A 3-layer temporal RAG memory library for Python. Zero external dependencies. just stdlib + sqlite3. 43KB installed. Core idea: a self-growing tag dictionary. Day 1, every tag extraction calls your AI. Day 90, ~90% resolve from a local keyword map instantly. Cost goes down over time. L1: 3-day Markdown logs. L2: 30-day AI summaries. L3: SQLite FTS5 permanent archive (~1ms search). Works with any AI, OpenAI, Claude, Gemini, or local models. pip install sandclaw-memory

HiBird — Smarter communication for growing businesses.

Smarter communication for growing businesses.

Hi everyone! I'm the maker of sandclaw-memory. I built this while working on a larger AI project. The memory system grew to 7,600 lines, and I realized the core concept, self-growing tags + temporal layers + zero dependencies, was useful beyond my own project. Every memory library I found needed a vector DB, a graph DB, or some external infrastructure. I wanted something that just works with pip install. So I extracted it into a standalone library. The part I'm most proud of: the keyword_map that learns over time. Your AI costs actually go down the more you use it, instead of staying flat. No GPU, no Docker, no vector DB. Just pip install on any machine. Happy to answer any questions about the architecture or trade-offs!

HiBird — Smarter communication for growing businesses.

Smarter communication for growing businesses.

sandclaw-memory

Zero-dep Python RAG memory that gets cheaper over time

Zero-dep Python RAG memory that gets cheaper over time