memvid/memvid: Rust-Based Serverless Memory Layer for AI Agents — 13.6K Stars

GitHub·medium signal

Memvid packages data, embeddings, search structures, and metadata into a single portable file, replacing traditional RAG pipelines with a serverless memory system claiming +35% SOTA on LoCoMo with sub-5ms local retrieval. Supports BM25 full-text, HNSW vector similarity, and CLIP visual search across text, PDFs, images, and audio (Whisper transcription). SDKs available for Rust, Node.js, Python, and CLI. The frame-based architecture inspired by video encoding uses append-only Smart Frames for versioning and crash safety — a novel approach to agent memory without external databases.