Microsoft adds native cross-session memory to Agent Framework via Cosmos DB

Microsoft's addition gives Agent Framework agents durable memory that survives across sessions, backed by Azure Cosmos DB as the storage layer. The developer experience is deliberately simple: install the agent-framework-azure-cosmos-memory Python package, attach a CosmosMemoryContextProvider to an agent once, and memory is automatically persisted and rehydrated beyond a single conversation — no bespoke persistence plumbing required.

Using Cosmos DB is a natural fit: its low-latency, globally distributed, schema-flexible design suits the read/write patterns of agent memory, and it plugs into Microsoft's broader Azure data ecosystem for governance and scale. This lets enterprises keep agent state within their existing Azure footprint rather than adopting a separate memory store.

The timing is the headline: this ships the same day as AWS AgentCore memory lifecycle guidance and NVIDIA's NemoClaw framework — three hyperscalers/vendors independently converging on persistent agent memory in a single 24-hour window. That convergence confirms the theme of the week: 2026's agent race has moved past raw model capability to the surrounding infrastructure (memory, identity, freshness) needed to run agents reliably in production.

Competitively, Microsoft's advantage is integration — memory as a native, one-line-attach capability inside Agent Framework, backed by a managed Azure database — which lowers adoption friction for teams already on Azure. As a preview, the caveats are maturity and cost: Cosmos DB pricing at scale, how memory is indexed and retrieved efficiently, and whether the provider offers the scoring/pruning governance AWS emphasized. Readers should watch the retrieval-relevance and cost characteristics as the preview matures toward GA.