Procedural Memory: The Key too Smarter,More Efficient AI Agents for Your Business
Artificial intelligence is rapidly evolving,and the quest for truly autonomous agents – AI systems that can learn,adapt,and execute complex tasks – is gaining momentum. A recent breakthrough centers around a concept called “procedural memory,” and it promises to unlock a new level of capability for AI in enterprise settings. Let’s explore what this means for you and your organization.What is Procedural Memory & why does it Matter?
Think about learning to ride a bike. You don’t consciously recall every step - balance, pedal, steer – you just do it. That’s procedural memory at work: the ability to perform tasks automatically, without intentional thought. Researchers are now successfully embedding this capability into AI agents.[Figure: [Figure: [Figure:
Transferable Knowledge: big Models Helping Small Models
One of the most exciting aspects of this research is the transferability of procedural memory. A team demonstrated that knowledge gained by the powerful GPT-4o could be successfully transferred to a much smaller model, Qwen2.5-14B. Here’s how it works:- Large Model Learns: GPT-4o generates procedural memory through experience.
- Knowledge transfer: This memory is then “given” to the smaller Qwen2.5-14B model.
- Performance Boost: The smaller model experiences a critically important improvement in task success and efficiency.
Continuous Learning & The Path to True Autonomy
The “Memp” framework, developed by the researchers, allows agents to continuously update and refine their procedural knowledge while actively working. This results in a steady, almost linear improvement in performance over time. however, achieving full autonomy requires addressing a key challenge: how do you evaluate the quality of an agent’s work when there isn’t a clear “right” or “wrong” answer? Many real-world tasks – like writing a thorough report – are subjective. The solution? Let LLMs be the judges. Currently,we often rely on manually-created rules to assess task completion. But these rules are often inflexible and tough to adapt. Instead, an LLM can provide nuanced, supervisory feedback, enabling agents to self-correct and improve on complex tasks. This creates a scalable and robust learning loop, bringing us closer to truly adaptable and resilient AI workers. As Fang, a researcher on the project, notes, “Hand-written rules are brittle and hard to generalize.” An LLM-as-judge offers a dynamic alternative.What this means for Your Enterprise
Procedural memory isn’t just a theoretical advancement. It has the potential to revolutionize how you deploy AI within your organization. Consider these possibilities: Automated Report Generation: AI agents that can learn to produce high-quality reports with minimal human intervention. * Streamlined Customer Service: Agents that can handle complex customer inquiries with greater accuracy and efficiency.A new technique from Zhejiang University and Alibaba group gives large language model (LLM) agents a dynamic memory, making them more efficient and effective at complex tasks. The technique,called Memp, provides agents with a “procedural memory” that is continuously updated as they gain experience, much like how humans learn from practice.
Memp creates a lifelong learning framework where agents don’t have to start from scratch for every new task. Instead, they become progressively better and more efficient as they encounter new situations in real-world environments, a key requirement for reliable enterprise automation.
The case for procedural memory in AI agents
LLM agents hold promise for automating complex,multi-step business processes. In practice, though, these long-horizon tasks can be fragile. the researchers point out that unpredictable events like network glitches, user interface changes or shifting data schemas can derail the entire process. For current agents, this often means starting over every time, which can be time-consuming and costly.
Meanwhile, many complex tasks, despite surface differences, share deep structural commonalities. Instead of relearning these patterns every time, an agent should be able to extract and reuse its experience from past successes and failures, the researchers point out. This requires a specific “procedural memory,” which in humans is the long-term memory responsible for skills like typing or riding a bike, that become automatic with practice.
Procedural Memory: The Key to smarter, More Efficient AI Agents for Your Business
Artificial intelligence is rapidly evolving, and the quest for truly autonomous agents – AI systems capable of autonomous problem-solving - is gaining momentum. A recent breakthrough centers around a concept called “procedural memory,” and it promises to dramatically improve the performance and cost-effectiveness of AI deployments within enterprises. Let’s explore what this means for you and your organization. (Image: Figure showing procedural memory improving task completion - as described in the original article. Include alt text: “Using procedural memory (right) helps agents accomplish tasks in fewer steps and using fewer tokens (source: arXiv)”)What is Procedural Memory & Why Does it Matter?
