Just as ChatGPT seems pretty capable at summarizing text, an AI with "unlimited memory" could potentially answer analytical questions about larger datasets and non-linear data (in the sense that prose is read from start-to-finish).
The OP is most excited about this ability to remember to create more structured longford outputs with internal consistency (e.g., asking questions about a fantasy universe that respects the characters that exist elsewhere in the story or universe).
I don't understand exactly how that would work. At some point, the generation would introduce new events and characters, new places or objects, and name them, but then when summarising, won't the names of some of them be lost, just because there's not enough space in the summary to name them all? The same goes for all sorts of detail, not necessarily named. At that point, what happens to the narrative about those forgotten characters, objects, etc?
The main idea, of continuously feeding the model a summary of its generation (and its dialog with the user of course) sounds interesting, but it's still not a memory. At some point, the continuous summarisation will have to grow big enough that it again exceeds the system's buffer (its "short term memory"). Either that, or it will drop so much detail from the summary that it will lose the plot.
So while this may result in longer generations, it doesn't look like it will really solve the problem of "long term memory", or long-distance dependency. It's a smart trick, but that's not enough.
I think once we have an LLM that can take books 6-12 of the Wheel of Time saga and turn them into a single book we'll be close enough for the size of the memory not to matter.
The OP is most excited about this ability to remember to create more structured longford outputs with internal consistency (e.g., asking questions about a fantasy universe that respects the characters that exist elsewhere in the story or universe).