PUBLISHEDDev Tooling
June 2026
Atomir
Atomic Memory Infrastructure for LLM Agents · pip package
A published Python package that gives LLM agents a real memory layer. Instead of dumping text blobs into fuzzy search, Atomir splits every incoming message into discrete facts on write and decomposes user queries into sub-questions on read — atomic on both ends.
- Atomic fact extraction + reconciliation (add / update-with-history / delete) with similarity gates to avoid over-merging
- Hybrid retrieval — dense embeddings + BM25 lexical search fused via reciprocal rank fusion
- Vendor-neutral: swappable LLM, embedder, and storage backends; runs fully offline by default (no API keys)
- Drop-in LangChain retrievers and LangGraph nodes, plus an HTTP server mode with an identical client interface
PythonLangChainLangGraphQdrantBM25EmbeddingsDocker
[01]
