V3 Memory Unification

Read all entries from the SQLite backend, generate an embedding for each entry's content, then store the entry with its embedding in AgentDB. The same pattern applies to Markdown files by reading file content and preserving the original path in metadata.

HNSW (Hierarchical Navigable Small World) is an approximate nearest-neighbor index that accelerates semantic similarity search over embeddings. This Skill configures it with parameters like efConstruction 200 and M 16 to target large search speedups over brute-force queries.

Yes, the unified memory service is designed for real-time cross-agent memory synchronization. Once legacy per-agent stores like SwarmMemory are migrated, all agents query the same AgentDB backend through the shared interface.

Learning patterns are stored as memory entries whose metadata includes the SONA mode, reward, and adaptation time, along with a content embedding. Similar patterns are retrieved through semantic queries filtered by the learning pattern type.

HNSW trades exact results for speed, so approximate search may miss some true nearest neighbors depending on index parameters. It also requires generating embeddings for every stored entry, adding upfront migration and storage cost.