What are Traits?
Traits are structured attributes extracted from user conversations. Unlike unstructured memory, traits are:
- Typed: String, number, boolean, array, or enum
- Labeled: Each trait has a defined key and category
- Scored: Confidence level from 0.0 to 1.0
- Tracked: Source messages are recorded (optionally)
Trait Structure
interface Trait {
id: string;
profileId: string;
key: string; // e.g., "preferred_language"
category: string | null; // e.g., "communication"
valueType: "string" | "number" | "boolean" | "array" | "enum";
value: unknown;
confidence: number; // 0.0 - 1.0
source: "extracted" | "manual" | "inferred";
sourceMessageIds: string[];
createdAt: Date;
updatedAt: Date;
}
Default Traits
GetProfile includes these default traits out of the box:
| Trait | Type | Category | Description |
|---|
name | string | identity | User’s name or preferred name |
preferred_language | string | communication | Communication language |
communication_style | enum | communication | formal, casual, technical, simple |
detail_preference | enum | communication | brief, moderate, detailed |
expertise_level | enum | context | beginner, intermediate, advanced, expert |
timezone | string | context | User’s timezone |
interests | array | preferences | Topics the user is interested in |
current_goals | array | context | What the user is working toward |
Confidence Scores
Confidence represents how certain GetProfile is about a trait value:
| Score | Meaning | Example |
|---|
| 0.9+ | Explicit statement | ”My name is Alex” |
| 0.7-0.9 | Strong indication | Uses technical jargon consistently |
| 0.5-0.7 | Moderate inference | Seems to prefer detailed responses |
| < 0.5 | Weak inference | Might be interested in Python |
Traits with low confidence are still stored but may not be injected into
prompts. Configure the confidenceThreshold in your trait schema.
Trait Sources
| Source | Description |
|---|
extracted | Automatically extracted from conversations |
manual | Set directly via API |
inferred | Derived from other traits or patterns |
Trait Updates
When new information conflicts with existing traits:
- Higher confidence wins: New extraction with 0.8 confidence overrides existing 0.5
- Manual takes priority: Manually set traits are preserved unless explicitly updated
- Contradictions delete: If new info directly contradicts a trait, it’s deleted
Given this conversation:
User: Hi, I'm Alex and I've been coding in Python for about 5 years now.
I prefer when explanations are concise and technical.
[
{
"key": "name",
"value": "Alex",
"confidence": 0.95,
"action": "create"
},
{
"key": "expertise_level",
"value": "advanced",
"confidence": 0.75,
"action": "create"
},
{
"key": "communication_style",
"value": "technical",
"confidence": 0.85,
"action": "create"
},
{
"key": "detail_preference",
"value": "brief",
"confidence": 0.7,
"action": "create"
}
]