Think about learning to ride a bike. You don’t consciously think through every movement; your body just knows what to do. That’s procedural memory at work – the ability to perform tasks automatically, without deliberate thought. researchers are now successfully embedding this capability into AI agents. Traditionally, Large Language Models (llms) excel at understanding and generating text, but struggle with complex, multi-step tasks. They frequently enough require a large number of “tokens” (units of text processed) and can be inefficient. Procedural memory addresses this by allowing agents to learn how to do things, rather than simply knowing about them.Transferable Skills: Big Models Helping Small Models
One of the most exciting findings is that procedural memory is transferable. This means the knowledge gained by a powerful LLM, like GPT-4o, can be passed down to smaller, more affordable models like Qwen2.5-14B. Here’s how it works: Large Model Learns: A state-of-the-art LLM performs a task and builds up procedural knowledge. Knowledge transfer: This knowledge is then transferred to a smaller model. Performance Boost: The smaller model experiences a significant improvement in its ability to complete tasks, requiring fewer steps and tokens. According to researchers, smaller models are frequently enough good at individual actions, but struggle with long-term planning. procedural memory effectively bridges this gap, allowing you to leverage cutting-edge AI capabilities without the high costs associated with running massive models. You can acquire knowledge with a top-tier model and then deploy it on a more budget-amiable platform.Continuous Learning with the Memp Framework
The “Memp” framework takes this a step further by enabling agents to continuously learn and refine their procedural knowledge in real-time. This results in a steady,almost linear improvement in task performance. Imagine an AI assistant that gets demonstrably better at its job with each interaction. Though, achieving true autonomy requires addressing a key challenge: how do you evaluate performance on tasks without clear-cut success criteria? Many real-world business problems – like writing a comprehensive report – are subjective.The Future: LLMs as Judges
The solution? Leverage LLMs themselves as evaluators. Instead of relying on hand-coded rules (which are often inflexible and difficult to generalize), an LLM can provide nuanced feedback, allowing agents to self-correct and improve on complex tasks. Consider these benefits: Scalability: Automated evaluation scales far more effectively than manual review. robustness: LLM-based judges can handle ambiguity and provide more insightful feedback. adaptability: The evaluation process can evolve alongside the task itself. This approach represents a critical step toward building resilient, adaptable, and truly autonomous AI workers capable of handling complex enterprise automation needs.Want to stay ahead of the curve on AI in business? Daily insights on business use cases with VB Daily If you want to impress your boss, VB Daily has you covered. We give you the inside scoop on what companies are doing with generative AI,from regulatory shifts to practical deployments,so you can share insights for maximum ROI. [Link to Newsletter Signup] read our Privacy PolicyProcedural Memory: The Key to Smarter, More Efficient AI Agents for Your Business
Artificial intelligence is rapidly evolving, and the quest for truly autonomous agents – AI systems capable of independent problem-solving – is gaining momentum. A recent breakthrough centers around a concept called ”procedural memory,” and it promises to dramatically improve the performance and cost-effectiveness of AI deployments within enterprises.Let’s explore what this means for you and your organization. (Image: Figure showing procedural memory improving task completion – as described in the original article. Caption: Using procedural memory (right) helps agents accomplish tasks in fewer steps and using fewer tokens (source: arXiv))What is Procedural Memory & Why Does it Matter?
