Memory Optimization

I study how memory and context policies can make long-horizon agents more reliable.

• Retrieve evidence and lessons that are relevant to the current decision.
• Record experiment outcomes and explicit failure modes for reuse across runs.
• Control what enters the working context through selective loading and summarization.

This direction connects production experience in agent harnesses and self-evolving experimentation with ongoing research. It is presented as a research direction, without a claim of a completed memory benchmark or a published paper.