src/memory/vector_memory.rs

This file is your semantic LoopMemory, under the vector_memory feature. Entries are embedded through an EmbeddingProvider and stored in a VectorIndex.

Retrieval blends vector similarity with keyword and tag overlap, which is the shared scorer’s lexical signal, and consolidation prunes and merges with index eviction.

Key items

VectorMemoryStore::new(embedder, index) takes a Box<dyn EmbeddingProvider> plus a Box<dyn VectorIndex>, so swapping tiers from linear to HNSW to Qdrant or pgvector costs you one constructor call.

VectorMemoryConfig and with_config expose your retrieval blend and eviction knobs.

GoldenSet, GoldenQuery, golden_set(), and precision_ratio(hits) carry your quality bar: a fixed query set with known-relevant entries, where the 80% precision gate runs against every backend.

Behavior notes

The golden set is public, so your custom backend proves itself against the same bar.

Real embedders arrive through provider::embeddings with openai or ollama plus vector_index, and HashingEmbedder covers your tests.

Pair MemoryDemotionSink with SearchMemoriesTool and this becomes the store half of the evict, demote, recall loop.

Deep dive: Memory: the semantic tiers