Reciprocal Rank Fusion Memory

Hiroshi utilizes a concurrent Reciprocal Rank Fusion (RRF) memory retrieval pipeline to achieve robust context matching across both semantic concepts and exact code symbols.

📊 The Fusion Formula

When a memory retrieval sweep is triggered, Hiroshi queries the vector similarity indexing and FTS5 keyword indexes concurrently, scoring matching records using reciprocal ranks:

\text{RRF Score}(d) = \sum_{m \in M} \frac{w_m}{k + r_m(d)}

Where:

(M) represents the retrieval models (Vector Cosine Similarity & SQLite FTS5).
(w_m) represents the model weight: 0.70 for Vector and 0.30 for FTS5.
(r_m(d)) is the 1-based rank index of document (d) within model (m).
(k) is a smoothing constant set to 60.0.

                     ┌──> Concurrent Vector Search (70% Weight) ──┐
[ Memory Query ] ────┤                                            ├──> [ Reciprocal Rank Fusion ] ──> Top-N Context
                     └──> Concurrent SQLite FTS5 (30% Weight) ────┘

This hybrid model guarantees that if your embedding server goes offline or has connection drops, the FTS5 keyword engine acts as an automatic fallback, keeping your memory layer functional.

​📊 The Fusion Formula

📊 The Fusion Formula