how did it work for a very long archival memory？

[moviwhat]

Q1：For example, he has a lot of external memory of pages, and the answers I need may be divided into some of them. Can the model collect all the corresponding pages? Or did he just find a partial page line and return the answer?

Q2：Is the page turning process code-controlled or memGPT intelligently controlled?

[vivi]

Q1: Yes, you can also modify the system prompt/user instructions to be more aggressive about searching. When we ran our DocQA eval, it would page through all the results until it found an answer. Function trace here.

Q2: Controlled by MemGPT.

[moviwhat]

for Q1: but i see in DummyArchivalMemory.search, matches just return some of them. when the answers are divided into more than count pages, Will he take the initiative to keep turning the pages?

if start is not None and count is not None: return matches[start : start + count], len(matches)

[CodeLifeX]

I'm also confused about this. How does MemGPT know it found a answer? what's the criterion of 'i found the answer'?