NanoVector
A roughly 120KB, zero-dependency vector search and episodic memory library for AI agents, written in pure C99 with hand-tuned AVX2, ARM NEON and x64 assembly.
NanoVector targets the common case of AI agents storing between 50 and 50,000 vectors, such as conversation turns, tool history or episodic facts, where graph-based indexes like HNSW add pointer indirection, memory overhead and non-deterministic recall for little benefit; it instead does exact, sub-millisecond brute-force SIMD search directly in CPU cache.
It avoids the multi-hundred-megabyte dependency chains that libraries like ChromaDB or FAISS pull in, and the multi-second cold-start penalty that comes with importing them, which matters for CLI tools, serverless functions and autonomous agent loops.
It installs via pip, uv, conda-forge, or an APT package for Debian and Ubuntu, and supports CPython, including free-threaded builds, PyPy, and Windows, Linux and macOS on both x86_64 and ARM64.