Chatbots Struggle with Short-Term Memory, But Solutions Emerge to Improve Retention
-
Chatbots like ChatGPT struggle to remember details between and even within conversations. After enough input, around 4 million words, their performance deteriorates.
-
MIT scientists found the cause of mid-conversation forgetting is how chatbots store short-term memories. They proposed a solution called StreamingLLM to improve retention.
-
StreamingLLM changes how chatbots encode and store early parts of conversations so key details are retained longer.
-
OpenAI is testing long-term memory for ChatGPT, allowing users to make it remember or forget specific details across conversations.
-
OpenAI will use ChatGPT's memories to improve its models but is taking steps to prevent retaining sensitive information without explicit user permission.