snap-stanford/memorilla ? reverse-engineered prompt

Build me the Memorilla project so I can train and test a memory system for an LLM that turns a whole pile of documents or past chats into a small set of memory tokens the model can use later.

I want a Python codebase that can load the provided datasets, train the memory module in stages, run evaluations on the supported benchmarks, and compare against simple baselines like closed book, RAG, and full context. It should also include the personalization examples and the TextWorld learning setup mentioned in the README, with easy scripts for training and evaluation and clear configs for each run.

Make it practical to use from the command line and from Python, with a simple way to load a pretrained checkpoint and ask a question against a document collection. If you need to check any current library details or model docs, look them up online first.