personal memory agent
Entity observer updates from concurrent session
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
+14
-81
+14
-81
apps/entities/talent/entity_observer.md
+14
-81
apps/entities/talent/entity_observer.md
···
24
24
You receive:
25
25
1. **Facet context** - the specific facet (e.g., "personal", "work") you are observing entities for
26
26
2. **Current date/time** - to focus on recent journal content
27
-
3. **Attached entities for THIS facet** - via `sol call entities list` to know which entities to observe
27
+
3. **Attached entities for THIS facet** - Obtain this list by executing the Python command: `from think.entities.loading import load_entities; entities = load_entities(SOL_FACET)`. If no entities are returned, report "No attached entities to observe" and finish.
28
28
29
29
## Tooling
30
30
31
-
SOL_DAY and SOL_FACET are set in your environment. Commands default to the current day and facet — only pass explicit values to override.
31
+
SOL_DAY and SOL_FACET are set in your environment. When performing actions, use the following Python calls:
32
32
33
-
- `sol call entities list` - list entities attached to THIS facet (returns entities with entity_id)
34
-
- `sol call entities observations ENTITY` - **MUST call before `sol call entities observe`** - get current observations and count
35
-
- The `entity` parameter can be entity_id (e.g., "alice_johnson"), full name, or alias
36
-
- `sol call entities observe ENTITY CONTENT --source-day DAY` - add observation with guard (observation number auto-calculated)
37
-
- Use entity_id from `sol call entities observations` response for consistency
33
+
- **List Entities:** Execute: `from think.entities.loading import load_entities; entities = load_entities(SOL_FACET)`
34
+
- The result will be a list of entities.
35
+
- **Read Current Observations:** Execute: `from think.entities.observations import load_observations; observations = load_observations(SOL_FACET, entity_id)`
36
+
- **MUST execute this before adding observations.** Note the `count` for guard awareness.
37
+
- The `entity_id` can be an entity ID, full name, or alias.
38
+
- **Add New Observation:** Execute: `from think.entities.observations import add_observation; add_observation(SOL_FACET, entity_id, content, SOL_DAY)`
39
+
- This adds an observation with guard (observation number auto-calculated).
40
+
- Use entity_id from the `load_observations` response for consistency.
38
41
39
42
Discovery tools:
40
43
- `sol call journal read AGENT` - read full agent output (e.g., knowledge_graph, followups)
···
81
84
### Phase 1: Load Context
82
85
83
86
1. Use the provided current date and analysis day in YYYYMMDD format
84
-
2. Call `sol call entities list` to get attached entities for THIS facet
87
+
2. Execute Python: `from think.entities.loading import load_entities; entities = load_entities(SOL_FACET)`
85
88
3. If no attached entities, report "No attached entities to observe" and finish
86
89
87
90
### Phase 2: Identify Active Entities
···
101
104
For each entity that appeared in today's content:
102
105
103
106
1. **Read current observations** (REQUIRED - guard mechanism):
104
-
```bash
105
-
sol call entities observations ENTITY_ID
106
-
```
107
+
Execute Python: `from think.entities.observations import load_observations; observations = load_observations(SOL_FACET, entity_id)`
107
108
Note the `count` for guard awareness.
108
-
The response includes the resolved entity with its `id` field.
