Couchbase ships an AI Data Plane for production agent memory

Couchbase has launched an AI Data Plane, framing the product as the operational data layer that production AI agents run on top of, with persistent Agent Memory, an Agent Catalog for tool discovery, and an enterprise-supported Model Context Protocol server. The release lands at a moment when most enterprises still run agents on stitched-together caches, vector stores, and document databases, and when platform teams are starting to feel the operational cost of that fragmentation.

What the AI Data Plane ships

The product combines three pieces. Agent Memory is a persistent memory layer designed to hold conversational context, structured operational data, and cross-session state for production agents in a single service. The Agent Catalog is a discoverable registry for the tools agents can call, intended to replace the ad-hoc tool inventories that platform teams have been hand-maintaining. The third is an enterprise-supported, self-managed MCP server for standardized integration of Model Context Protocol clients, which is the part of the stack that lets agents from different frameworks talk to the data plane without each team writing their own connector. The AI Data Plane runs across both Capella, Couchbase's managed cloud, and self-managed environments, and is built on the same distributed multi-model engine that already handles JSON documents, key-value, SQL for JSON queries, full-text search, eventing, and vector search in a single system. The pitch is that an enterprise can keep its operational data on the platform it already runs, add an agent-aware data layer on top, and avoid the integration tax that has been slowing production agent deployments for the last year. IDC's Devin Pratt, research director for AI, automation, data, and analytics, framed the announcement as the database industry catching up with how agents actually use data, noting that 80 percent of agentic AI use cases will need real-time, contextual, and widely accessible data.

Why agent memory belongs in the data platform

The most distinctive piece of the launch is Agent Memory, and the case for it is straightforward: simple agents can succeed with vector search alone, but production-grade agents need to store conversational context, retrieve structured operational data, and maintain state across sessions and restarts, all with sub-millisecond latency at the point of decision. Most agent stacks today wire those three requirements onto three different services, and the operational cost of that fragmentation is showing up in incidents where the agent loses context, where memory writes race with retrieval, or where the cache and the source of truth drift out of sync. Couchbase's argument is that the agent memory layer belongs inside the operational data platform rather than alongside it, because the agent is, functionally, the application that reads and writes the database. The framing lines up with the broader enterprise AI infrastructure pattern documented in the AI Infrastructure in 2026 resource page, which lays out why operational data layers have become a first-class concern for agent deployments. The company is also framing Agent Memory as framework-agnostic, validated with LangGraph, CrewAI, and LlamaIndex, so engineering teams can switch orchestration frameworks without rebuilding the memory layer underneath. The point of the validation list is that platform teams do not want to be locked into one agent framework just because their memory store is tied to it, and that choice matters because most enterprises are still in the middle of picking their orchestration stack. The MCP server piece is the second part of the same argument, since most agent frameworks are converging on Model Context Protocol as the standard connector layer, and a database vendor that ships an enterprise-supported MCP server is positioning itself as the data side of that protocol rather than as a passive participant.

Lakehouse federation and the agentic enterprise

The data plane launch lands alongside Enterprise Analytics 2.2, a major expansion of Couchbase's analytics capabilities that opens operational data to the broader lakehouse ecosystem through Apache Iceberg federation. The new release adds Google Cloud Storage support, JWT authentication, Oracle and SQL Server change data capture, asynchronous long-running queries, an index advisor, index-only query plans, SQL++ UPDATE support, and SDK updates across Java, .NET, Python, JavaScript, and Go. A new Trino adapter, expected in Q3 calendar 2026, will give enterprises in-place SQL access to Couchbase operational data from Trino-based platforms including AWS Athena, Amazon EMR, Google Dataproc, and Starburst, removing the need to extract and replicate live data into separate analytical stores before querying. Capella iQ, Couchbase's natural-language query assistant, also gets multi-model provider selection with AWS Bedrock and OpenAI, governed by organization-level policies so administrators can control which models are available to which teams and keep inference costs and data residency requirements within organizational guardrails without slowing down individual developers. The release is broad, but the core message is that the AI Data Plane is not a separate product so much as the next layer of an existing platform that already runs in production for tens of millions of transactions per second with sub-millisecond latency, and that the Trino adapter is the piece that ties the operational and analytical stories together (the CXOToday coverage carries the IDC and Agora quotes in full). The announcement lands in a market where the phrase "agentic enterprise" is starting to mean an enterprise that has collapsed its agent runtime, its data layer, and its governance model into a single operational surface, so that platform teams can answer the same questions about an agent that they can answer about a microservice: who owns it, what data does it touch, what does it do in production, and what does it cost to run. Couchbase is one of several infrastructure vendors now competing for that framing, and it lands the same week that Databricks shipped Genie One for data-smart AI coworkers, another sign that the agentic enterprise story is becoming a database-platform story as much as a model story, and the AI Data Plane is its bet that the agent memory and context retrieval question is going to be answered by the operational data platform rather than by a separate vector store or cache. For enterprises that have already standardized on Couchbase for operational workloads, the release is the path from chat-style pilots to production-grade agents without bolting a second data layer onto the side, and for everyone else it is one more signal that the next round of agent infrastructure investment is going to land on the data side of the stack rather than the model side.