Join our everyday and weekly newsletter for the latest updates and exclusive content for top -class AI in the field. More information
AI can automate many tasks they want to perform. One disadvantage, however, is that they tend to be oblivion. Without long -term memory, you must either complete the role in one session or to be constantly told again.
Since businesses continue to explore used boxes for AI agents and how to carry out them safely, companies to develop agents must consider how to make them less forgiveness. Long -term memory will make much more valuable agents in the workflow, capable of remembering instructions for complex tasks that require completion of several moves.
Manvinder Singh, Vice President for AI product management in Redis, said Venturebeat that memory makes agents more robust.
“Agency memory is essential for increasing (agents’) effectiveness and abilities, because LLM is by their very nature without nationality – remember that as the speed or history of chat,” Singh said in -mail. “Memory allows AI agents to remind past interactions, maintain the context of information and maintenance to provide more cohesive, personalized responsibilities and more impacts of caroma.”
Companies like Langchain began to offer the possibilities of expanding agent memory. Langchain’s Langmem SDK helps developers to build agents with tools “for extraction of conversation information, optimize the behavior of the agent through fast updates and long -term memory of behavior, facts and events”.
Other options include Memobase, Open-Source running in January, which provides “users’ memory”, so the applications remember and adapt. Crewai also has long -term agent instruments, while OpenI’s Swarm requires users to bring their memory model.
Mike Mason, Chief Director of Tech Consultancy Throudworks, said Venturebeat in -mail that better agent memory is changing how companies use agents.
“Memory transforms AI agents from simple reactive instruments into dynamic and adaptive assistants,” Mason said. “Without it, it has to rely solely on what is provided in a one -time session, which limits their ability to improve interactions over time.”
Better memory
Along-la-la-la-la-la-la-la-lshmory V could be found in various flavors.
Langchain works with the most common types of memory: semantic and procedural. Semantic refers to the facts, while processdural concerns processes or how to perform tasks. The company said they already have good short -term memory agents and can respond in the current conversation thread. Langmem imposes procedural memory as updated instructions in the challenge. Banking has identified the interaction formulas for its work on quick optimization and updated the “system call to strengthen efficient behavior. This creates a loop of feedback, where the basic instructions of the agent evolve on the basis of the observed performance.
Scientists working on ways to expand the memories of AI models and subsequently AI agents have found that long -term memory agents can learn from mistakes and improve. The Octuber 2024 contribution has explored the concept of self -term AI through long -term memory, showing that models and activation really improve, the more they remember. Models and agents are starting to adapt to more individual needs because they remember more of their own instructions for length.
In another post, scientists from Rutgers University, Ant Group and Salesforce introduced themselves to a new memory system called A-MEM, based on the method of getting to know the remarks Zettelkasten. In this system, create knowledge networks that allow “more adaptive and context -based memory”.
Redis’s Singh said that long -term memory function agents such as hard drives, “hold a lot of information that persists in several runs or conversations, hosting agents learn from feedback and adapt to users’ preferences.” When agents are integrated into workflows, this kind of adaptation and self -taught allows organizations to maintain the same set of work on the task long enough to complete it without having to re -perform them.
Reflections on memory
At the same time, it is not to get agents to remember more; Singh said that the organization must also decide what agents need to forget.
“There are four high -level decisions that you have to make when designing memory management: Which type of memories do you store? How do you store and update memories? How to load when you get under memories? How do you break down memories? “Singh said.
He stressed that businesses must be these questions because they will make sure that “the speed of maintenance of the agent system, scalabibility and flexibility is the key to creating fast, efficient and user experience”.
Langchain also said that the organization must be clear about the behavior of the people of my people and which should learn through memory; What types of Nowlege should constantly watch agents; And what triggers memory remembers.
“In Langchain, we have found that it is first useful to identify the capacity that your agent has to read, map them into specific types of memory or approaches, and then implements them in your agent,” the company said in the blog.
Recent research and these new offers only take up the start of tools to provide long -term memory agents. And because businesses are planning to deploy agents on a larger scale, memory is present on the occasion of companies to distinguish their products.
