Memory3

Lightweight persistent memory system built on SQLite with sqlite-vec for SIMD-accelerated vector operations and optional HNSW indexing for O(log n) approximate nearest neighbor search. Supports hybrid search combining vector similarity with BM25 full-text search via Reciprocal Rank Fusion, Late Chunking (encoding full documents through a single Transformer pass before mean-pooling by chunk boundaries), and context window expansion that automatically retrieves adjacent chunks during search. Provides 9 tools covering memory save, search, delete, update, list, import/export, stats, and similarity operations. Features include TTL-based expiration, tag-based organization with a dedicated index table, embedding caching, syntax-aware file import for 20+ programming languages, graceful degradation across search backends, and thread-safe concurrent access.