We were inspired by the Stanford "Generative Agents" paper and decided to make an implementation that anyone could run.
1 Run setup.py 2 Enter openai api key 3 Customize config.json 4 And run it!
Each agent runs this loop: - Observe new events - Decide how to react - Plan, if we need to - Carry out the top plan - Reflect, if its time to
When it observes new events, it assigns an importance score based on how poignant the event is, then stores it in memory.
When an agent is executing a plan, it first gathers relevant memories based on Recency (exponential decay function), Relevancy (based on semantic embeddings), and Importance.
The agents also generate their own reflection’s allowing them to achieve higher level thinking. This is triggered every time the importance of recent memories passed a threshold of 100. The prompt is: “Given the most recent 100 memories, what are the 3 most salient high-level questions we can answer about the subjects involved?” After getting the questions, we use the memory retrieval system to generate answers. This allows the agents to come to conclusions such as “John is very talkative” or “Sally is often late to work”.
This is just an experiment for us, but it was a lot of fun. I think this could serve as a template to folks who are making entertainment or gaming experiences with LLMs.
I dont think this particular setup is optimal for productive output.
We welcome forks and contributions!