How does memory management work?

[ProjCRys]

I know that it stores memories into two different types (the context and the archive).

After playing around with MemGPT, I noticed that in loading previous discussion, it needs to use the lastest saved json file. So based from that, it only loads the previous discussion and it's contents accessible are only what's inside the json file.

And if I wanted to ask a question to MemGPT with the answer coming from another json load file, I would have to load a json file that contains that answer. (Correct me if I'm wrong but this is what I learned from chatting with it.)

Also, there is also a bug that it crashes whenever I try to load a save file so I never get to tests how it access through it's archive memories.

From a single run with long conversation, it does it job well as it can even recall memories from the beginning of the conversation. But if I closed the discussion and reopened the program, then ask what happened. All I get is that MemGPT couldn't load the previous discussion and say ran out of input. And if does load, can I ask some questions that the information is another json load file?

[sarahwooders]

If you load from a previous conversation JSON, then it should be exactly as if though you're continuing the conversation from that point and have all the same memories. If you're not seeing this, could you please post an issue that includes the version you are running, the command you ran, and screenshots of the problematic outputs?

[cpacker]

As @sarahwooders mentioned, if you can provide an example (eg screenshots) of starting a conversation, saving it, then loading it (or failing to load it), that would help us out a ton!

Also letting us know how you installed MemGPT and what the version is if you installed via pip (pip list | grep pymemgpt).

[ProjCRys]

@sarahwooders @cpacker After I updated the memgpt (through pip install pymemgpt and by git cloning it), I couldn't do run it anymore. Before it was working perfectly fine except the /load. Since /load had been remove this is the problems I've been getting.

I posted the issue #248 as an update as I think tackling this is more important as it handles the basic running of MemGPT using LM Studio. I tried running the main.py from the git cloned repo but all I got is this error.

python main.py

On the issue of #248 Here's what I got:
create venv and do 'memgpt run'

[osgiliath-stone]

It's not clear to me if the error I got is related, but I got a similar error message and temporarily bypassed it by removing "async" from "async def get chat completion(..." on line 19 of chat_completion_proxy.py

My error message is not the same, but I did not upgrade the library like you did, I just tried to implement the revised code in a piece meal fashion found on the latest commit to chat_completion_proxy.py, and had to make that change to make it work for me. Unfortunately, I am still getting a JSON error once I hit about 5,500 submitted tokens (which excludes tokens generated by MemGPT).

If you want to continue working with latest MemGPT code while the memory issues are being worked out, you might want to try out this little hack. But this hack would only make sense if your upgrade of the library included the recent changes to chat_completion_proxy.py, and I am not sure that's been approved yet.

[ProjCRys]

from the updated version on the MemGPT I cloned, "async" from "async def get chat completion(..." on line 19 of chat_completion_proxy.py had been already removed. However, I did get some occasional working responses from my MemGPT but I had to retry it couple of times to work, unlike the previous version that it works the first time it ran.

Still I'm thankful for your answer.