There's something clear about starting a separate conversation in e.g.

There's something clear about starting a separate conversation in e.g. ChatGPT

You can see exactly what the LLM sees: what's in the conversation.

You can start a fresh conversation if you want fresh context.

The new ChatGPT memory feature kind of confuses this feature.

Now there are certain things that the LLM chooses to remember.

What precisely it chooses to remember, and why, is nondeterministic.

If you peek into the memories it's stored for you, you'll find all kinds of odd things it thinks you find generally important.

This is state that is kind of smooshed across conversations.

It's included in some way for every conversation.

In some ways it would be easier to reason about if all the LLM knew was always simply what you can see on the screen in that conversation.

The one-conversation model isn't perfect, either.

Conversations are append only; they can't be modified.

This allows you to have an iterative conversation with the LLM. "No, not like that, like this."

But once you get to the right understanding and intermediate state, you want to elide the confusing discussion that it took to get there.

It would be better to be able to post-hoc cut out parts of the conversation that turned out to not be useful so it doesn't get confused by that meandering path to the right intermediate answer.

Less like an append-only log, more like a whiteboard.