Skip to content

Home

Jump to bottom
Klein Panic edited this page Apr 4, 2026 · 3 revisions

Memory-Spark

A high-performance RAG (Retrieval-Augmented Generation) pipeline for OpenClaw agents, built on NVIDIA's Nemotron models and achieving SOTA-level retrieval accuracy on BEIR benchmarks.

Quick Links

  • Architecture - System design, component breakdown, data flow, pool isolation, LanceDB schema
  • 15-Stage Pipeline - Detailed retrieval flow with code snippets, I/O types, and performance
  • API Reference - All memory tools, manager API, storage backend, reranker, embeddings
  • Configuration - Full 19-block configuration reference
  • Benchmarks - BEIR evaluation results
  • Examples - Real-world queries, multi-pool searches, temporal queries, MMR tuning, gate behavior
  • Troubleshooting - Common errors, Spark connection, embedding failures, LanceDB issues
  • Deployment - Setup and deployment guide

Key Features

  • Dynamic Reranker Gate: Skips cross-encoder when vector confidence is high, reducing latency by 50%
  • Reciprocal Rank Fusion: Scale-invariant hybrid merging of vector and BM25 results
  • HyDE Support: Hypothetical Document Embeddings for improved recall
  • MMR Diversity: Post-reranking diversity filtering
  • Temporal Decay: Time-weighted relevance scoring
  • Multi-Pool Search: Isolated memory pools per agent

Performance Highlights

Dataset NDCG@10 Recall@10 vs 2026 SOTA
SciFact 0.7889 0.9243 +16.5%
FiQA 0.5920 0.8234 +68.0%
NFCorpus 0.4443 0.7012 +35.5%

Infrastructure

Fully self-hosted on NVIDIA DGX Spark:

  • Embedding: llama-embed-nemotron-8b (4096-dim)
  • Reranker: llama-nemotron-rerank-1b-v2
  • LLM: Nemotron-Super-120B
  • Storage: LanceDB with IVF_PQ indexing

Clone this wiki locally