Mihaji Memory

Key features

• Double-track hybrid search (Hybrid Seach + RRF): Combined with the ChromaDB Multilingual Condensity Vector (MUS) and Jieba BM25 (compatible to precise proprietary terms), by a non-supervised integration sorting of RRF (Reciprocal Rank Fusion), recall quality for semantics versus precision hits.
• DSH-mihaji with DSHThe default q8 quantitative version, more than half of the memory in the fp32 province, accuracy measurements consistent with fp32 ⁇ 0.992, old fp32 vector need not be reset.
• Invalid shredded noise filtering < 20 characters
• Two-part automatic injection (Two-Pass Prefetch & A-Res Sampling): Before each round of dialogue, two-part search is performed automatically at the bottom: Top-3 hit with precision + Top-2 memory strength weighted sample (A-Res algorithm, W2 square weight), both to ensure that key facts are not forgotten and to avoid seeing a context where machines repeat.
• Automatically suture around context (Context Chunk Stitching): Multi-section long memories automatically trace the cunk idx ± 1 section adjacent to the cunk idx when retrieving a hit, so as not to interrupt the break.