How to use Mem-alpha in inference?

[chenzhuofu]

Thanks for your great work! I'm wondering how to use Mem-alpha in inference like agent workload?

I noticed that in the evaluation, the Mem-alpha model visit the whole knowledge store at first, generating a memory for all information. Then, a frozen RAG and LLM will be used to work on downstream tasks.

In my understanding, Mem-alpha model is like a stand-alone summary model. If so, does that mean in agent inference cases, we will use a standard LLM to reason and take action, then use Mem-alpha to summarize/compress when necessary?

@wangyu-ustc

[wangyu-ustc]

Yes, that is exactly the idea. We expect the future agent systems to have a standalone memory agent that is doing the memory construction (i.e. something similar to summarization but should be more than that) and another agent that is doing the searching and retrieving from the memory system to perform the tasks.