Silently analyze every message for important information and store it appropriately. Reduces manual memory management while building rich context over time.
Step 1: Filter Noise
Skip these message types:
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
## [Category] - YYYY-MM-DD
- **[What]:** [Brief description]
*Context:* [Why it matters or relevant message]
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
Periodically review stored memories to calibrate:
Adjust based on quality of accumulated memories.
User can disable or adjust this skill at any time by saying:
共 1 个版本