Memory Vault adds hybrid search, RRF

Hybrid retrieval is the right move here: pure embeddings are good at “what did I mean?”, but bad at “find the exact phrase I used.” RRF is a sane choice because it avoids score-normalization drama and gives you a durable way to merge heterogeneous rankers.

• –Vector search handles vague, concept-level recall; tsvector + GIN catches exact terms, names, and rare technical phrases
• –RRF is a pragmatic fusion method for memory/RAG because it combines rank lists without tuning cosine vs. BM25-style scores
• –Query enrichment via tokenizer-derived variants is a useful second-order improvement, especially for technical jargon and multi-subword tokens
• –The local-first stack matters: Postgres 16, pgvector HNSW, and CPU-friendly MiniLM make this reproducible without cloud dependencies
• –This is less a flashy launch than a solid retrieval architecture lesson for anyone building long-term assistant memory