Concepts

A concept is a label for a cluster of related memories. Concepts give a query another way to narrow semantic results without moving or changing the memories.

From memories to concepts

A concept labels a group of related memories. The memories stay where they are, and the clustering is deterministic.

The consolidation job clusters embeddings with deterministic k-means. It records which memories belong to each concept. With the same inputs and settings, clustering produces the same groups.

Filter by concept

Pass concept IDs through profile.conceptFilter. The filter applies to semantic search only. Concepts do not change the topic tree or the memory lifecycle.

The dream job creates concepts. The default settings are clusterK: 8 and clusterMinSize: 3. The model used for clustering must produce 384-dimensional embeddings.