Save state / continue later

[TheAIAdopter]

Duplicates

• I have searched the existing issues

Summary 💡

I searched and couldn't find anything similar while i'm pretty sure it been requested before, I appologize in advance if i'm duplicating
Would be great to have a way to restart where it left when there is an issue. I couldn't find howe to do this

Examples 🌈

No response

Motivation 🔦

Would be time saver to be able to restart where it left. Almost every long project i tried failed before it finished

[psvensson]

I second this. There's a PR for caching and essentially memofying already sent requests to the GPT API, but it would also be nice if we could choose when/if to erase the Pinecone data (and the Redis erase would be nice to have as a separate command or CLi flag too)

[TheAIAdopter]

I second this. There's a PR for caching and essentially memofying already sent requests to the GPT API, but it would also be nice if we could choose when/if to erase the Pinecone data (and the Redis erase would be nice to have as a separate command or CLi flag too)

Can you please point me to some info about that? I would really like to be able to do something like that. what can be done exactly with this PR? sorry i'm new to this and i'm on a learning curve....

[psvensson]

The PR was created today and it is here; #816
I don't know Python myself so I'm not sure myself. But the above PR would at least speed up the time restarting by quite a bit.

[GoMightyAlgorythmGo]

My project always exit or crash before finishing. They are longer. It is impossible for me to get smth done in 1 day. I think a system where Auto GPT writes down issues milestones and priority sorts in 3 txt files and then always every 30 minutes sorts and reads them would maybe help for longer projects. Even if AutoGPT loses some progress at least he can continue!

[Stunspot]

I have yet to see auto-gpt successfully complete anything. I've seen it claim to, but it always falls over at the end and never writes anything to disk. I'm pretty annoyed at the compute money I spent.

[Slowly-Grokking]

and never writes anything to disk. I'm pretty annoyed at the compute money I spent.

I def feel your pain on the compute money without yielding results. You can make it write to disk more reliably by explicitly telling it to do so in your goals.

I can make autogpt do simple tasks, more complex tasks have thus far ended in donothing loops, or had charmap encoding issues. If I could reliably resume progress, i believe that I could be successful.

If anyone has completed complex tasks, they must have fixed bugs like the do_nothing loop without sharing them, or they are not sharing their prompting/goal crafting strategy that is yielded them good results.

[Stunspot]

Oh I can get it to do all kinds of crap, I just can't make it write a file to disk that isn't some kind of hallucination.

[Boostrix]

This could probably be renamed "Checkpointing Support" ?

I can make autogpt do simple tasks, more complex tasks have thus far ended in donothing loops, or had charmap encoding issues.
[...]
If anyone has completed complex tasks, they must have fixed bugs like the do_nothing loop without sharing them, or they are not sharing their prompting/goal crafting strategy that is yielded them good results.

At least for me, it's been primarily a matter of goal crafting strategy + priming the agent with sufficient context, often repeating redundant constraints (some of the deeper plans, the agent will lose context/constraints and will try to create files outside of its workspace)

That is why I suggested elsewhere (#1359 (comment)) to create new directories as part of the project so that the community can share working/non-working "ai_settings.yaml" files to see/compare results - just imagine the degree of unit testing and regression testing that would be faciltated by having this community of ~100k users provide stuff that works/isn't working for them.

The degree of self-improvement that would be possible this way, even using automated tools, would be thrilling !
Especially if Auto-GPT could also internally keep track of the number/type of steps it needed to complete a task:

Some of the things that I could make it complete with little effort were:

• create a directory named FOO, use this as the working directory
• create a new file named FOO/main.cc to compute the first 1024 prime numbers
• create the corresponding FOO/CMakeLists.txt file for cmake including unit tests
• run/test the program, exit if successful

[github-actions]

This issue was closed automatically because it has been stale for 10 days with no activity.