Researchers at the University of Illinois Urbana-Champaign and Google Cloud AI Research have developed a framework that enables large language model (LLM) agents to organize their experiences into a memory bank, helping them get better at complex tasks over time.
The framework, called ReasoningBank, distills “generalizable reasoning strategies” from an agent’s successful and failed attempts to solve problems. The agent then uses this memory during inference to avoid repeating past mistakes and make better decisions as it faces new problems. The researchers show that when combined with test-time scaling techniques, where an agent makes multiple attempts at a problem, ReasoningBank significantly improves the performance and efficiency of LLM agents.
Their findings show that ReasoningBank consistently outperforms classic memory mechanisms across web browsing and software engineering benchmarks, offering a practical path toward building more adaptive and reliable AI agents for enterprise applications.
The challenge of LLM agent memory
As LLM agents are deployed in applications that run for long periods, they encounter a continuous stream of tasks. One of the key limitations of current LLM agents is their failure to learn from this accumulated experience. By approaching each task in isolation, they inevitably repeat past mistakes, discard valuable insights from related problems, and fail to develop skills that would make them more capable over time.
The solution to this limitation is to give agents some kind of memory. Previous efforts to give agents memory have focused on storing past interactions for reuse by organizing information in various forms from plain text to structured graphs. However, these approaches often fall short. Many use raw interaction logs or only store successful task examples. This means they can't distill higher-level, transferable reasoning patterns and, crucially, they don’t extract and use the valuable information from the agent’s failures. As the researchers note in their paper, “existing memory designs often remain limited to passive record-keeping rather than providing actionable, generalizable guidance for future decisions.”
How ReasoningBank works
ReasoningBank is a memory framework designed to overcome these limitations. Its central idea is to distill useful strategies and reasoning hints from past experiences into structured memory items that can be stored and reused.
According to Jun Yan, a Research Scientist at Google and co-author of the paper, this marks a fundamental shift in how agents operate. "Traditional agents operate statically—each task is processed in isolation," Yan explained. "ReasoningBank changes this by turning every task experience (successful or failed) into structured, reusable reasoning memory. As a result, the agent doesn’t start from scratch with each customer; it recalls and adapts proven strategies from similar past cases."
The framework processes both successful and failed experiences and turns them into a collection of useful strategies and preventive lessons. The agent judges success and failure through LLM-as-a-judge schemes to obviate the need for human labeling.
Yan provides a practical example of this process in action. An agent tasked with finding Sony headphones might fail because its broad search query returns over 4,000 irrelevant products. "ReasoningBank will first try to figure out why this approach failed," Yan said. "It will then distill strategies such as ‘optimize search query’ and ‘confine products with category filtering.’ Those strategies will be extremely useful to get future similar tasks successfully done."
The process operates in a closed loop. When an agent faces a new task, it uses an embedding-based search to retrieve relevant memories from ReasoningBank to guide its actions. These memories are inserted into the agent’s system prompt, providing context for its decision-making. Once the task is completed, the framework creates new memory items to extract insights from successes and failures. This new knowledge is then analyzed, distilled, and merged into the ReasoningBank, allowing the agent to continuously evolve and improve its capabilities.
Supercharging memory with scaling
The researchers found a powerful synergy between memory and test-time scaling. Classic test-time scaling involves generating multiple independent answers to the same question, but the researchers argue that this “vanilla form is suboptimal because it does not leverage inherent contrastive signal that arises from redundant exploration on the same problem.”
To address this, they propose Memory-aware Test-Time Scaling (MaTTS), which integrates scaling with ReasoningBank. MaTTS comes in two forms. In “parallel scaling,” the system generates multiple trajectories for the same query, then compares and contrasts them to identify consistent reasoning patterns. In sequential scaling, the agent iteratively refines its reasoning within a single attempt, with the intermediate notes and corrections also serving as valuable memory signals.
This creates a virtuous cycle: the existing memory in ReasoningBank steers the agent toward more promising solutions, while the diverse experiences generated through scaling enable the agent to create higher-quality memories to store in ReasoningBank.
“This positive feedback loop positions memory-driven experience scaling as a new scaling dimension for agents,” the researchers write.
ReasoningBank in action
The researchers tested their framework on WebArena (web browsing) and SWE-Bench-Verified (software engineering) benchmarks, using models like Google’s Gemini 2.5 Pro and Anthropic’s Claude 3.7 Sonnet. They compared ReasoningBank against baselines including memory-free agents and agents using trajectory-based or workflow-based memory frameworks.
The results show that ReasoningBank consistently outperforms these baselines across all datasets and LLM backbones. On WebArena, it improved the overall success rate by up to 8.3 percentage points compared to a memory-free agent. It also generalized better on more difficult, cross-domain tasks, while reducing the number of interaction steps needed to complete tasks. When combined with MaTTS, both parallel and sequential scaling further boosted performance, consistently outperforming standard test-time scaling.
This efficiency gain has a direct impact on operational costs. Yan points to a case where a memory-free agent took eight trial-and-error steps just to find the right product filter on a website. "Those trial and error costs could be avoided by leveraging relevant insights from ReasoningBank," he noted. "In this case, we save almost twice the operational costs," which also improves the user experience by resolving issues faster.
For enterprises, ReasoningBank can help develop cost-effective agents that can learn from experience and adapt over time in complex workflows and areas like software development, customer support, and data analysis. As the paper concludes, “Our findings suggest a practical pathway toward building adaptive and lifelong-learning agents.”
Yan confirmed that their findings point toward a future of truly compositional intelligence. For example, a coding agent could learn discrete skills like API integration and database management from separate tasks. "Over time, these modular skills… become building blocks the agent can flexibly recombine to solve more complex tasks," he said, suggesting a future where agents can autonomously assemble their knowledge to manage entire workflows with minimal human oversight.