Memrust is an open-source memory engine for AI agents built in Rust. Instead of treating memory as just vector search, it combines semantic search, BM25, entity graphs and recency-aware ranking to retrieve information the way agents actually need it. With agent-native APIs like remember(), recall() and forget(), built-in memory lifecycle management, multi-agent support, MCP integration, and Python, TypeScript and Rust SDKs, memrust gives AI agents a true memory layer-not just a vector database.
Hey Product Hunt! š I'm Sonu, creator of memrust.
After building AI agents for production, I kept running into the same problem: vector databases are great at semantic search, but agents don't just search-they need to remember. They need to recall exact identifiers, relationships, recent decisions, and knowledge accumulated over time.
That's why I built memrust.
Instead of relying on vectors alone, memrust combines semantic search, BM25, entity graphs, and recency-aware ranking into a single memory engine. It also introduces agent-native concepts like `remember()`, `recall()`, and `forget()`, along with memory lifecycle management, MCP support, and SDKs for Python, TypeScript, and Rust.
The entire project is open source (Apache-2.0), and every benchmark on the website is reproducible from scripts in the repository. If you think the approach can be improved-or you've hit memory problems in your own agents-Iād genuinely love to hear your thoughts.
You can also connect with me on [LinkedIn](https://www.linkedin.com/in/sonu...).
Happy to answer any questions about the architecture, benchmarks, design decisions, or the roadmap. Thanks for checking out memrust! š
Report
the rust sdk alone makes me want to try this, honestly tired of memory layers that feel like vector db wrappers. one thing that would help though is some kind of memory summarization or consolidation step for really long-running agents, so old interactions get condensed instead of just piling up in storage.
the rust sdk alone makes me want to try this, honestly tired of memory layers that feel like vector db wrappers. one thing that would help though is some kind of memory summarization or consolidation step for really long-running agents, so old interactions get condensed instead of just piling up in storage.