Context

Two independent sources converge on the same architecture for next-gen agent memory retrieval:

1. Agentic RAG with Hybrid Search (TDS) -- Dense + BM25 sparse retrieval with RRF fusion and agentic query reformulation
2. NVIDIA NeMo Retriever -- ReACT loop with think/retrieve/final_results triad, RRF fallback. Add design specification and project setup #1 on ViDoRe v3.

RRF rank fusion is already noted as adopted in the research log. Qdrant natively supports sparse vectors + RRF.

Action Items

• Implement BM25/sparse search alongside dense retrieval in Mem0/Qdrant config
• Wire RRF fusion for merging retrieval rounds
• Add agentic query reformulation (query rewriting, sufficiency checking) for TOOL_BASED injection strategy
• Evaluate whether memory hot path should bypass MCP bridge (in-process retriever for latency)

References

• TDS article
• NVIDIA NeMo blog
• Research log entries Set up type checking with mypy or pyright and strict config #25, Implement task assignment strategies (role-based, load-balanced, manual) #26 (2026-03-14)