Yesterday OpenAI announced that it is carrying out experiments regarding the introduction of long-term memory on ChatGPT.
The team, specifically, will work to enable the chatbot to remember details between one conversation and another, making the interaction between them even deeper Artificial intelligence and user. The tests are involving a limited number of users but, if successful, the implementation should be available on a large scale soon.
So far, i large language models (LLM) were based on two types of memory: one integrated into the AI model and during the training process and a memory contextual (i.e. conversation history) that persists for the duration of the session. Until now, in fact, ChatGPT forgot a user’s interactions once a new session was started.
Various projects have experimented with the possibility of giving LLMs a memory that persists, through different dynamics. Techniques include dynamic management of context historythe history compression through a sort of summary, links to vector databases that store information externally or simply periodically insert information into a prompt the system.
OpenAI did not explain what technique it uses in this case, but the implementation reminds us of the Custom Instructions (Personalized instructions), a feature introduced in July 2023 that allows users to make custom additions to the ChatGPT system prompt to change its behavior.
Long-term memory for ChatGPT: OpenAI working to mitigate privacy risks
It goes without saying that the introduction of long-term memory on ChatGPT could bring significant benefits to those who use the chatbot.
If the tests are successful, it will be possible to create specific contexts, recalling data and information previously provided to the AI. Additionally, OpenAI says this introduction will help subscribers ChatGPT Enterprise e Team to work better together as the team’s shared memories may recall specific document formatting preferences or other data useful in the work context.
A separate discussion, however, must be made regarding the issue privacy. This type of implementation, in fact, inevitably also involves risks. According to OpenAI, long-term memory can be deactivated completely. Furthermore, the company is studying techniques to mitigate risks related to data leaks or other unpleasant situations.