Think about learning to ride a bike.You don’t consciously recall every step; your body just knows what to do. That’s procedural memory at work – the ability to perform tasks automatically, without deliberate thought. Researchers are now successfully embedding this capability into AI agents. Traditionally, Large Language Models (LLMs) excel at understanding and generating text, but struggle with complex, multi-step tasks. They often require a large number of “tokens” (units of text processed) and can be inefficient. Procedural memory changes that. Here’s how it benefits you: Faster Task Completion: Agents learn to accomplish goals in fewer steps. Reduced Costs: Fewer tokens used translates directly into lower operational expenses. Improved Efficiency: AI handles more complex tasks with greater autonomy.Transferable Knowledge: Big models Helping Small Models
One of the most exciting findings is that procedural memory is transferable. Researchers demonstrated this by training a powerful GPT-4o model to perform tasks, then transferring that learned “experience” to a much smaller model, Qwen2.5-14B. The results? The smaller model experienced a significant performance boost, completing tasks more successfully and efficiently. This is crucial because: cost Savings: You don’t always need the most expensive, largest models. Scalability: Knowledge acquired by state-of-the-art models can be deployed on more affordable infrastructure. Capability gap Bridging: Smaller models often handle simple actions well, but struggle with long-term planning.Procedural memory fills that gap.The Memp Framework: continuous Learning for AI Agents
The “Memp” framework,developed by the researchers,allows agents to continuously build and refine their procedural knowledge while operating in a real-world surroundings.This leads to a remarkable level of improvement – a “continual, almost linear mastery of the task.” Imagine an AI agent that gets better and better at a specific job the more it does it, without constant human intervention. That’s the power of Memp.the Next frontier: LLMs as Judges
While significant progress has been made, achieving true autonomy requires addressing a key challenge: evaluating performance on tasks without clear-cut success signals. Many real-world scenarios, like writing a comprehensive research report, are subjective. How do you teach an AI to know if it’s done a good job? The answer, according to researchers, lies in leveraging LLMs themselves as judges. Currently, we rely on hand-crafted rules to assess completion, but these rules are frequently enough inflexible and difficult to generalize. here’s why LLM-as-judge is a game-changer: Nuanced Feedback: LLMs can provide sophisticated, contextual evaluations. Scalability: Automated judging eliminates the bottleneck of manual review. Robustness: The learning loop becomes more adaptable and resilient.Building the Future of Autonomous AI Workers
By combining procedural memory with LLM-based self-evaluation, we’re taking a critical step toward building AI agents that are truly autonomous, adaptable, and capable of handling complex enterprise automation tasks. This isn’t just about automating tasks; it’s about creating AI workers* that can learn, improve, and deliver real business value. Stay Ahead with VB Daily Daily insights on business use cases with VB Daily If you want to impress your boss,VB Daily has you covered. We give you the inside scoopProcedural Memory: The Key to Smarter, More Efficient AI Agents for Your Business
Artificial intelligence is rapidly evolving, and the quest for truly autonomous agents – AI systems that can learn, adapt, and execute complex tasks – is gaining momentum. A recent breakthrough centers around a concept called “procedural memory,” and it promises to unlock a new level of capability for AI in enterprise settings. let’s explore what this means for you and your organization.What is procedural Memory & Why Does it Matter?
Think about learning to ride a bike.You don’t consciously recall every step – balance, pedal, steer – you just do it. That’s procedural memory at work: the ability to perform tasks automatically, without deliberate thought. Researchers are now successfully embedding this capability into AI agents.[Figure:[Figure: [Figure:
Transferable Knowledge: Big Models Helping Small Models
One of the most exciting aspects of this research is the transferability of procedural memory. Experiments demonstrated that knowledge gained by the powerful GPT-4o could be successfully transferred to a smaller model,Qwen2.5-14B. The smaller model experienced a significant performance boost, completing tasks more effectively.This is crucial because: Smaller models are more cost-effective: You don’t always need the most powerful (and expensive) LLM. Knowledge can be democratized: Advanced capabilities can be deployed across a wider range of applications. It bridges the capability gap: Smaller models often handle simple actions well, but struggle with long-term planning. Procedural memory fills that void.Building agents That Continuously Learn
The “Memp” framework, developed by researchers, allows agents to continuously build and refine their procedural knowledge while operating in a real-world environment. This results in a steady, almost linear improvement in performance. However, achieving true autonomy requires addressing a key challenge: how do you evaluate the quality of an agent’s work when there isn’t a clear “success” or “failure” signal? Such as, how do you assess the quality of a research report?The Future: LLMs as Judges
The answer, according to researchers, lies in leveraging LLMs themselves as evaluators. Instead of relying on brittle,hand-crafted rules to score task completion,an LLM can provide nuanced,supervisory feedback. Consider these benefits: Scalability: LLMs can evaluate a vast amount of work quickly and efficiently. Robustness: LLMs can handle subjective tasks and provide more accurate assessments. * Adaptability: LLMs can learn and improve their evaluation skills over time. This approach, exemplified by Meta’s self-taught evaluator, promises to create a more scalable and robust learning loop, paving the way for resilient, adaptable, and truly autonomous AI workers capable of handling sophisticated enterprise automation tasks. Stay ahead of the curve. Get daily insights on business use cases with VB Daily. We deliver the inside scoop on generative AI, regulatory shifts, and practical deployments, so you can share impactful insights and maximize ROIProcedural Memory: The Key to Smarter, More Efficient AI Agents for Your Business
Artificial intelligence is rapidly evolving, and the quest for truly autonomous agents – AI systems that can learn, adapt, and execute complex tasks – is gaining momentum. A recent breakthrough centers around procedural memory,a concept borrowed from cognitive science,and its submission to Large Language models (LLMs). this innovation promises to unlock significant benefits for enterprise applications, making AI more effective and cost-efficient.What is Procedural Memory & Why Does it Matter?
