Permanent Memory

Your personal expert AI with permanent memory

Large Language Models are often constrained by the length of the context, once the context length is exceeded, they will lose their previous conversation memory. This has always been a significant flaw for various important applications such as personal assistants, chat-bots, and virtual companions.

One potential solution is to continuously strengthen the length of the context to adapt to longer content understanding and memory. Starting from the release of Claude 2.1 200w , OpenAI subsequently proposed permanent memory, allowing large-scale models to remember the conversation history of each individual and form memories. However, actual testing found that OpenAI's so-called permanent memory is simply to form a user's own conversation knowledge base based on the history of questions and answers. At the beginning of a new conversation, it searches and retrieves previous chat content from the knowledge base to awaken the so-called memory. This approach does not include the model's analysis and thinking process in the conversation knowledge base.

In fact, when simulating human dialogue, the memory of a large-scale model about a conversation with a person is not limited to what the other party asked and what the large-scale model said. Humans also form memories of the thinking process during the conversation. In simple terms, even if I didn't say it, when I heard your question, I thought of many things, which is also human memory. This method is less efficient and requires more storage space.

Therefore, when establishing permanent memory to help our expert bots with each conversation object, BitMind overcame this by introducing a method called "BitMemo". BitMemo uses a multi-layer knowledge graph to help any bot to establish a memory graph with any conversation object. The advantages of BitMemo include:

• Using the network structure of the knowledge graph, each conversation can find the content of nodes and the relationship between nodes closer to one or more nodes. Then provide it as important content of the context to the robot, helping the robot to always connect to the previous chat content during the conversation.

• In addition to assisting in chatting, BitMemo can also simultaneously put other answers with close relevance into the memory graph. This means not only recording what was said, but also recording what was thought before answering. This is the only way to more completely record the memory of bots.

• BitMemo adopts a multi-level knowledge graph structure, which saves overall storage space more than OpenAI's knowledge vector space memory method. Because even humans do not need to remember all the details of every conversation, just remember the key points and connections between key points during the conversation.

• The so-called permanent memory is viewed from the perspective of the chat participant. That is to say, anyone chatting will feel that BitMemo remembers the content of the conversation. However, from the perspective of the bot, a multi-level memory graph (which we call the bot's memory palace) can achieve that every bot can remember the key points and relationships of the conversation with everyone. In this way, the bot has its own most complete memory.