Couchbase launches AI Data Plane to deliver persistent agent memory from cloud to edge

Couchbase brings persistent agent memory to the edge | VentureBeat

The competitive edge in enterprise AI is shifting to context: which platform can give an agent the right memory, the right retrieval and the right data at the moment of decision. Couchbase on Tuesday announced its AI Data Plane, combining persistent agent memory, real-time context retrieval and an enterprise-managed MCP server in a single operational platform. Couchbase's roots are in caching and high-transaction databases -- an architecture the company argues makes it better suited for agent memory than vendors that came to the problem from search or analytics. The AI Data Plane runs identically across cloud, on-premises and disconnected edge environments, extending agent memory and local vector search to devices with no network connection. "How do you make sure that the intelligence that you get out of these models are the ones that databases specialize in?" Gopi Duddi, CTO at Couchbase, told VentureBeat. "How can you get that value out of storage systems, which are still going to be databases?" What the AI Data Plane delivers The AI Data Plane packages three components designed to replace the fragmented stacks most enterprises are currently running. Agent memory: A unified persistence layer for conversational context, structured operational data and vector embeddings. Couchbase says the guardrails are what distinguish it from standalone memory services: token constraints per session, time-to-live limits on stored memories and metering controls that cap compute consumption per agent session. Enterprise MCP server: An enterprise-supported self-managed server for standardized model-context protocol integration, shipping as part of the platform rather than requiring a separate service. Agent catalog: A function-level catalog of discoverable agent tooling built by Couchbase. Duddi distinguished it from metadata catalogs like Databricks Unity or AWS Glue -- describing it, in his words, as closer to a glorified MCP that surfaces agent functions as callable tools within the platform. Memory-first architecture takes agent context to the disconnected edge The lineage of Couchbase and its core architectural foundation is what Duddi says gives it an edge when it comes to context. "We were a cache before we became a database," Duddi said. Writing to memory is 10x faster than writing to disk, Duddi said -- a speed advantage he argues separates Couchbase from NoSQL databases that layer memory workloads on top of disk-based storage. Couchbase isn't the only data technology that has its roots in a caching layer. Redis similarly is rooted in cache and also recently announced an agentic AI context layer. Duddi argued that Couchbase is different in that it maintains an ACID (Atomicity, Consistency, Isolation, and Durability) compliant database which matters for transactional workloads. Couchbase also has a long history across multiple deployment modalities. That architecture extends to the edge through Couchbase Lite, the platform's on-device runtime. It runs SQL, full-text search and vector search locally without a network connection, using a proprietary sync mechanism to replicate bidirectionally back to cloud or between edge nodes when connectivity returns. The target environments are retail floor operations, field service, industrial deployments and regulated settings where agent data cannot leave the device. Duddi cited hotel reservations as an early example: multiple agents serving customers concurrently, each pulling local context and running vector search on-device, with shared session memory synchronizing centrally. The practical benefit is token efficiency. Rather than every agent independently retrieving and processing the same data, the platform caches shared context so concurrent sessions draw on it without burning tokens repeatedly. Agora's view from production Agora, a platform that helps developers embed real-time voice, video and conversational AI into enterprise applications, has run Couchbase in production since February 2024. The initial use case was its Signaling product, managing channel setup and state synchronization for live calls. Expanding into conversational AI agents brought stricter requirements: memory-first architecture, full JSON support for storage and query, cross-datacenter replication for high availability and enterprise-grade vendor support. "Couchbase was the best fit based on these criteria," Patrick Ferriter, SVP of Product at Agora, told VentureBeat. Agora is now extending that relationship to support context retrieval for conversational AI agents. "This will simplify the architecture and deliver enterprise grade RAG with predictable lower latency required for conversational AI use cases," Ferriter said. For data professionals trying to figure out the best approach to context, there is no one answer. On platform selection, Ferriter was direct. "It depends on the preference and goals of the organization, including timing," Ferriter said. "If they want something enterprise grade and optimal for immediate production and scale vs. having to optimize and maintain an open-source solution with community support. We wanted the former and that is why we looked at an expanded partnership with Couchbase." Competitive context: following the right trend The context layer has become a crowded space in 2025. Oracle put a memory core in its database back in March providing a context layer. Redis added a context layer in May as did vector-native database vendor Pinecone. "Couchbase is following this trend, not setting it, but it's the right one to follow," Devin Pratt, Research Director for AI, Automation, Data and Analytics at IDC, told VentureBeat. "Its real edge is reach, running the same platform from cloud to edge to mobile, which is how enterprises actually operate. The test now is to scale against bigger names." For teams navigating the vendor landscape, Pratt's framing is direct. "Match the tool to the workload. Consolidate where it makes sense, use a specialized engine like a graph database where relationship-heavy reasoning earns it, and let governance drive the call rather than treating memory as plumbing," Pratt said.

Couchbase AI Data Plane tests whether NoSQL's cloud-era strengths can carry into agent memory

Oh the halcyon days of the early cloud. Arguing whether operating at global scale was just an extension of what we already did within the enterprise -- private cloud, application servers, relational databases but just a bit... bigger -- or something that required fundamentally different practices -- public cloud, serverless infrastructure, key-value stores. People fighting it out online and willing to shed intellectual blood to metaphorically die on one hill or the other. And as usual the disappointingly mundane answer was a bit of both. But one of the areas that did see a huge blossoming of innovation was data management -- because it turned out that the challenge of serving a billion external users wasn't the same as serving a few thousand internal ones. Assumptions around co-location, vertical scaling, schema stability, ACID transactions, and response time no longer held. Applications had to be distributed across machines, data centers and regions. They had to keep running while data structures changed. And they had to deliver fast responses even when the underlying systems were spread across the world. SQL databases were not created for that world -- which is what created the opening for the NoSQL movement to emerge. So the broad lesson is: shift in scale and complexity -> shift in practice and architecture. And Couchbase was one of the companies at the vanguard of that cloud-era shift. But as the world turns again, we find ourselves at the start of another shift -- AI agents. A shift that will likely bring a host of new challenges as billions of agents start to analyze, create, and remember via trillions of continuous data retrievals and updates -- with often partial, fragmented, or rapidly changing data -- at inhuman speed and scale. And the core contention in Couchbase's new AI Data Plane is that as more and more agents move from pilots to production, we won't only need to provide persistent memory -- but persistent memory that is fast, scalable, tolerant of ambiguity, and easily distributed to where it is needed. All of which, to Couchbase at least, appears to rhyme with its previous experience -- and lead it to argue that the data challenges of enterprise AI agents may not be that different from the data challenges of distributed clouds. And so in launching its AI Data Plane, Couchbase is effectively pushing the idea that agent memory and context management need to shift from artisanal, ad hoc combinations of files, documents and databases to become foundational data infrastructure -- and that the capabilities it built for the last shift -- fast storage, straightforward scaling, unified access, and increasing distribution -- are the best foundation for the next one. Turning agent memory into enterprise infrastructure In Couchbase's telling, agent memory is currently a barrier to production agents because it still relies on a patchwork of files, databases, and ad-hoc system integrations chosen on a project-by-project basis -- something pilot agents can get away with, but which rapidly spirals out of control as agents multiply across workflows, customers and operational systems. Because production agents need more than just data in the moment. Instead the memory they create needs to survive restarts, cross sessions, remain accessible to future agents, and avoid being trapped inside whichever orchestration framework happened to be used when the project began -- especially in a field as fast-moving as AI. And so the company argues that for agents to scale, memory management needs to become real enterprise infrastructure -- which is why it has put Agent Memory at the center of its AI Data Plane launch. Because its point is not just that agents need somewhere to store things -- but that the memory layer they use to do so needs to be sufficiently resilient, scalable and independent to underpin sustainable, long-running production ecosystems. What Couchbase is not doing, however, is pretending that it is also going to be a full agent stack. Instead the company has had its Agent Memory compatibility verified by a range of popular agent harness frameworks -- including LangGraph, CrewAI and LlamaIndex -- in a sign that it believes that memory infrastructure has a defensible value all of its own. And this makes sense given that many organizations are also starting to understand the interplay between model, harness, and memory -- and the need to protect institutional memory and operating context over the long term as agents absorb more work. In that sense, Couchbase is trying to position itself in the layer below the noisy agent framework debate -- not claiming that it will decide how every agent is built, but that customers should be free to change how those agents are built without having to either suck it up or lose and rebuild their institutional memory every time they fall out with a vendor. Which means that, in simple terms, Couchbase wants to be the plumbing for agents -- a way to store, transfer, and access the continuous flow of data necessary to sustain themselves. A critical but unremarkable infrastructure that is rarely considered but always there when you need it. Because Couchbase's argument is that if enterprises build enough agents, memory can no longer be something teams constantly have to worry that they must touch, tweak, extend or repair with each new agent -- instead it needs to become another kind of unremarkable infrastructure that is always there when they need it. Which is interesting -- but also opens another question. If memory infrastructure is plumbing then how do we make sure it has the capacity necessary to support increasing demand? Making agent data infrastructure scale In any growing town infrastructure always faces pressure -- and if the sometimes-breathless industry predictions of agent growth hold true then, boy, could agent town explode. And so it feels like our agent data infrastructures need to be designed from the outset to cope with that growth -- and the speed, scale and intensity of data access that will result as the number of agents increasingly outstrips the number of people. But if you follow Couchbase's argument through, the issue is not simply that agents need scalable access to more data -- which