Memv open-source library enables AI agents to build persistent memory by extracting knowledge from conversations using predict-calibrate mechanisms.

v0.1.2 adds production-ready features including PostgreSQL backend with pgvector for vectors and asyncpg pooling, plus SQLite support

Supports multiple embedding providers: OpenAI, Voyage, Cohere, and local fastembed (ONNX)

Implements bi-temporal validity tracking (event time vs transaction time) and hybrid retrieval combining vector similarity with BM25 search

Handles contradictions by invalidating old facts when new ones emerge while maintaining full audit trail

Includes episode segmentation to group messages before knowledge extraction from agent conversations