em-llm/EM-LLM-model — reverse-engineered prompt

Build me a Python research project for EM LLM, a long context language model wrapper that gives an existing Hugging Face model an episodic memory system instead of fine tuning it. I want it to read very long inputs in chunks, split them into meaningful memory events, store those events efficiently, and retrieve the most useful past events when answering later questions.

Please include a simple setup flow, editable install support, requirements, example config files, and scripts to download benchmark data and run evaluations. The user should be able to choose common base models like Mistral, Llama, Phi, pick LongBench, Infinite Bench, or a passkey style test, and adjust GPU usage, memory sizes, chunk sizes, disk offloading, and logging from config or command line flags.

Also include clear README instructions, result logging, and enough comments so someone can reproduce the paper style experiments. Look up current docs online if you need to.