What I did

• —Co-developed C3M, a multimodal memory framework maintaining a 36-entry, 6,144-token active index for long-horizon AI agents.
• —Achieved 73%/68% accuracy on the MemLens benchmark at 32K/64K token histories, beating the next best baseline by 9/3%.
• —Built the evaluation pipeline that feeds MemLens' 789 questions in session order, testing 5 task types across 3 context lengths.
• —Implemented the confidence-thresholded rule deciding if new memories merge, update, or stay separate from existing entries.

Tools

Python · PyTorch · Hugging Face Transformers · NumPy