Synaptic Memory v0.9.0
Core Improvements
- BM25 hybrid scoring — corpus-size adaptive (large: BM25 80%, small: substring 70%)
- Supersede detection — same-title nodes prefer newest (by updated_at)
- Auto-chunking API —
add_document(chunk_size=1000)with sentence-boundary splitting + PART_OF edges - PPR edge type weights — CAUSED 1.0 > RELATED 0.4 (reduces noise from S2 ablation -14~-32%)
- Kind soft boost — hard filter → 1.5x score boost (MRR +9%)
- Phrase node filter —
_phrasenodes excluded from search results - Evidence threshold — 0.2→0.3 + first-sentence position bias
Benchmarks
- 14 external datasets (6 new: NFCorpus, SciFact, FiQA, MIRACLRetrieval, MultiLongDoc, XPQARetrieval)
- LongMemEval (ICLR 2025) — 500-question long-term memory evaluation
- PPR / Evidence / PhraseExtractor unit tests (67 new)
- Ablation S8 LLM Full stage
Results (FTS only, MemoryBackend)
| Dataset | MRR | Change |
|---|---|---|
| KLUE-MRC | 0.731 | +20% |
| SciFact | 0.536 | +29% |
| Ko-StrategyQA | 0.393 | +24% |
| FiQA (57K) | 0.209 | +58% |
| MultiLongDoc | 0.154 | +120% |
Install: pip install synaptic-memory==0.9.0