A user has developed and open-sourced a long-term memory pipeline for AI assistants that runs its storage and retrieval components on the Qwen3 235B A22B Instruct 2507 model. The system achieved a score of 470 out of 500 on the LongMemEval-S benchmark, ranking first among known systems while reportedly being approximately ten times more token efficient than competitors.
- The pipeline uses Qwen3 235B A22B Instruct 2507 for memory operations, allowing users to select a separate model for final answer generation.
- The developer spent roughly 90% of their time refining prompts and implementing safeguards to ensure reliability with smaller models.
- The project includes a detailed write-up, an interactive benchmark viewer, and an open-source repository (c137-runner) for verifying scores against official grading scripts.
The author provides the code and benchmarks to allow others to reproduce the results and offers assistance for those interested in testing or integrating the system.