reshuibuduo/TMCRA-Agent-Memory ? reverse-engineered prompt

Build me a long term memory engine for an AI agent that can keep track of conversations across multiple turns and even multiple sessions, but still stay scoped so one question does not leak into another.

I want it to save the original conversation as traceable source memory, turn the useful bits into compact current memory, and organize lasting facts into a graph that can be searched later. When a new question comes in, it should look through all of that, rank the best evidence, handle time related details, avoid duplicates, and return a small evidence packet the agent can use to answer.

Also include the benchmark setup so I can reproduce the LongMemEval run from the repo, load the released scoring artifacts, and export the score results. Keep it focused on the memory engine and the evaluation pipeline, and make sure the design is clear enough that I can inspect how each memory item came from the source conversation.