We achieved 94.7% in Longmemeval with no hacks lol, everything open sourced (github.com)

A breakthrough in AI semantic memory has been achieved with a new open-source tool called Swafra, which attained an impressive 94.7% recall rate on the LongMemEval benchmark, a standard for evaluating long-term memory in AI assistants. This innovative tool integrates seamlessly with Claude Desktop, Claude Code, and VS Code Copilot, enabling these AI platforms to effectively remember and retrieve information through simple commands. Users can easily install Swafra via pip or npm and configure it with minimal setup. This advancement is significant for the AI/ML community as it offers a robust solution for managing long-term memory in AI systems without relying on complex database architectures or cloud services. Swafra employs advanced techniques such as semantic chunking using graph algorithms, local embeddings, and a hybrid retrieval method that fuses various scoring signals, which ensures diverse and non-redundant context retrieval. The tool's performance in categories such as knowledge updates and single-session preferences reached perfect scores, showing its potential in enhancing AI's ability to learn and adapt over time.