Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More
Researchers at Rutgers University, Ant Group and Salesforce Research have proposed a new framework that enables AI agents to take on more complicated tasks by integrating information from their environment and creating automatically linked memories to develop complex structures.
Called A-MEM, the framework uses large language models (LLMs) and vector embeddings to extract useful information from the agent’s interactions and create memory representations that can be retrieved and used efficiently. With enterprises looking to integrate AI agents into their workflows and applications, having a reliable memory management system can make a big difference.
Why LLM memory is important
Memory is critical in LLM and agentic applications because it enables long-term interactions between tools and users. Current memory systems, however, are either inefficient or based on predefined schemas that might not fit the changing nature of applications and the interactions they face.
“Such rigid structures, coupled with fixed agent workflows, severely restrict these systems’ ability to generalize across new environments and maintain effectiveness in long-term interactions,” the researchers write. “The challenge becomes increasingly critical as LLM agents tackle more complex, open-ended tasks, where flexible knowledge organization and continuous adaptation are essential.”
A-MEM explained
A-MEM introduces an agentic memory architecture that enables autonomous and flexible memory management for LLM agents, according to the researchers.
Every time an LLM agent interacts with its environment— whether by accessing tools or exchanging messages with users — A-MEM generates “structured memory notes” that capture both explicit information and metadata such as time, contextual description, relevant keywords and linked memories. Some details are generated by the LLM as it examines the interaction and creates semantic components.
Once a memory is created, an encoder model is used to calculate the embedding value of all its components. The combination of LLM-generated semantic components and embeddings provides both human-interpretable context and a tool for efficient retrieval through similarity search.
Building up memory over time
One of the interesting components of the A-MEM framework is a mechanism for linking different memory notes without the need for predefined rules. For each new memory note, A-MEM identifies the nearest memories based on the similarity of their embedding values. The LLM then analyzes the full content of the retrieved candidates to choose the ones that are most suitable to link to the new memory.
“By using embedding-based retrieval as an initial filter, we enable efficient scalability while maintaining semantic relevance,” the researchers write. “A-MEM can quickly identify potential connections even in large memory collections without exhaustive comparison. More importantly, the LLM-driven analysis allows for nuanced understanding of relationships that goes beyond simple similarity metrics.”
After creating links for the new memory, A-MEM updates the retrieved memories based on their textual information and relationships with the new memory. As more memories are added over time, this process refines the system’s knowledge structures, enabling the discovery of higher-order patterns and concepts across memories.
In each interaction, A-MEM uses context-aware memory retrieval to provide the agent with relevant historical information. Given a new prompt, A-MEM first computes its embedding value with the same mechanism used for memory notes. The system uses this embedding to retrieve the most relevant memories from the memory store and augment the original prompt with contextual information that helps the agent better understand and respond to the current interaction.
“The retrieved context enriches the agent’s reasoning process by connecting the current interaction with related past experiences and knowledge stored in the memory system,” the researchers write.
A-MEM in action
The researchers tested A-MEM on LoCoMo, a dataset of very long conversations spanning multiple sessions. LoCoMo contains challenging tasks such as multi-hop questions that require synthesizing information across multiple chat sessions and reasoning questions that require understanding time-related information. The dataset also contains knowledge questions that require integrating contextual information from the conversation with external knowledge.
The experiments show that A-MEM outperforms other baseline agentic memory techniques on most task categories, especially when using open source models. Notably, researchers say that A-MEM achieves superior performance while lowering inference costs, requiring up to 10X fewer tokens when answering questions.
Effective memory management is becoming a core requirement as LLM agents become integrated into complex enterprise workflows across different domains and subsystems. A-MEM — whose code is available on GitHub — is one of several frameworks that enable enterprises to build memory-enhanced LLM agents.
Source link
