mcxross/memora-rs

Memora-rs is a Rust implementation of a memory system for AI agents, based on a research paper called Memora that describes a way to store information so it balances being general with being specific. In plain terms, it gives an AI application a place to save facts and later retrieve the ones relevant to a question, rather than the AI forgetting everything between conversations. It can connect to different AI providers to generate the underlying embeddings and text completions: OpenAI or Azure OpenAI in the cloud, or Ollama, which runs entirely on your own computer with no data leaving your machine. For storing the memory data itself it can use ChromaDB or Redis as the underlying vector database. It can also pull text directly out of PDF, plain text, and Markdown files so those documents can be added as memories. For finding relevant memories again, it supports plain semantic search, a hybrid search method called BM25, and a more advanced multi step approach that follows a chain of related memories guided by prompts. The README is direct about one thing the paper describes that this implementation does not include: a reinforcement learning training process for improving the retrieval strategy over time. Instead, Memora-rs uses a simpler prompt guided version of that retrieval strategy without the training step. You use it either as a command line tool, where you add and query memories from the terminal, or as a library you add to your own Rust project's dependencies, with feature flags to choose which storage backend and AI provider you want compiled in. There is also a browser based web interface, built with a Rust web framework called Axum, for visually adding files and querying memories once ChromaDB is running. The project includes an integration test suite for checking Ollama and ChromaDB compatibility, and it is released under the Apache 2.0 license, which allows free use including commercial use as long as you keep the copyright notice.