Intelligent Intent Validation™ v2.1
Every memory is classified into an intent category before storage. This classification is mandatory, always active, and cannot be disabled.
On this page
Intent Categories
The Intent Engine classifies every memory into one of six categories:
| Category | Description | Example |
|---|---|---|
| preference | User preferences and settings | "I prefer dark mode" |
| fact | Factual information and knowledge | "The API rate limit is 100 req/s" |
| procedural | Step-by-step instructions | "To deploy, run git push main" |
| bio | Biographical and identity data | "My name is Alex, I’m a data engineer" |
| ephemeral | Temporary or session-scoped data | "Current task: review PR #42" |
| critical | Safety-critical information (auto-promoted) | "User has a peanut allergy" |
Classification Process
Classification follows a two-step process:
Step 1: Keyword Heuristics
The content is scanned against a curated keyword registry for each intent category. Matches are scored based on keyword frequency and category specificity. The top-scoring category becomes the initial classification.
Step 2: Confidence Gate
The initial classification must exceed a confidence threshold to be accepted. If confidence is below the gate, the classification may be adjusted or flagged with a warning. The confidence score is included in the store result.
Critical Promotion
A separate set of critical keywords (safety, emergency, allergy, medication, and similar terms) triggers automatic promotion to the criticalcategory. This promotion overrides the normal classification process.
Warning
Confidence Gate
The confidence gate is a calibrated threshold that determines classification acceptance. The gate incorporates:
- Agent-reported confidence (passed at store time)
- Keyword saturation depth (number of distinct keyword hits)
- Confidence boost from critical keyword detection
- Self-healing adjustments when classifications conflict
The specific threshold values and boost factors are proprietary.
API Surface
Intent classification is exposed in the StoreResult:
{
"validation_details": {
"final_intent": "preference",
"source_type": "validated",
"confidence": 0.92,
"confidence_boost": 0.0,
"is_critical": false,
"is_valid": true,
"warning": null,
"self_healing_applied": false
}
}Constraints
- Intent validation is always active — it cannot be disabled via configuration
- Classification is based on keyword heuristics — it is not an LLM-based classifier
- The keyword registry and scoring weights are not publicly documented
- Self-healing adjustments are logged but details are not exposed to the caller