LanceDB-backed long-term memory system with hybrid retrieval, reranking, multi-scope isolation, and management CLI.

• Hybrid retrieval pipeline combining vector search, BM25 full-text search, reciprocal rank fusion, cross-encoder reranking, and adaptive scoring stages (recency, importance, length normalization, time decay, noise filtering, MMR)
• Multi-scope memory isolation for access control, auto-capture with category detection (preference/fact/decision/entity), and optional auto-recall injection into agent context
• OpenAI-compatible embedding abstraction with task-aware requests, LRU caching, and support for multiple embedding providers; reranking via Cohere or Jina
• CLI commands for memory management: list, search, stats, delete, bulk operations, export/import, reembedding, and migration from legacy memory-lancedb
• Agent tools for recall, store, forget, update, and stats; noise filtering for low-quality content; JSONL session distillation pipeline for data preprocessing