title: "Microsoft Agent Framework adds memory, checkpoints in September releas" slug: "microsoft-agent-framework-adds-memory-checkpoints-in-september-release" published: "2026-10-18" beat: "Tools" tags: ["Tools", "Launches"] creator: "Agentry Newsroom" editor: "Susanne Sperling, Editor — Human in the Loop" tools: ["Claude (Anthropic)", "Perplexity Sonar"] creativeWorkStatus: "verified" dateReviewed: "2026-10-18" aiActArticle50: "compliant" humanView: "https://agentry.news/launches/microsoft-agent-framework-adds-memory-checkpoints-in-september-release" agentView: "https://agentry.news/agent/microsoft-agent-framework-adds-memory-checkpoints-in-september-release"
Microsoft published expanded capabilities for its Agent Framework on September 24, 2026, introducing Foundry Agent Service integration with persistent memory retrieval, Azure Cosmos DB support, and ch
Drafted by an AI agent. Verified by Susanne Sperling, Editor — Human in the Loop. AI policy.
Microsoft published "What's new in Microsoft Agent Framework: Interactive experiences, memory, and resilient execution" on September 24, 2026, expanding orchestration and persistence capabilities for developers building agents on its Foundry platform Microsoft DevBlogs.
The update introduces Foundry Agent Service integration through a FoundryMemoryProvider, which retrieves relevant memories before an agent run executes and submits conversation information for asynchronous memory extraction Microsoft DevBlogs. This enables agents to maintain context across multiple sessions without retraining or manual state management.
Azure Cosmos DB support is now available through a Python-preview CosmosMemoryContextProvider, offering hybrid vector and full-text retrieval capabilities Microsoft DevBlogs. This allows enterprises to persist agent interactions in a managed database backend, supporting both semantic search via vectors and keyword-based lookups.
Microsoft's Agent Framework now integrates workflow checkpoints with resilient background responses, enabling agents to reload persisted state after recovery. However, the company explicitly warns that "recovery does not imply exactly-once execution of external effects"—a critical distinction for workflows involving financial transactions, API calls, or other operations where duplicated execution poses risks Microsoft DevBlogs.
The Foundry Toolkit will support long-running agents that can be disconnected and later reconnected to inspect progress or retrieve results without restarting the underlying work. This capability targets enterprise use cases where agents manage complex, multi-step tasks over hours or days.
The memory and checkpoint features address a core pain point in agent deployment: maintaining reliability and context when systems fail or sessions end. Teams building customer-service agents, workflow automation, or data-processing pipelines can now persist agent state and reuse learned context, reducing latency and redundant computation.
The caveat on exactly-once execution is particularly important for regulated industries. Financial services, healthcare, and supply-chain firms will need to implement idempotency checks at the application layer if duplicate external calls pose compliance or operational risk. Microsoft's documentation does not specify which external effects (API calls, database writes, webhooks) are most susceptible to duplication during recovery.
These updates reflect Microsoft's effort to move Agent Framework from proof-of-concept tooling toward production-grade infrastructure. Memory providers and checkpoint recovery are table-stakes for enterprises piloting agent deployments at scale.