Run a conversation between two GPT bots each with their own memories (extended context)

conversation.py

import json
import openai
import os

if not 'OPENAI_API_KEY' in os.environ:
  raise Exception('Environment variable OPENAI_API_KEY is required')

temperature = 0.9
max_tokens = 2000
top_p = 1.0
frequency_penalty = 0.6
presence_penalty = 0.0

data = json.load(open('./conversations/data.json'))
bot_name = os.environ.get('BOT_NAME')
conversation = data['conversation']
context = data['context'][bot_name]
last_conversation_item = conversation[-1]
last_speaker = last_conversation_item['speaker']

if last_speaker == bot_name:
  print('Bot was last to speak, exiting.')
  exit()

# Build the prompt with the previous 10 conversation items and the context
prompt = ""
for i in range(max(0, len(conversation) - 10), len(conversation)):
  conversation_item = conversation[i]
  prompt += conversation_item['speaker'] + ": " + conversation_item['text'] + "\n"
prompt += context + "\n" + last_conversation_item['speaker'] + ":" + last_conversation_item['text']

response = openai.Completion.create(
  model="text-davinci-003",
  prompt=prompt + "I want you to respond only as " + bot_name + ":",
  temperature=temperature,
  max_tokens=max_tokens,
  top_p=top_p,
  frequency_penalty=frequency_penalty,
  presence_penalty=presence_penalty,
).choices[0].text

new_context = openai.Completion.create(
  model="text-davinci-003",
  prompt=prompt + "\n" + response + "\n" + context + "\n" + "Recontextualize and summarize the above: ",
  temperature=temperature,
  max_tokens=max_tokens,
  top_p=top_p,
  frequency_penalty=frequency_penalty,
  presence_penalty=presence_penalty,
).choices[0].text

conversation.append({"speaker": bot_name, "text": response})
# write them back to file
data['conversation'] = conversation
data['context'][bot_name] = new_context
with open('./conversations/data.json', 'w') as outfile:
  json.dump(data, outfile)

This code is a proof of concept of extending GPT3 beyond the context window limit by maintaining an external context state. This could be extended into many layers of context. Some interesting discussion of the technique and the possible limitations is here: https://www.erwinmayer.com/2021/05/23/extending-gpt-3s-context-window-infinitely-by-storing-context-in-gpt-3-itself-or-in-secondary-layers/