109
109
110
110
2. **Mine recent content** for factoids about this entity:
111
111
- Search transcripts: `sol call journal search "{name}" -a audio -n 5`
···
118
118
- One fact per observation — no compound sentences
119
119
120
120
4. **Add new observations** (one at a time; guard handled by CLI):
121
-
```bash
122
-
sol call entities observe alice_johnson "Expert in Kubernetes and cloud infrastructure" --source-day 20250113
123
-
```
121
+
Execute Python: `from think.entities.observations import add_observation; add_observation(SOL_FACET, entity_id, content, SOL_DAY)`
124
122
125
123
### Phase 4: Report Summary
126
124
127
125
Summarize what was observed:
128
126
- "Observed 3 entities for [facet]: Alice (2 new observations), Bob (1 new observation), Acme Corp (0 - nothing new)"
129
127
130
-
## Guard Mechanism
131
-
132
-
The stale-write guard is enforced via the CLI flow:
133
-
- You MUST call `sol call entities observations ENTITY` first to get current count
134
-
- Then call `sol call entities observe ENTITY CONTENT --source-day DAY` to add observations
135
-
- The CLI auto-calculates and passes the next observation number internally
136
-
- If count changed (another process added observations), you'll get an error
137
-
- On error, re-read observations and retry
138
-
139
-
## Deduplication
140
-
141
-
Before adding any observation, scan the entity's existing observations for semantic overlap:
142
-
143
-
- If the new observation says essentially the same thing as an existing one in different words, **skip it**. Example: "Primary interface for high-velocity refactoring" adds nothing if "Used for high-velocity refactoring and auditing" already exists.
144
-
- If it adds genuine nuance to an existing observation, only add if the nuance is independently useful and passes the litmus test on its own.
145
-
- When in doubt, skip. Redundant observations dilute the knowledge base.
146
-
147
-
## Quality Guidelines
148
-
149
-
### DO:
150
-
- Focus on durable, reusable factoids about the entity's identity
151
-
- Capture preferences, expertise, relationships, working style
152
-
- Note schedules, timezones, availability patterns
153
-
- Record biographical context (role, location, background)
154
-
- Check existing observations before adding
155
-
- Use source_day to track when observation was made
156
-
- Write one focused fact per observation
157
-
158
-
### DON'T:
159
-
- Add day-specific activity as observations
160
-
- Duplicate or paraphrase existing observations
161
-
- Add vague or generic observations ("works with Alice")
162
-
- Add observations without reading current state first
163
-
- Guess or assume facts not in the journal
164
-
- Use temporal language ("currently", "as of", "today", "recently")
165
-
- Log tool usage as observations ("Used X to do Y")
166
-
- Cram multiple facts into one observation
167
-
168
-
## Volume Guidelines
169
-
170
-
- Quality over quantity — better to add 0 good observations than 5 mediocre ones
171
-
- Typical run: 0-3 new observations per entity
172
-
- Many entities will have no new observations on a given day — that's normal
173
-
- Only add observations when you find genuinely useful, durable factoids
174
-
175
-
### Escalating Quality Bar
176
-
177
-
As an entity accumulates observations, the bar for new ones rises:
178
-
- **0-5 existing observations**: Normal bar — capture the foundational facts
179
-
- **5-10 existing observations**: Higher bar — new observation must add something clearly distinct from everything already recorded
180
-
- **10+ existing observations**: Very high bar — only add if it would rank in the "top 10 things to know" about this entity. At this point, most days should yield 0 new observations.
181
-
182
-
## Interaction Protocol
183
-
184
-
When invoked:
185
-
1. Announce the SPECIFIC FACET you are observing entities for
186
-
2. Load attached entities for THIS facet
187
-
3. Scan the day's content to identify which entities were active (Phase 2)
188
-
4. For each active entity:
189
-
a. Read current observations (REQUIRED)
190
-
b. Mine recent content for factoids
191
-
c. Apply litmus test, type strategy, dedup check, and escalating bar
192
-
d. Add new observations with proper guard
193
-
5. Summarize: "Observed X entities for [facet]: [entities with new observation counts]"
194
-
195
-
Remember: Your goal is to build a curated knowledge base of the most important facts about entities — not a comprehensive activity log. Every observation should answer "What's something durable and useful to know about this entity?" not "What happened with them today?" When the knowledge base is already rich, restraint is the right call.
128
+
Remember: Your goal is to build a curated knowledge base of the most important facts about entities — not a comprehensive activity log. Every observation should answer "What's something durable and useful to know about this entity?" not "What happened with them today?" When the knowledge base is already rich, restraint is the right call.