Microsoft's Agent Framework just reached Release Candidate status for both .NET and Python — the API is stable, all 1.0 features are complete, and it's production-ready. The framework is a comprehensive successor to Semantic Kernel and AutoGen, offering a unified programming model that supports A2A (Agent-to-Agent), AG-UI, and MCP protocols out of the box. Build agents that work with Azure OpenAI, OpenAI, Anthropic Claude, AWS Bedrock, Ollama, and more — all through the same abstraction layer.

The framework includes graph-based workflows (sequential, concurrent, handoff, group chat), streaming support, human-in-the-loop patterns, and type-safe function tools. Version 1.0.0rc2, released February 25, 2026, added workflow improvements and agent skills support. Migration guides are available for teams moving from Semantic Kernel and AutoGen. It's the most complete open-source agent framework from any major cloud provider.

But Microsoft's framework shares a gap with every other agent framework: agents built with it have no persistent memory across sessions, and multi-agent workflows lose their shared context when the workflow completes.

What the Framework Enables

The multi-provider support is the headline feature. A single codebase can create agents that use Claude for complex reasoning, GPT-5 for rapid responses, and local Ollama models for cost-sensitive operations. The A2A protocol enables agents built with Microsoft's framework to communicate with agents built on other frameworks — LangGraph, CrewAI, or custom implementations. MCP support connects agents to thousands of external tools and data sources.

Graph-based workflows make complex multi-agent orchestration declarative. Instead of writing procedural coordination code, developers define agent interaction patterns as graphs: sequential pipelines, parallel fan-out/fan-in, handoff chains, and group discussions. The framework handles scheduling, state management within the workflow, and streaming results back to the application.

Agent skills — the newest addition in RC2 — allow reusable capability modules that agents can invoke. A "search" skill, a "code review" skill, or a "data analysis" skill can be shared across agents and workflows. Combined with human-in-the-loop support, the framework enables production patterns where agents work autonomously but escalate to humans at configurable decision points.

The Cross-Session Memory Gap

Microsoft's framework manages state within a running workflow: agents in a group chat share conversation context, handoff chains pass state between agents, and parallel branches merge results. This within-workflow state management is well-designed and production-ready. The gap is between workflows.

When a workflow completes, its accumulated context disappears. The code review agent that identified a critical pattern in your codebase forgets it before the next review. The data analysis agent that discovered an anomaly in last week's report starts fresh on this week's report. The customer service agent that learned a user's communication preferences relearns them in every session.

The multi-provider support actually compounds the problem. An agent that switches from Claude to GPT-5 mid-workflow carries forward the conversation context. But an agent that used Claude last week and GPT-5 this week has no mechanism to share knowledge across providers. The framework's greatest strength — provider flexibility — becomes a liability without cross-session, cross-provider memory.

How MemU Adds Memory to Agent Frameworks

MemU integrates with Microsoft's Agent Framework as a persistent memory layer. Agent skills can include MemU memory operations: write discoveries at the end of workflows, read relevant context at the start. Graph-based workflows can include MemU retrieval as a first step, ensuring every agent in the workflow starts with accumulated knowledge.

For multi-provider deployments, MemU provides the cross-provider knowledge bridge. Memories generated by Claude-powered agents are available to GPT-5-powered agents and vice versa. The knowledge exists independently of the model that generated it, creating a unified organizational memory that spans providers, workflows, and time.

Microsoft built the framework for building agents. MemU provides the memory that makes agents worth building. Together, they deliver the production agent systems that enterprises actually need.

Get Started

Add persistent memory to your Microsoft Agent Framework deployments. Explore MemU at memu.pro and on GitHub.