Worked on the facebookresearch/faiss repository to optimize memory usage during inverted lists loading under IO_FLAG_MMAP_IFC. Addressed a peak memory consumption issue by redesigning the loading process, merging two loops into a single-pass per-list strategy. This approach allowed memory to be released promptly, reducing peak heap usage from approximately 165 GB to 10 GB and decreasing load times from 99 seconds to 16 seconds on large indices. Ensured consistent behavior across both FileIOReader and MappedFileIOReader without altering API or ABI. Utilized C++ and Python, applying skills in memory management, performance optimization, and unit testing to validate stability and reliability.