Retrieval Tuning

Optimize memory retrieval for your use case.

Relevance Threshold

Control result quality vs. quantity:

{
  "context": "user preferences",
  "relevance_threshold": 0.7
}

Threshold	Results
0.3-0.5	More results, lower precision
0.5-0.7	Balanced
0.7-0.9	Fewer results, higher precision

Max Memories

Limit the number of results:

{
  "context": "...",
  "max_memories": 5
}

Use fewer memories for:

• Quick lookups
• Specific questions

Use more memories for:

• Comprehensive answers
• Context building

Search Depth

Control how thoroughly the system searches for relevant memories:

{
  "context": "...",
  "search_depth": 0.7
}

Depth	Search Coverage	Best For
0.0-0.3	Fast (5-15% of memories)	Simple queries, large datasets, latency-sensitive apps
0.3-0.7	Balanced (15-50% of memories)	Most use cases, good accuracy/speed tradeoff
0.7-1.0	Thorough (50-100% of memories)	Critical queries, high accuracy requirements

Trade-offs:

• Higher depth = More accurate, slower, higher cost
• Lower depth = Faster, cheaper, may miss some relevant memories

When to increase search depth:

• Critical business decisions
• Complex or nuanced queries
• When initial results aren't satisfactory
• Small to medium memory stores (< 10,000 memories)

When to decrease search depth:

• Simple factual lookups
• Large memory stores (> 100,000 memories)
• Latency-sensitive applications
• High-volume query scenarios

Query Phrasing

How you phrase queries affects results:

Query Style	Best For
Questions	"What does the user prefer?"
Keywords	"user preferences dark mode"
Statements	"Finding information about user preferences"

Metadata Filtering

Pre-filter memories by metadata:

{
  "context": "preferences",
  "metadata_filter": {
    "category": "preference"
  }
}

Tips

1. Be specific - Detailed queries get better results
2. Use context - Include relevant background information
3. Test thresholds - Find the right balance for your data
4. Monitor relevance - Check relevance_score in responses