Are there effort to build cognitive agent which learns and self-improves by doing some complex task (and which have own memory)

[alexandre-emmanuel]

As I understand, then current approach is to pack command/observation/nextstep and send all this to the LLM and expect that LLM will do reasoning. But there is no change to improve the cognitive abilities of the external LLM.

The agent should be able to see the result of command. If it is failure or suboptimal then agent should decide whether it should seek for new knowledge (e.g. go to the web, read additional documentation or reserach monograph) or new skills (e.g. go to the web and seek and appropriate new library) or it should even do some kind of automatic reasoning and discovery (like AI for Science - e.g. https://www.amazon.com/Automated-Formation-Mathematics-Distinguished-Dissertations/dp/1447111133 but for the algorithms).

All this requires that the agent should have some internal structure:

• it should have internal knowledge base and controller aka BDI - Belief/Desires/Intention - agent architecture. It can be implemented as LLM or as konwledge graph or expert database or some kind of NeSy (NeuroSymbolic mix);
• it should have internal memory. E.g. when the Desire/Intention part means, that agent has some goal hierarchy and some current plan of the execution of goals and continous update of them. This should be placed in the internal memory along with things it has learned additionally.
• if OpenAI allows the fine-tuning of the private models which are hosted on OpenAI, then the agent can move this internal structure to the OpenAI private model with all the caveats of loss of control, loss of access to the most important part of the agent and with the costs, of course.

I am not sure about the OpenDevin scope. Maybe the development of such agent is outside the scope of OpenDevin but OpenDevin may welcome and integrate such project if it starts to exist somewhere?

Or maybe there is alredy similar project?

I am aware of the old school cognitive architectures like Soar, OpenCog, Clarion, Workspace Memory etc. There was BICA conference for them. Maybe one can make those agent more slim and more expressible and integrate.

I would like to contribute in such effort but maybe there is place and community already where I can participate?

[alexandre-emmanuel]

I did some web search and it seems to me, that there is already ongoing effort in integrating generative models with cognitive architectures, e.g. by https://integratedcognition.ai/fss-2023/ "FSS2023: Integration of Cognitive Architectures and Generative Models". But this is effort to join the two quite disparate worlds together.

So - it may be possible that agent, about which I did proposal, may be cognitive architecture like https://soar.eecs.umich.edu/ (or others) which in turn integrates both old school expert system and LLM and maybe neural memory, also https://www.goodai.com/long-term-memory/ and https://superagi.com/towards-agi-part-1/ and https://learn.microsoft.com/en-us/azure/cosmos-db/ai-agents

Actually - agent sophistication can be the key to success of OpenDevin.

[rezzie-rich]

Gptswarm has self learning capability and a graph knowledge base. There's a long-term effort to incorporate that. Idk the current stage other than there's a PR open for that.