Think about learning to ride a bike. You don’t consciously recall every step; your body just knows what to do.That’s procedural memory at work – the ability to perform tasks automatically, without deliberate thought. Researchers are now successfully embedding this capability into AI agents using a framework called Memp. Here’s why this is a game-changer: Faster Task Completion: Agents with procedural memory complete tasks in fewer steps. Reduced Costs: Fewer steps translate to fewer “tokens” used,lowering the operational costs associated with LLMs. (Tokens are the units of text LLMs process, and you’re typically charged based on usage.) Improved Efficiency: agents become more adept at handling complex scenarios.
Transferable Knowledge: Big Models Helping small Models
One of the most exciting findings is that procedural memory is transferable. Researchers demonstrated this by training a powerful GPT-4o model to develop procedural knowledge,then transferring that knowledge to a much smaller model,Qwen2.5-14B. The results? The smaller model experienced a significant performance boost, achieving higher success rates and requiring fewer steps to complete tasks. This means you don’t always need the most expensive, cutting-edge LLM to benefit from advanced AI capabilities. You can leverage the knowledge acquired by larger models and deploy it on more cost-effective solutions.Essentially, smaller models often excel at simple actions but struggle with complex planning. Procedural memory bridges this gap, providing the reasoning skills they need.Continuous Learning & The Path to True Autonomy
the Memp framework doesn’t just give agents procedural memory; it allows them to continuously build and refine it while operating in real-world environments. This leads to a “continual, almost linear mastery of the task,” according to researchers. However, achieving full autonomy presents another challenge. Many real-world tasks – like writing a comprehensive research report – lack a clear “success” signal. How does an agent know if it’s doing a good job? The answer,researchers believe,lies in using LLMs as judges. Currently, we rely on hand-crafted rules to evaluate AI performance. but these rules are often inflexible and difficult to generalize. An LLM,with its nuanced understanding of language and context,can provide more sophisticated and accurate feedback,enabling agents to self-correct and improve on complex,subjective tasks.the Future of AI-Powered Automation
Using LLMs as judges will make the learning loop more scalable and robust. This is a critical step toward building resilient, adaptable, and truly autonomous AI workers capable of handling sophisticated enterprise automation. Here’s what this means for you: Increased ROI: More efficient AI agents translate to lower costs and higher productivity. Greater Adaptability: AI systems that can learn and improve continuously are better equipped to handle changing business needs. Reduced Reliance on Manual Intervention: Truly autonomous agents require less human oversight, freeing up your team to focus on strategic initiatives.Stay informed about the latest advancements in AI and how theycurrent agent systems frequently enough lack this capability. Their procedural knowledge is typically hand-crafted by developers, stored in rigid prompt templates or embedded within the model’s parameters, which are expensive and slow to update. Even existing memory-augmented frameworks provide only coarse abstractions and don’t adequately address how skills should be built, indexed, corrected and eventually pruned over an agent’s lifecycle.
Consequently, the researchers note in their paper, “there is no principled way to quantify how efficiently an agent evolves its procedural repertoire or to guarantee that new experiences improve rather than erode performance.”
How Memp works
Memp is a task-agnostic framework that treats procedural memory as a core component to be optimized.It consists of three key stages that work in a continuous loop: building, retrieving, and updating memory.
