Does Auto-GPT remember what it did in previous runs?

[TungstenX]

Does Auto-GPT remember what it did in previous runs?
Can it pick up where it left off if it was interrupted?
It managed to get it self either in a loop or bombed out. Does it learn from those actions?

[Wladastic]

As of now it does rewrite the previous memory but the re-arch movement aims to rebuild that process.
Maybe a persistent memory will be there soon, as it bugs me too when it has to do a task again because it either ran into token limits or openai servers are down for a minute.

[TungstenX]

In theory it will then run into the same issues every time it restarts the same tasks

[Boostrix]

Can it pick up where it left off if it was interrupted?

for the time being there is no scheme for proper workspace persistence, aka an agent cannot be easily suspended/resumed or restarted while also being able to make heads and tails of its workspace artifacts or previous steps.

There is some experimental work related to improving things, but nothing specific at this point
Some folks have tinkered with schemes by ingesting the previous log file into vector memory - most of these efforts were using 0.2.2 (with vector memory currently being disabled due to ongoing refactorings)

It managed to get it self either in a loop or bombed out. Does it learn from those actions?

not really for the time being, there are some experimental patches floating around on discord and github to improve the situation a little, but nothing specific: #3668

[Rbosial]

No, Auto-GPT, or any other instance of the GPT language model, does not have memory of previous runs or interactions. Each interaction with the model is treated as an isolated prompt, and the model does not retain any knowledge of previous prompts or actions.

If an Auto-GPT session is interrupted, it cannot pick up where it left off. When you start a new session with the model, it begins from a blank state without any context of the previous session. Just like you cannot start a game from where you left it.

[TungstenX]

I've not been able to have a run till the end. One issue is when it wants to install a package and a prompt for user interaction is presented then it get stuck. Ctrl + C does get it out of it.
What will the final output look like?

[Boostrix]

You can constrain it not to install stuff or pre-install any requirements

[W3Warp]

Does Auto-GPT remember what it did in previous runs?
Can it pick up where it left off if it was interrupted?
It managed to get it self either in a loop or bombed out. Does it learn from those actions?

Yes
Yes
Yes

Does it work in the current implantation?
No

[Boostrix]

Should be added to the FAQs if it isn't already there.

[W3Warp]

Should be added to the FAQs if it isn't already there.

what now? not sure if you are talking with me or the OP?
If, ...I am not part of the dev team, if I was we would be much further ahead right now ROFL, am kidding, or I'm I ? ;) Given my statement above I look like am waaay ahead, but who knows really, it's not like there is any evidence of that put out there by me, or is there? ;)

[GoMightyAlgorythmGo]

Does Auto-GPT remember what it did in previous runs? Can it pick up where it left off if it was interrupted? It managed to get it self either in a loop or bombed out. Does it learn from those actions?

that is the problem with autoGPT and language models the memory and remembering how to remember. Using different systems for it, instead they just remember and push out other improtant info to remember creating a cycle of droping the last thing then picking it up and droping another thing and derailing like a toddler/person with a 1 minute memory.

[W3Warp]

Does Auto-GPT remember what it did in previous runs? Can it pick up where it left off if it was interrupted? It managed to get it self either in a loop or bombed out. Does it learn from those actions?

that is the problem with autoGPT and language models the memory and remembering how to remember. Using different systems for it, instead they just remember and push out other improtant info to remember creating a cycle of droping the last thing then picking it up and droping another thing and derailing like a toddler/person with a 1 minute memory.

Yes, but it isn't as hard as it seems to work around this until a better option exists to deal with this limitation.
One way you can give the AI a better fighting chance so to speak is to give it more precise goals where the role is better linked to the goals. If correctly understood by AutoGPT the joining Agent will know what to do every session. Sure there will still be the problem of, if a action failed over and over the context window will in the end don't have enough and the agent will start to revert back to defaults that should IMHO be already defined as the current goals not some random yunk to research business models. (no offense to you whoever defined that, I understand it's just something to use before it is decided to use something that would be useful, I assume)