Memory-First Retrieval (CONCEPT:KG-2.12)

Overview

Memory-First Retrieval is the multi-stage recall policy layered over the KG-2.3 hybrid retriever: HyDE query expansion → dual thresholds → self-correcting two-pass → quantitative-fidelity ledger. Assimilated from Quarq Agent (agent-oss/agent.py), made graph-native so every retrieved hit carries backlink-boost + positional encodings. Extends KG-2.3 (Unified Retrieval).

How it works

• HyDE expansion. The ORCH-1.27 planner role emits a structured plan — multiple vector formulations (baseline / entity / action / literal-unit) + keywords + a search_mode — and each sub-query runs through the existing retrieve_hybrid; results merge by id-dedup + max-score.
• Dual thresholds. standard (0.38) for point facts, deep (0.28) for aggregations / temporal spans (hyde_planner.HYDE_THRESHOLDS).
• Self-correcting two-pass. If the KG-2.6 quality gate reports gate_passed=False after the first pass, a second pass re-runs at the deep threshold and merges — an evidence-based trigger, stronger than Quarq's model-self-report REQUIRED_DATA.
• Quantitative-fidelity ledger. build_evidence_ledger emits an ACCEPT/REJECT table with extracted numbers so a generator aggregates a complete ledger rather than the single most salient row.

Key files / API

Piece	Location
Pure HyDE helpers	knowledge_graph/retrieval/hyde_planner.py (HydePlan, parse_hyde_plan, merge_retrievals, build_evidence_ledger, threshold_for_mode)
Orchestration	knowledge_graph/retrieval/hybrid_retriever.py (plan_and_retrieve, _generate_hyde_plan)
Entry surface	knowledge_graph/orchestration/engine_query.py (search_hybrid(mode, self_correct)); MCP graph_search(mode="hyde"|"deep", self_correct=True)

Wiring (≤3 hops)

graph_search → search_hybrid → plan_and_retrieve → retrieve_hybrid (exactly 3 hops — plan_and_retrieve stays a retriever method, not a service).

Research provenance

Quarq Agent retrieval stack — agent-oss/agent.py:1817-2825, 2435, 3211.