If the big model is the "brain" of an AI agent, then memory is its "hippocampus". An intelligent agent that only handles the current conversation is difficult to handle complex tasks that span across days and sessions; To make the agent truly remember and use it, there is a whole set of engineering problems that are far more complex than imagined.
Firstly, it is necessary to distinguish between two types of memory. Short term memory usually refers to the content in the current context window, with limited capacity and increasing cost with length; Long term memory refers to information that is persisted to external storage and can be retrieved on demand. The former is responsible for 'what you are thinking at the moment', while the latter is responsible for 'what you will remember in the long term'. How to connect the two is the first difficulty in design.
The second difficulty is that more memories are not necessarily better. Putting all the historical dialogues into the search database at once can easily bring out a lot of noise when recalling, which can actually drag down the quality of the answers. How to determine whether a piece of information may be useful in the future lacks reliable standards. Moreover, memories may expire - the preferences expressed by users last week may have changed this week.
The common practices in the industry currently include hierarchical memory (working memory, episodic memory, semantic memory), periodic summary compression, and letting agents actively decide when to write and read memory. These plans each have their own advantages and disadvantages, and overall they are still in a rapid evolution stage, far from mature.
For the industry, memory ability is likely to become a key indicator for distinguishing between "toy level agents" and "usable level agents". Whoever can make the intelligent agent remember what needs to be remembered and forget what needs to be forgotten is closer to the universal assistant that people expect.
【 Reference Source 】 Comprehensive compilation of industry information released publicly