personal memory agent
Add speakers segment insight for meeting participant detection
New segment-level insight that analyzes audio + screen context to:
- Detect active video meetings (Zoom, Teams, Meet, etc.)
- Extract participant names from visible UI and conversation
- Handle audio-only imports via conversation pattern detection
- Support segments spanning two distinct meetings
Output structured for occurrence extraction with participant lists.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
+85
+8
think/insights/speakers.json
+8
think/insights/speakers.json
···
1
+
{
2
+
"title": "Meeting Speakers",
3
+
"description": "Detects video meetings in the segment and extracts participant names from screen and conversation.",
4
+
"frequency": "segment",
5
+
"occurrences": "For each meeting detected, extract start/end times and the list of participant names. If segment spans two meetings, create separate occurrences.",
6
+
"contains": "Meeting detection with time bounds and participant name lists extracted from visible participant panels and conversation.",
7
+
"color": "#ff5722"
8
+
}
+77
think/insights/speakers.txt
+77
think/insights/speakers.txt
···
1
+
# Meeting Speaker Detection
2
+
3
+
## Objective
4
+
5
+
Identify meetings in this segment and extract participant names. This output feeds into occurrence extraction to populate meeting participant lists.
6
+
7
+
## Input Types
8
+
9
+
### With Screen Activity (Live Capture)
10
+
11
+
A meeting is detected when screen activity shows a video conferencing application:
12
+
- Zoom, Google Meet, Microsoft Teams, Webex, or similar
13
+
- Participant panels or video tiles visible
14
+
- Meeting controls or UI elements present
15
+
16
+
Extract participant names from:
17
+
1. **Visible participant list/panel** on screen (highest confidence)
18
+
2. **Names spoken in conversation** - direct address ("Thanks, Sarah"), mentions ("John was saying...")
19
+
3. **Self-introductions** ("Hi, I'm Alex from...")
20
+
21
+
### Audio Only (Import)
22
+
23
+
When no screen activity is present, a meeting is detected when audio shows:
24
+
- Multiple speakers with conversational turn-taking
25
+
- Meeting-style patterns: agenda items, discussion, decisions, action items
26
+
- Professional/work context evident from topics
27
+
28
+
Extract names from:
29
+
1. **Names spoken in conversation** - direct address and mentions
30
+
2. **Self-introductions**
31
+
3. **Setting/topics metadata** if available from transcription
32
+
33
+
## Boundary Detection
34
+
35
+
If the segment spans two distinct meetings (one ending, another starting):
36
+
- Look for farewell patterns followed by new greetings
37
+
- Different participant set appears
38
+
- Significant topic shift with new introductions
39
+
40
+
Report each meeting separately with its own time range and participants.
41
+
42
+
## Output Format
43
+
44
+
Structure your output clearly for occurrence extraction:
45
+
46
+
### Meeting Detection
47
+
48
+
State whether meeting(s) were detected and the detection basis:
49
+
- **Detected**: Yes / No
50
+
- **Source**: Screen (video call UI visible) / Audio-only (conversation patterns)
51
+
52
+
### Meeting 1
53
+
54
+
- **Time**: HH:MM:SS - HH:MM:SS
55
+
- **Platform**: Zoom / Meet / Teams / etc. (if visible, omit for audio-only)
56
+
- **Participants**:
57
+
- Name (evidence: where/how detected)
58
+
- Name (evidence: where/how detected)
59
+
60
+
### Meeting 2 (if segment contains two meetings)
61
+
62
+
Same structure as above.
63
+
64
+
## What to Ignore
65
+
66
+
- Podcasts, videos, or streaming content being watched
67
+
- Music or background audio
68
+
- Single-speaker recordings without meeting context
69
+
- Phone calls (no video conferencing UI)
70
+
71
+
## No Meeting Case
72
+
73
+
If no meeting is detected, output:
74
+
75
+
### Meeting Detection
76
+
- **Detected**: No
77
+
- **Reason**: [Brief explanation - e.g., "No video call UI visible, single speaker only"]