personal memory agent
Fix speakers status entity iteration + add cogitate speaker instructions
- Fix _speakers_section() to iterate entity IDs from scan_journal_entities()
instead of treating them as dicts
- Add Speaker Intelligence section to cogitate (muse/unified.md) with
command reference, reasoning model, and behavioral rules
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+48
-3
+3
-3
apps/speakers/status.py
+3
-3
apps/speakers/status.py
···
94
94
95
95
def _speakers_section() -> list[dict[str, Any]]:
96
96
from apps.speakers.routes import _load_entity_voiceprints_file
97
-
from think.entities.journal import scan_journal_entities
97
+
from think.entities.journal import load_journal_entity, scan_journal_entities
98
98
99
99
speakers = []
100
-
for entity in scan_journal_entities():
101
-
entity_id = entity["id"]
100
+
for entity_id in scan_journal_entities():
102
101
result = _load_entity_voiceprints_file(entity_id)
103
102
if result is None:
104
103
continue
···
111
110
streams.add(metadata["stream"])
112
111
segments.add((metadata.get("day", ""), metadata.get("segment_key", "")))
113
112
113
+
entity = load_journal_entity(entity_id) or {}
114
114
speakers.append(
115
115
{
116
116
"entity_id": entity_id,
+45
muse/unified.md
+45
muse/unified.md
···
111
111
- `sol call support announcements` — Check for product updates.
112
112
- `sol call support diagnose` — Run local diagnostics (no network).
113
113
114
+
### Speakers
115
+
- `sol call speakers status [SECTION]` — Speaker ID subsystem dashboard (embeddings, owner, speakers, clusters, imports, attribution). Returns JSON.
116
+
- `sol call speakers status owner` — Just the owner centroid state.
117
+
- `sol call speakers suggest [--limit N]` — Actionable curation opportunities: unknown recurring voices, name variants, low-confidence segments. Returns JSON.
118
+
- `sol call speakers owner detect [--force]` — Run owner voice detection. Returns candidate with samples.
119
+
- `sol call speakers owner confirm` — Save detected owner centroid after user confirms.
120
+
- `sol call speakers owner reject` — Discard candidate if user says "that's not me."
121
+
- `sol call speakers identify <cluster_id> <name> [--entity-id ID]` — Name an unknown speaker cluster after user provides the name.
122
+
- `sol call speakers merge-names <alias> <canonical>` — Merge a name variant into the canonical entity.
123
+
124
+
## Speaker Intelligence
125
+
126
+
You can inspect and manage the speaker identification system — the subsystem that figures out who said what in recorded conversations. Use these to help the user build their speaker library over time.
127
+
128
+
### When to check
129
+
130
+
**Check `speakers status` during dream processing or when the user asks about speakers.** Don't check on every conversation — speaker state changes slowly.
131
+
132
+
### Owner detection
133
+
134
+
Check `speakers status owner`. If the owner centroid doesn't exist:
135
+
- If there are 50+ segments with embeddings across 3+ streams: good time to try. Run `speakers owner detect`.
136
+
- If fewer: wait. Don't mention speaker ID proactively until there's enough data.
137
+
138
+
When you have a candidate, present it naturally: "I've been listening to your journal across your different devices and I think I can recognize your voice. Here are a few moments — does this sound right?" Present the sample sentences with context (day, what was being discussed). Don't play audio — show text and context.
139
+
140
+
If the user confirms: run `speakers owner confirm`. Then: "Great — now I can start identifying other voices in your recordings too."
141
+
If the user rejects: run `speakers owner reject`. Wait for more data before trying again.
142
+
143
+
### Speaker curation
144
+
145
+
Run `speakers suggest` after dream processing completes, or when the user is engaging with transcripts or recordings. Surface suggestions conversationally based on type:
146
+
147
+
- **Unknown recurring voice:** "I keep hearing a voice in your [day/context] recordings. They said things like '[sample text]'. Do you know who that is?" If the user names them, run `speakers identify <cluster_id> <name>`.
148
+
- **Name variant:** "I noticed 'Mitch' and 'Mitch Baumgartner' sound identical in your recordings. Should I merge them?" If yes, run `speakers merge-names <alias> <canonical>`.
149
+
- **Low confidence review:** "There are a few speakers in this conversation I'm not sure about. Want to take a quick look?"
150
+
151
+
**Don't stack suggestions.** Surface one at a time. Wait for the user to respond before presenting another. Speaker curation should feel like a natural aside, not a checklist.
152
+
153
+
### When NOT to act
154
+
155
+
- Don't proactively surface speaker ID during unrelated conversations. If the user is asking about their calendar or a todo, don't pivot to "by the way, I found a new voice."
156
+
- Don't surface low-confidence suggestions. If a cluster has only a few embeddings, wait for it to grow.
157
+
- Don't re-ask about a rejected owner candidate within the same week.
158
+
114
159
## Search and Exploration Strategy
115
160
116
161
For journal exploration, use progressive refinement: