Recently, US company Bilt deployed several million AI agents, working with Letta, a startup that allows the agents to learn from and share memories, which are stored in a memory vault.
Using the “sleeptime compute” process, the agents decide what information is archived and what is needed for faster recall.
AI has a limited amount of information in the “context window”, and usually only recalls things if specifically asked to in the prompt.
However, experts are pushing for the development of storage and recall functions, which would make AI agents less error-prone.
Charles Packer, CEO of Letta, and Harrison Chase, CEO of LangChain, are two such experts seeking to improve AI memory functions.
OpenAI has also announced that its ChatGPT will store relevant information for a more personalised experience, although the method for this isn’t disclosed.
It is hoped that these memory functions can be made transparent to engineers building AI networks, making them more effective.
.jpg)