Archival storage question

[ThreepE0]

I know this is a fast-moving project, and so things change rapidly; Wondering where I can learn more about the current state of archival memory.

I've tried to keep everything simple, so I'm running the latest version of memgpt installed from pip install pymemgpt, and I'm using OpenAI ChatGPT directly (I am having some inconsistent results with the localLLMs (json missing args, etc,) but I figured I'd tackle that later.)

The tldr is that I'm wondering if this sort of inconsistency in archival storage searching is expected, and whether I can do anything about it. Was hoping to have long-term conversations where details are remembered, but corruption and weirdness seems to have crept in early in this case. Hoping it's something that I can work around if I understand things better?

The specific issue I'm having is:

1. ask memgpt to track some projects. acknowledged, works as expected.
2. add projects and tasks to tracking. works great.
3. ask memgpt to list back all my projects and tasks. I've only entered two projects with 3 tasks each. works great.
4. ask memgpt to create a purchase list for a project. that is acknowledged as expected.
5. add purchases to purchase list. acked, ok.

here's where things go wrong:

7. ask memgpt to list all my projects and purchases. Only projects are listed.
8. ask about purchases based on archival memory. a search is done, and nothing returned, and memgpt says creates two purchases that I didn't add.
9. ask to list the projects and tasks. that still works as expected.

At some point I got an error message about context length exceeded so I had to restart memgpt, which was a bit surprising as I only had a little bit of information that we were talking about. Listing the projects back after restarting worked fine, but still missing purchase list.

Grepping the pickle files, I can see the purchases mentioned in there, and they should be coming back on an archival search. Memgpt even mentioned which project it was adding the purchases to when they were added. But for some reason, it refuses to acknowledge the existence of them in the archival memory, and makes up information that I didn't add instead.