too much noise?
- Am I missing important things?
- Are categories correct?
Adjust based on quality of accumulated memories.
Override
User can disable or adjust this skill at any time by saying:
- "Don't store that"
- "Clear recent memories"
- "Adjust memory curation"