personal memory agent
talent(sense): tighten entity extraction guardrails
Backfill spot-check on 2025 gap-zone segments surfaced five recurring
edge cases the prompt rules implied but didn't explicitly enforce.
Hardening before a planned ~67k-segment historical backfill:
- Person: strengthen first-name-only skip; explicitly ban generic
speaker labels ("Speaker 1", "Colleague") from entities (they belong
in the speakers array when meeting_detected=true)
- Project: exclude generic git/file identifiers ("main", "dev",
"staging") and one-word lowercase tokens
- speakers ↔ meeting_detected consistency: empty speakers + true
meeting_detected is invalid; use generic labels rather than emptying
- New rule 7: skip placeholder names with speaker-uncertainty markers
("Museum something", "whatever-thing-it's-called")
- New rule 8: summary-mentioned named entities must appear in entities
No structural / schema changes. Only adds explicit guardrails the model
already implicitly violated under the prior wording.
+7
-3
+7
-3
talent/sense.md
+7
-3
talent/sense.md
···
61
61
62
62
### entities
63
63
Extract ALL named entities mentioned in the content. Be thorough — extract every entity you can identify, not just the most prominent ones. Four types only:
64
-
- **Person**: Individual people by name. Prefer full names. Consolidate variants ("JB" + "John Borthwick" → one entity "John Borthwick"). Skip ambiguous first-name-only references. Include historical figures, authors, scientists, politicians — anyone mentioned by name.
64
+
- **Person**: Individual people by name. Prefer full names. Consolidate variants ("JB" + "John Borthwick" → one entity "John Borthwick"). ALWAYS skip first-name-only references unless the same segment locks the identity with surrounding context (role, organization, or full-name introduction). NEVER include generic speaker labels like "Speaker 1", "Speaker 2", "Colleague", "Person A" — these belong only in the `speakers` array when `meeting_detected=true`. Include historical figures, authors, scientists, politicians — anyone mentioned by full name.
65
65
- **Company**: Businesses and organizations. Include companies, government agencies (NASA, NOAA), universities, media outlets.
66
-
- **Project**: Named projects, products, or codebases. Include missions (OSIRIS-REx), initiatives, specific product models.
66
+
- **Project**: Named projects, products, or codebases. Include missions (OSIRIS-REx), initiatives, specific product models. EXCLUDE generic git/file identifiers ("main", "dev", "staging", "src", "tmp"), file extensions, path components, and one-word lowercase tokens that are likely branch or directory names rather than named projects.
67
67
- **Tool**: Software applications and services. Include websites (Fox News, Wikipedia, Amazon), browser extensions, developer tools, hardware products mentioned by name.
68
68
69
69
**For screen content specifically:** Extract entities from visible text in screen descriptions — article headlines, page titles, product names, people mentioned in articles, organizations referenced. If the user is browsing a website about the Renaissance, extract the specific historical figures, art movements, and institutions mentioned.
···
107
107
108
108
Prefer complete canonical forms (full names when identifiable). Do NOT include the journal owner's name. Return `[]` if no meeting or no names identified.
109
109
110
+
**Consistency rule:** If `meeting_detected=true`, this array must have at least one entry. If you cannot identify any names, use generic labels ("Speaker 1", "Speaker 2", "Colleague A") rather than emptying the array — an empty `speakers` array with `meeting_detected=true` is invalid.
111
+
110
112
### recommend
111
113
Processing recommendations for downstream agents:
112
114
- **screen_record**: `true` if density is "active" AND there is meaningful screen content worth documenting (not just a static/repetitive screen)
···
127
129
## Rules
128
130
129
131
1. Every field is required. Never omit a field.
130
-
2. `entities` and `speakers` may be empty arrays `[]`.
132
+
2. `entities` and `speakers` may be empty arrays `[]` (subject to rule 8 for speakers when `meeting_detected=true`).
131
133
3. `facets` always has at least one entry — the closest configured facet for the activity. Empty array is not allowed.
132
134
4. Be precise with density — misclassifying active segments as idle is the worst error.
133
135
5. For `content_type`, choose the single best match — the dominant activity in the segment. If two activities are roughly equal, pick the one with more durable continuation evidence (entities, repeated screen content); the `facets[]` array's `level` field already encodes secondary activity.
134
136
6. Activity summary must describe observable actions, not inferred states.
137
+
7. Skip entities whose name contains a speaker-uncertainty placeholder. If the transcript says "a game called Museum something" or "the new whatever-thing-it's-called", the speaker is signaling they don't know the actual name — do not extract a placeholder name as an entity.
138
+
8. If `meeting_detected=true`, `speakers` must contain at least one entry (use generic labels if no names are identifiable). If `activity_summary` mentions specific named people, projects, or tools, those names should also appear in `entities` (subject to the per-type rules above) — don't reference an entity by name in the summary and then omit it from `entities`.
135
139
136
140
Return ONLY the JSON object, no other text or explanation.