Core Concepts

Retrieval

Databaset uses semantic search: find memories by meaning, not keywords.

Semantic vs keyword

| Query | Keyword match | Semantic match | |-------|---------------|----------------| | budget | "budget" only | "User has 50k to spend" | | language | "language" only | "User speaks Hindi" |

Ranking factors

1. Similarity score: cosine distance in embedding space
2. Recency: newer memories weighted higher
3. Memory type: preferences ranked higher for preference queries
4. Contradictions: newer facts replace older conflicting ones

Parameters

await memory.recall({
  userId: 'user_123',
  query: 'what are their preferences?',
  limit: 5,        // max results
  minScore: 0.7,   // minimum similarity threshold
})

Output formats

• memory.recall() returns a formatted string for prompts
• memory.recallRaw() returns an array with scores and metadata