HowieHwong/MemoHarness ? reverse-engineered prompt

Build me a Python project called MemoHarness that runs an agent harness which learns from past runs and adapts itself for new tasks.

I want it to read a config file, connect to a model provider, and use a Harbor plus Daytona runtime to run experiments. It should support a training mode that searches over different harness settings, keeps a memory of past cases and general patterns, then uses that experience to pick a better harness for each new case. It should also support an eval only mode that reuses the saved experience without doing more search.

Please make the code organized cleanly with separate parts for the experience bank, model and embedding helpers, runtime integration, controllers, and shared data models. Include simple scripts and config examples so I can run the full loop from the command line. If you need to check the latest docs for any runtime or provider details, look them up online. Make it feel like a real research project that can be configured easily and extended later.