- Pick an existing chat, preferably with a longer history
- Prompt this (or similar):
Summarise this conversation in a compact symbolic form that an LLM can interpret to recall the full content. Don't bother including human readable text, focus on LLM interpretability only - To interpret the result, open a new chat and try a prompt like:
Restore this conversation with an LLM based on the compact symbolic representation it has produced for me: ...
For bonus points, share the resulting symbolic form in the comments! I'll post some examples below.
I can't say it's super successful in my tests as it results in a partially remembered narrative that is then badly restored, but it's fascinating that it works at all, and it's quite fun to play with. I wonder if functionality like this might have some potential uses for longer-term memory management / archival / migration / portability / etc.
NB this subreddit might benefit from a "Just for fun" flair ;)
[link] [comments]