Memories are built from an agent’s past experiences, or “trajectories.” The researchers explored storing these memories in two formats: verbatim,step-by-step actions; or distilling these actions into higher-level,script-like abstractions. For retrieval, the agent searches its memory for the most relevant past experience when given a new task. The team experimented with different methods, such vector search, to match the new task’s description to past queries or extracting keywords to find the best fit.
The most critical component is the update mechanism. Memp introduces several strategies to ensure the agent’s memory evolves.As an agent completes more tasks, its memory can be updated by simply adding the new experience, filtering for only successful outcomes or, most effectively, reflecting on failures to correct and revise the original memory.
This focus on dynamic, evolving memory places Memp within a growing field of research aimed at making AI agents more reliable for long-term tasks. The work parallels other efforts, such as Mem0, which consolidates key information from long conversations into structured facts and knowledge graphs to ensure consistency. Similarly, A-MEM enables agents to autonomously create and link “memory notes” from their interactions, forming a complex knowledge structure over time.
However, co-author Runnan Fang highlights a critical distinction between Memp and other frameworks.
“Mem0 and A-MEM are excellent works… but they focus on remembering salient content within a single trajectory or conversation,” Fang commented to VentureBeat. In essence, they help an agent remember “what” happened. “Memp, by contrast, targets cross-trajectory procedural memory.” It focuses on “how-to” knowledge that can be generalized across similar tasks, preventing the agent from re-exploring from scratch each time.
“By distilling past successful workflows into reusable procedural priors, Memp raises success rates and shortens steps,” Fang added.“Crucially, we also introduce an update mechanism so that this procedural memory keeps improving— after all, practice makes perfect for agents too.”
Overcoming the ‘cold-start’ problem
While the concept of learning from past trajectories is powerful, it raises a practical question: How does an agent build its initial memory when there are no perfect examples to learn from? The researchers address this “cold-start” problem with a pragmatic approach.
fang explained that devs can frist define a robust evaluation metric rather of requiring a perfect “gold” trajectory upfront. This metric, which can be rule-based or even another LLM, scores the quality of an agent’s performance. “Once that metric is in place, we let state-of-the-art models explore within the agent workflow and retain the trajectories that achieve the highest scores,” Fang said. This process rapidly bootstraps an initial set of useful memories, allowing a new agent to get up to speed without extensive manual programming.
Memp in action
to test the framework, the team implemented Memp on top of powerful LLMs like GPT-4o, Claude 3.5 Sonnet and Qwen2.5, evaluating them on complex tasks like household chores in the ALFWorld benchmark and information-seeking in TravelPlanner. The results showed that building and retrieving procedural memory allowed an agent to distill and reuse its prior experience effectively.
During testing,agents equipped with Memp not only achieved higher success rates but became much more efficient. They eliminated fruitless exploration and trial-and-error, leading to a substantial reduction in both the number of steps and the token consumption required to complete a task.
One of the most significant findings for enterprise applications is that procedural memory is transferable.In one experiment, procedural memory generated by the powerful GPT-4o was given to a much smaller model, Qwen2.5-14B. The smaller model saw a significant boost in performance, improving its success rate and reducing the steps needed to complete tasks.
According to Fang, this works because smaller models often handle simple, single-step actions well but falter when it comes to long-horizon planning and reasoning. The procedural memory from the larger model effectively fills this capability gap. This suggests that knowledge can be acquired using a state-of-the-art model, then deployed on smaller, more cost-effective models without losing the benefits of that experience.
Toward truly autonomous agents
By equipping agents with memory-update mechanisms, the Memp framework allows them to continuously build and refine their procedural knowledge while operating in a live environment. The researchers found this endowed the agent with a “continual, almost linear mastery of the task.”
However, the path to full autonomy requires overcoming another hurdle: Many real-world tasks, such as producing a research report, lack a simple success signal. To continuously improve, an agent needs to know if it did a good job. Fang says the future lies in using LLMs themselves as judges.
“Today we often combine powerful models with hand-crafted rules to compute completion scores,” he notes. “This works, but hand-written rules are brittle and hard to generalize.”
An LLM-as-judge could provide the nuanced, supervisory feedback needed for an agent to self-correct on complex, subjective tasks.This would make the entire learning loop more scalable and robust, marking a critical step toward building the resilient, adaptable and truly autonomous AI workers needed for sophisticated enterprise automation.
Worth a look