insights

name	insights
description	Generate higher-order insights from the memories table — patterns, contradictions, blind spots Jay hasn't articulated. Used by /insights slash command.
triggers	/insights, generate insights, what patterns, what should I know

Insights

V3 transcript ch.14 — Gemini scans memories and surfaces things Jay doesn't know about himself or his business.

When to use

• /insights slash command (manual)
• Weekly cron mission weekly_insights
• After a Suggestions run flags an unusual pattern

Inputs

• period: 7d | 30d | 90d (default 30d)
• agent: optional — scope to one agent's memories
• min_confidence: float 0–1 (default 0.6)

Behavior

1. Pull memories from SQLite where created_at >= now - period
2. Send to Gemini 2.5 Flash with prompt:

   "From these memories, emit JSON of insights: {observation, confidence 0-1, evidence_memory_ids[]}. Look for contradictions, recurring blockers, taste shifts, productive tensions. Skip the obvious."

3. Filter confidence >= min_confidence
4. Upsert into insights table
5. Render top 5 to user

Output format

INSIGHTS — last {period}
━━━━━━━━━━━━━━━━━━━━━
1. {observation}
   confidence: {0.XX} · evidence: {n memories}
2. ...

Tools used

• SQLite read on memories and insights tables
• Gemini 2.5 Flash (cheap workhorse)

Cost notes

• ~$0.01 per /insights invocation (1 Flash call, 30 days of memories ≈ 50K tokens)

See also

• memory/extractor.py — populates the memories this skill reads
• skills/suggestions/SKILL.md — agent-overburden detector, sibling skill