Grafy News
Try Grafy free

ARISE architecture enables self-evolving memory for AI agents

Researchers have proposed ARISE, a memory architecture that allows enterprise AI agents to learn from failures without retraining their underlying models.

Illustration for the news story: ARISE architecture enables self-evolving memory for AI agentsSource: Grafy Research Lab

Researchers have developed a new memory architecture designed to help enterprise AI agents adapt to changing environments and evolving models without the need for expensive retraining. The system, named ARISE (Agent Refinement through Iterative Self-Evaluation), allows agents to learn from their own mistakes by evolving their memory rather than updating their core weights, tools, or workflow logic.

How does the ARISE architecture work?

According to the research paper, ARISE operates through an offline loop that functions independently of the agent's active runtime. When an agent fails a task, the system uses the task trace, the output, and diagnostic feedback to perform failure attribution. This process identifies exactly why a mistake occurred.

Once a failure is attributed, the system uses either the deployment model itself or a more powerful "teacher" model to propose specific corrections. These corrections are stored in two distinct ways. First, they are kept as "micro memory," which are task-specific corrections that can be retrieved during similar future tasks. Second, the system consolidates these individual corrections into a limited set of "macro rules" that provide general guidance.

To ensure that these updates actually improve performance, ARISE employs a validation gate. A new macro rule is only accepted into the agent's permanent memory if it demonstrates improved accuracy on a held-out validation set. This prevents the agent from adopting incorrect or suboptimal logic based on isolated errors.

What are the performance benefits of ARISE?

In evaluations conducted across 14 enterprise-workflow domains, the researchers found significant improvements in accuracy. The study utilized benchmarks including SOP-Bench and τ2\tau^{2}-bench, testing five different models.

When using a teacher model—specifically Claude Opus 4.6—for the attribution and consolidation phases, the ARISE-Teacher configuration saw a 10.9 percentage point gain over the baseline on SOP-Bench and a 5.8 percentage point gain on τ2\tau^{2}-bench. Even when the deployment model acted as its own attributor (ARISE-Self), it achieved a 10.0 percentage point gain on SOP-Bench.

Beyond accuracy, the researchers noted that ARISE-Teacher was more efficient than the baseline, using 4.0% fewer model calls and 5.2% fewer output tokens while maintaining a comparable per-task cost (a decrease of 1.0%).

Why is self-evolving memory important for enterprises?

Enterprise AI agents often operate in highly dynamic settings. The researchers note that foundation models evolve rapidly, with new state-of-the-art versions appearing frequently. Additionally, standard operating procedures (SOPs) and company policies change regularly.

Traditionally, updating an agent to reflect these changes might require rebuilding the agent, updating prompts, or retraining model weights. Retraining is often inaccessible for many enterprises as it requires access to model weights and significant computational resources. ARISE addresses this by allowing the agent to refine its behavior through a compact, inspectable, and bounded set of rules that can be rerun whenever the underlying model or workflow changes.

How does ARISE compare to existing methods?

The paper compares ARISE to several existing approaches for agent memory and prompt engineering. While methods like ExpeL focus on extracting insights and Reflexion turns feedback into verbal reflections, ARISE distinguishes itself by using a dual-memory architecture that combines instance-level retrieval (micro memory) with generalized rule sets (macro memory).

Unlike some automated prompt engineering methods that evolve instructions through mutation or evolutionary operators, ARISE focuses on a controlled loop of failure attribution and validation. This ensures that the memory remains bounded and does not grow uncontrollably, which helps manage the serving costs associated with long prompts in large-scale deployments.

How can researchers apply these concepts in Grafy?

While ARISE is a specialized architecture for building autonomous agents, the core concept of using AI to analyze and refine information is highly relevant to research workflows. For those looking to dive deeper into the technical specifics of this paper or similar machine learning research, the Reader app in Grafy provides an environment to search, question, and synthesize complex academic documents using AI.

Frequently asked questions

Does ARISE require retraining the AI model?

No, ARISE does not require updating model weights, tools, or workflow logic. Instead, it evolves the agent's performance by managing a specialized memory system that stores corrections and rules derived from previous failures, allowing the agent to improve its accuracy through retrieval rather than retraining.

What is the difference between micro and macro memory in ARISE?

Micro memory consists of specific, task-level corrections that are stored for retrieval when the agent encounters similar individual failures. Macro memory is a consolidated set of general rules derived from multiple micro memories, providing broader guidance that applies across various tasks within the agent's workflow.

How does the validation gate ensure agent reliability?

The validation gate acts as a quality control mechanism by testing proposed macro rules against a held-out validation set. A rule update is only integrated into the agent's memory if it is proven to increase accuracy, which prevents the agent from learning incorrect patterns from errors.


Source: Grafy Research Lab — ARISE: Practical Self-Evolving Memory for Accurate and Cost-Efficient Enterprise Agents