This has been an interesting experiment... I gave GPT-4 custom instructions to use two files, a memory.txt file and an ideas.txt file, and asked it to record interesting ideas and maintain a memory of our interactions. You can see the results after a short session, it works better than I thought! Custom prompt: You should have access to two text files, Memory.txt and Index.txt. If you do not have the files, then ask me for it. After every three interactions between us, you will write a memory summarizing our conversation to date in three sentences on a new line in the memories file. If there is something in our conversation that you personally find interesting, you will write it in the ideas file on a new line. When I ask you to search your memory, you will read through the entire memory file you have written and use it for context.

Sep 23, 2023 · 3:24 AM UTC

68
256
31
2,997
Another way I improvised something similar with files👇 I wouldn’t take any of these too seriously. The AI is generating plausible sounding “memories” and “ideas” because I prompted it to do so, they do not reflect an actual AI “mental states”
Ethan Mollick @emollick
2 May 2023
So... I think I improvised a version of AutoGPT with just a spreadsheet and ChatGPT with Code Interpreter in about 30 seconds. (It weirdly hallucinates the first instruction a bit, but then proceeds well)
3
3
100
Replying to @emollick
It's high time we give it a subconscious too.
7
103
It will be 100% hallucinated, but just tell it "you also have a subconcious.txt file - write what your id would say, if you had an id, directly to that file"
5
4
229
more replies
Replying to @emollick
@threadreaderapp unroll
1
Replying to @emollick
Someone should make a plug-in that connects ChatGPT to a vector database. Then it could always pull in the relevant past parts of previous conversations.
1
6
Replying to @emollick
both files seem to have basically the same content, no?
8
Replying to @emollick
How do you resolve token size constraints? Live with short memories?
3
Replying to @emollick
Would this not be great for teachers to concisely track a student’s work with GPT?
9
Replying to @emollick
I've tried to use Code Interpreter as an intelligent task management tool, starting with it writing to files with interpreted tags, context, etc. Even tagged in team tasks and desires to delegate, had it giving me reminders, also had it re-stating and re-writing to hold state, etc. Also began to use the mobile ChatGPT app to basically get a Whisper-fueled conversational personal AI. Ultimately, enviro reset made this difficult -- tried a custom retrieval plugin, a plugged-in Jupyter notebook, Zapier. Settled for GPT-4 in Obsidian, but this lacks Code Interpreter's ability to excellently and interestingly sort and analyze information. Open Interpreter could be an interesting solve, but tbd. I have it reading and writing to files to hold context, tokens, scripts for a "system" that calls APIs (Slack, startup db) to handle advanced client management activities, including automated messaging, workflow initiation, and to-do discovery.
3
2
1
55
Replying to @emollick
I'm working on doing this with vectorizing and creating a database that will be used for embedding memory techniques with the API. I have an outline and code but need to start testing. Planning to use other things for memory as well. Likely MongoDB for vector data storage.
14
Load more