Agents can learn with Memory in Microsoft Foundry Agent Service

2026-07-16 · Source: Microsoft Foundry Blog articles · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, short

Summary

Microsoft Foundry Agent Service now incorporates "Procedural Memory," a new capability designed to prevent AI agents from repeating past mistakes and inefficient workflows. Unlike traditional memory systems that store facts or conversation history, procedural memory captures actionable learnings from agent trajectories, such as specific actions to take under certain contexts. This system uses an LLM-as-a-judge approach to identify valuable learning signals, particularly from failures and near-failures, which are considered richer sources of improvement than successful patterns. Benchmarking on enterprise workflow tasks like STATE-Bench and Tau-Bench shows consistent performance improvements. On STATE-Bench, Pass¹ increased by 5.0% to 58.3% and Pass⁵ by 4.6% to 37.3%. On Tau-Bench, Pass¹ rose by 4.7% to 79.8% and Pass⁵ by 6.3% to 55.4%. Future enhancements include learning from population-level patterns, continuous memory evolution, and consolidation of accumulated memories.

Key takeaway

For AI Engineers developing robust agent systems, integrating procedural memory is critical to overcome repetitive failures. Your agents will learn from past experiences, reducing wasted tokens and improving task reliability. Consider adopting Microsoft Foundry Agent Service's new procedural memory feature to capture actionable learnings from agent trajectories, especially from failures, ensuring your agents execute complex workflows more efficiently and accurately.

Key insights

Microsoft Foundry's Procedural Memory enables AI agents to learn from past actions and avoid repeating errors, improving reliability.

Principles

• Failures offer richer learning signals than successes.
• Procedural knowledge is crucial for reliable agent execution.
• Real-world learning requires handling noisy, unlabeled data.

Method

Procedural Memory uses an LLM-as-a-judge to extract actionable learnings from agent trajectories, storing them for retrieval in similar future contexts.

In practice

• Implement procedural memory to prevent agents from repeating workflow errors.
• Prioritize agent failures and near-failures for learning signal extraction.
• Use STATE-Bench or Tau-Bench to evaluate agent memory improvements.

Topics

• AI Agents
• Procedural Memory
• Microsoft Foundry
• LLM-as-a-judge
• Agent Reliability
• Enterprise Workflows

Code references

• microsoft/STATE-Bench
• sierra-research/tau2-bench
• orgs/microsoft-foundry

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Teaching AI to Remember: Exploring Memory Store in Microsoft Foundry — Microsoft Foundry's Memory Store provides managed, user-scoped long-term memory for AI agents, enabling personalized, continuous interactions.
• AI Agents of the Week: Papers You Should Know About — A unified taxonomy for agent memory clarifies forms, functions, and dynamics, enabling systematic design of autonomous AI.
• What's New in Microsoft Foundry Labs – May 2026 — AI agent usefulness extends beyond task completion to include effective advocacy for user interests.
• A Practical Guide to Memory for Autonomous LLM Agents — Memory architecture, particularly its management, is more critical to LLM agent performance than model choice.
• Learning from Experience: Searchable Long-Term Memory for Data Engineering Agents — Searchable long-term memory, with automatic capture and context-level recall, enables AI agents to learn from experience, enhancing consistency and efficiency.
• Governed Memory: A Production Architecture for Multi-Agent Workflows — Governed Memory provides a shared memory and governance layer for multi-agent systems, enhancing recall and reducing token usage.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Microsoft Foundry Blog articles.