egoist-machines/lodedb

LodeDB is an embedded vector database designed to store and search through text, images, and other data that has been converted into numerical vectors, which is the format many AI applications use for memory and search. It runs directly inside your Python program with no separate server or account required, and stores everything on disk in a compact format. The core use case is as a memory backend for AI frameworks like LangChain, LlamaIndex, and mem0, which are tools developers use to build AI applications that need to remember things. LodeDB positions itself as a faster and smaller replacement for the default stores those frameworks ship with. The benchmarks in the README show large speed advantages over alternatives like ChromaDB, Qdrant, and sqlite-vec, with much smaller on-disk file sizes. A few things make it distinctive among embedded options. Writes are incremental: instead of rewriting the entire file on each save, it only writes the rows that changed, so adding a memory entry takes under a millisecond even with a million existing entries. On machines with an NVIDIA GPU, batch search can run on the GPU instead of the CPU, reaching tens of thousands of searches per second. Storage is compressed using low-precision vector codes plus zstd compression on retained text. All data stays local by default with no telemetry on raw content. Beyond text, LodeDB supports multimodal search: you can index both images and text in the same database using a CLIP model and search across both types at once. It also ships adapters for the major AI frameworks, a command-line tool, an MCP server for coding assistants like Claude Code and Cursor, and native Swift bindings for iOS and macOS apps. Installation is a single pip command with prebuilt wheels for Linux, macOS, and Windows, requiring no compilation. Optional extras add GPU acceleration, image search, and framework adapters. The full README is longer than what was shown.