Skip to content

Multi-Dimensional Retrieval

Semantic (Base Layer)

  • Vector similarity search
  • Over-fetches initially for refinement
  • Uses Vectorize with cosine similarity

Temporal Filtering

We use exponential decay with exponential boost algorithm and multiple parameters to tune the experience of temporal features based on access count, lastAccessed Date and per agent configurations.

// Exponential decay with frequency boost
function calculateRecencyScore(accessCount, lastAccessed, config) {
  const hoursSinceAccess = (now - lastAccessed) / (1000 * 60 * 60);
  const timeDecay = Math.exp(-hoursSinceAccess / config.halfLifeHours);
  const frequencyBoost = Math.log10(accessCount + 1);
  return combinedScore; // 0-100 scale
}

Presets

  • Balanced: 7-day half-life , suitable for balanced approach to decay.
  • Aggressive: 3-day half-life suitable for aggresive decay think ( chatbots, short-context window)
  • Long-term: 30-day half-life long term decay, slow update of temporal feature ( think like codebases, AI assisted coding ).

Contextual Filtering

  • Developer-supplied context (task, environment)
  • Included in response if semantic match
  • No auto-generation (future feature)

Behavioral Filtering

// Wilson score confidence interval
function wilsonScore(success, failure, confidence = 0.95) {
  // Calculates confidence-based success rate
  return score; // 0-1 scale
}

Zero-sum mechanism: All memories in response share success/failure 

## 7. `configuration.md`
```markdown
# Configuration Guide

## Agent Creation
```javascript
const agentConfig = {
  // Temporal features
  halfLifeHours: 168,      // 7 days
  timeWeight: 0.6,
  frequencyWeight: 0.4,
  decayCurve: 'hybrid',
  decayFloor: 0.15,

  // Behavioral features  
  successRate: 0.0,        // Start at 0 for passive learning
  stabilityThreshold: 0.0   // Start at 0 for passive learning
};

Preset Configurations

Balanced (Default)

{
  halfLifeHours: 168,      // 7 days
  timeWeight: 0.6,
  frequencyWeight: 0.4,
  decayCurve: 'hybrid',
  decayFloor: 0.15
}

Aggressive Decay

{
  halfLifeHours: 72,       // 3 days  
  timeWeight: 0.7,
  frequencyWeight: 0.3,
  decayCurve: 'exponential'
}

Long-Term Memory

{
  halfLifeHours: 720,      // 30 days
  timeWeight: 0.5, 
  frequencyWeight: 0.5,
  decayCurve: 'polynomial'
}

Important Notes

  • Choose carefully: Parameters affect memory filtering significantly
  • Avoid changes: Don't modify after agent creation
  • Eventual consistency: Config changes apply gradually
  • Monitor: Use Admin UI to observe filtering effects

Behavioral Thresholds

  • Start with successRate: 0.0 to avoid false positives
  • Gradually increase as agent matures (e.g., 0.5, 0.7)
  • Use Wilson score for confidence-based filtering