Why it matters

Retrieval quality shapes memory usefulness. Indexing shapes retrieval.

Advertisement

The architecture

Vector index: embeddings + ANN.

Keyword: BM25 + inverted.

Graph: entity + relation edges.

Memory indexingVector ANNsemantic simBM25 keywordlexical matchGraphstructured relationsHybrid: combine vector + keyword; graph for structure
Memory index.
Advertisement

How it works end to end

Hybrid retrieval: vector + BM25 fusion.

Reranking: cross-encoder for precision.

Graph for structured queries.