Improvement of AI agents with long-term memory: insights into long SDK, Memobase and A-mem-framework

AI agents can Automate many tasks that are involved wants to run. One disadvantage, however, is that they tend to be forgetful. Without long -term memory, agents either have to do a task in a single session or are constantly being commissioned.

While companies continue to investigate applications for AI agents and how they can safely implement, companies that enable the development of agents must consider how they can be less forgetful. Long -term memory makes agents in a workflow much more valuable and able to remember instructions, even for complex tasks that require several curves.

Manvinder Singh, VP of the AI product management at Redis, told Venturebeat that memory makes agents more robust.

“Agent memory is crucial for the improvement (agent) efficiency and skills, since LLMs are naturally stateless -they do not remember things such as requests, answers or chat stories,” said Singh in an e -mail. “The memory enables AI agents to remember earlier interactions, to keep information and to maintain the context in order to provide more coherent, personalized answers and more effective autonomy.”

Like companies Praise Keep starting to offer options for expanding the agent storage. Langchain’s long SDK helps developers create agents with tools “to extract information from the conversation, to optimize the behavior of the agent through quick updates and to maintain long -term memory about behaviors, facts and events”.

More options are MemobaseAn open source tool that started in January to give the agent “user-oriented memory” so that apps remember and adapt. Crewai also has tools in long-term agent memory, while Openais swarm Requires users to bring their memory model.

Mike Mason, Chief Ai officer at Tech Consultancy ThinkWorks, said Venturebeat in an e -mail that better agent storage changes the way in which companies use agents.

“The memory transforms AI agent of simple, reactive tools into dynamic, adaptive assistants,” said Mason. “Without them, the agents only have to rely on what is provided in a single session, and the ability to improve interactions over time.”

Better memory

Long -lasting memory in agents could come in different flavors.

Langchain works with the most common types of memory: semantic and procedural. Semantics refers to facts, while procedures refer to processes or the execution of tasks. The company said agents already had a good short -term memory and can react in the current thread. Langmem stores the procedural memory as updated instructions in the command prompt. Long -mem gently on his work on the immediate optimization and identifies interaction patterns and updates “the system request to strengthen effective behaviors. This creates a feedback loop in which the nuclear instructions of the agent develop based on the observed performance. “

Researchers who are working on opportunities to expand the memories of AI models, and consequently AI agents have found that agents can learn and improve with long-term memory. A Paper From October 2024, the concept of the AI self-evolution examined by long-term memory and shows that the more they remember models and agents. Models and agents begin to adapt to more individual needs because they remember customized instructions for longer.

In another paper, researchers from Rutgers University, the ants group and the Salesforce presented a new one Storage system called A-MemBased on the slip box note. In this system, agents create knowledge networks that enable “more adaptive and context -related storage management”.

With Redis’ Singh said that agents with long -term memory function work such as hard disks, “hold a lot of information about several task races or conversations in their hands, learn agents from feedback and adapt to the user preferences”. When agents are integrated into workflows, this type of adaptation and self -learning enables organizations to keep the same agents that work long enough on a task to do them without arranging them again.

Memory considerations

But it is not enough that agents can be remembered more; Singh said organizations must also make decisions What the agents have to forget.

“There are four high -ranking decisions that you have to make when designing a memory management architecture: What kind of memories do you store? How do you save and update memories? How do you pick up relevant memories? How do you fall into memories? “Said Singh.

He emphasized that companies have to answer these questions, since ensuring that an “agent system maintains speed, scalability and flexibility, is the key to creating a quick, efficient and exact user experience”.

Langchain also said that organizations have to be clear which behaviors must be set and which should be learned by memory. Which types of knowledge agents should pursue continuously? And what triggers the recall recall.

“At Langchain, we first felt useful to identify the functions that your agent needs to learn, assign them to certain storage types or approaches and only then implement them in their agent,” said the company in one Blog post.

The latest research and these new offers are only the beginning of the development of tool sets in order to give agents longer -lasting memory. And since companies plan to use agents on a larger scale, the memory company offers the opportunity to differentiate their products.