agent-v3-memory-specialist

Implement a UnifiedMemoryService that wraps an AgentDB adapter behind a common IMemoryBackend interface, then migrate each legacy backend (SQLite, Markdown, distributed stores) in phases while keeping a backward compatibility layer during transition.

HNSW (Hierarchical Navigable Small World) is an approximate nearest-neighbor index that turns O(n) linear memory scans into O(log n) vector searches. This skill targets 150x-12,500x search improvements and sub-100ms queries on datasets over one million entries.

Extract rows from the existing memory_entries table, generate embeddings for each content field, and insert them into AgentDB with the original metadata preserved. The skill provides a phased migration plan covering SQLite and Markdown sources.

Yes. The unified query interface routes semantic queries through the HNSW vector index with similarity thresholds, while structured queries use filters on metadata fields like agent type and timestamp directly against AgentDB.

HNSW provides approximate rather than exact nearest neighbors, so results can miss marginally relevant entries. It also requires embedding generation for every stored entry and tuning parameters like M and efConstruction for the dataset size.