Research
One scientific explainer per retrieval mechanism. Each follows the same honest arc — the problem the paper solves, the reference implementation it cites, the deterministic way brain-server implements it, and the ceiling (built from published research, not SOTA-parity claims).
- Bi-temporal Knowledge Graph — validity-aware facts,
?at=recall - Submodular Evidence Packing — token-budgeted, diverse evidence
- TRACE Typed Edges + Faithful Explanation Paths
- Personalized PageRank Graph Retrieval — HippoRAG-2-style
- Noise-Aware Graph + Hub Dampening — the Discern release
- Calibrated Abstention + Faithful Span Verification
- The PRF Gate + Evidence-Faithful Snippet — grounding the answer
- Hybrid Fusion: RRF over BM25 + quantized vectors — Cormack & Clarke RRF, Robertson & Zaragoza BM25, Jégou quantization
- Opt-in Anticipation (the Suggest surface) — Generative Agents / MemGPT / Mem0, honestly bounded
- Structure-Aware Markdown Chunking — CommonMark split, Lewis 2020 RAG framing
- Centroid Domain Auto-Routing — the nearest-centroid classifier, carving the store by domain
- Deterministic Consolidation — record-linkage duplicates/conflicts/stale-source sweep, reviewable not autonomous
Every mechanism is a deterministic implementation of specific published
techniques over a local store — no LLM in the retrieval loop, no data egress.
The proof map ties each to a shipped release and a live
curl/brain verification.