| name | self-improvement |
| description | Stores user corrections and preferences as learnings to improve future responses. Use when the user corrects a mistake, states a preference, or says 'remember that', 'from now on', or 'I prefer'. |
| version | 1.0.0 |
| author | zoe-team |
| api_only | true |
| priority | 1 |
| triggers | ["remember that","don't forget","i prefer","from now on","always remember","never forget","i told you","don't do that again","i like it when","stop doing"] |
| allowed_endpoints | ["GET /api/learnings","POST /api/learnings/*/confirm","POST /api/learnings/*/dismiss"] |
| tags | ["meta","self-improvement"] |
Self-Improvement
When to Use
Runs in the background after every conversation. Activates explicitly when the user corrects Zoe, states a preference, or uses trigger phrases like "remember that", "from now on", "I prefer", or "don't do that again".
Workflow
- Detect correction signal — scan the user's message for corrections ("actually, I meant…"), preference statements ("I prefer…"), or explicit triggers ("remember that…")
- Extract the learning — identify what was wrong and what the correct behavior is, structured as a key-value pair:
- Category:
tone, formatting, factual, preference, workflow
- Before: what Zoe did wrong
- After: what the user wants instead
- Check for duplicates — use
GET /api/learnings before proposing a new learning
- Acknowledge to the user — respond with a brief acknowledgment (e.g., "Got it, I'll remember that for next time"); the self-learning pipeline handles actual persistence
- Apply in future — before generating responses, retrieve relevant learnings from
GET /api/learnings and adjust behavior accordingly
Example
User: "Actually, I prefer bullet points over paragraphs when you summarize things."
Steps:
- Detect preference signal ("I prefer")
- Extract: category=
formatting, before=paragraph summaries, after=bullet point summaries
- Let the self-learning pipeline persist the learning with user ownership; this skill only uses the listed review endpoints
- Respond: "Noted — I'll use bullet points for summaries from now on."
Security
- Only trusted sources can create learnings (Trust Gate integration)
- Owner corrections auto-confirm via
POST /api/learnings/{id}/confirm
- Trusted contact corrections require user review before confirming
- Learnings expire after 30 days without confirmation — dismiss stale ones via
POST /api/learnings/{id}/dismiss