is already a significant challenge -- but that agent memory is more like a living fabric that is constantly accessed, updated and synchronized at a cadence ordinary enterprise systems were never designed to support. Because agents will be distributed, working with messy context, and constantly reading, writing and overwriting things to memory in ways that clash, overlap and diverge -- effectively using storage not only for data at rest but as a scratchpad for processing, building, and remembering. And that likely means lots of small, fast, stateful interactions happening continuously -- which, to be fair, gives Couchbase's attempt to transition into the agent memory market an understandable logic. Because when viewed side-on, the challenges of supporting agent memory and context management can indeed start to look suspiciously like the kind of problems that cloud-era NoSQL platforms were built to solve -- albeit with yet another step up in scale and intensity. And when viewed through that lens, Couchbase's memory-first, scale-out, cloud-era architecture -- which it claims already supports demanding operational workloads with sub-millisecond latency -- at least looks like an interesting foundation for the agentic era. The question, however, is how far that adjacency will really stretch -- and whether enterprise agents just represent more of the same or a genuinely new pattern of access, storage and memory manipulation -- one that will inevitably diverge over time from the cloud-era problems Couchbase already understands. But whatever the answer to that question, scale is only one part of Couchbase's argument -- because even if you can solve the issue of scale, the problem of fragmentation potentially remains, because different kinds of data are currently served by different forms of storage. Consolidating agent data infrastructure One of the perhaps less obvious points Couchbase raises in its AI Data Plane push is the increasing challenge of agent data integration. Because when agents rely on a variety of fragmented data sources -- such as separate vector stores, file systems, document repositories, operational data systems and reporting layers -- then agent memory and context can quickly become an integration problem. Couchbase's answer is therefore to collapse some of that sprawl back into its consolidated Data Plane -- arguing that, as much as possible, these data sources should be accessed through a common layer to avoid making every agent implementation an energy-sapping integration problem. Instead the company argues that the core data services agents repeatedly depend on should be managed together. That they should become plumbing. And while this may be an argument that finds favor within teams who spend more time trying to connect different infrastructures than they do building agents, it is also a faintly self-serving argument because this kind of consolidation has long been an aspect of Couchbase's strategy. Its distributed multi-model architecture already brings JSON documents, key-value access, SQL for JSON queries, full-text search, eventing and vector search into one system -- which means pushing the idea that an AI Data Plane should offer this range of capabilities feels like part argument and part canny competitive positioning. But it also doesn't make it wrong. And it has the advantage of enabling the company to present Agent Memory less as a new bolt-on than as an extension of a set of capabilities it has already been telling customers belong together. Which is why the stronger part of Couchbase's argument feels categorical rather than technical -- that vectors, documents, caches and operational data become more useful to agents -- and less overhead for teams -- when they do not have to be wired together from scratch every time. But this is still not the whole picture -- because while bringing those services closer together may complement Couchbase's integration play, it still leaves one significant question unaddressed -- where agent memory itself needs to live when agents themselves are distributed by nature. Taking agent data to the edge So one thing I think it's increasingly safe to predict is that we aren't heading towards a world in which there are only three or four global US AI model companies that are used by everyone. I mean even if that ever looked likely -- before other regions started dropping competitive open weight models -- increasing geopolitical pressures, and the realization that models you don't own can always be taken away from you without notice, have likely put the final nail into that particular coffin. Which means agents won't only run in the cloud -- but in our data centers, employee workstations, mobile devices, robots, vehicles, specialist equipment and edge servers. Or to put it more simply they'll run everywhere -- and there will be no single source of omniscient agent memory sitting in a data center at the heart of the world. And so in each of those environments the same basic questions return -- what data does the agent need, how quickly does it need it, how much can be stored locally, and how should state be synchronized with the wider enterprise as memory and context are manipulated? Couchbase's answer is to fold its existing distribution, caching, and synchronization capabilities -- Couchbase Lite, Edge Server, React Native support and Sync Gateway -- into its AI Data Plane so that agents running in mobile and edge environments can manage memory and perform local vector search, even in situations of intermittent or missing connectivity. Because if memory and context need to follow agents into distributed environments then, Couchbase suggests, synchronization, local access, and edge controls must go with them. This doesn't mean that Couchbase's synchronization and edge data capabilities can answer the higher-level design question of which context should move, how much of it should move, how long it remains useful, or how local memory should connect back to wider systems of knowledge -- and the company doesn't claim that it does. Which means the edge story is best read as enablement rather than solution -- because while Couchbase can help put memory closer to where agents operate, the harder architectural choices around context design, data selection and memory boundaries still sit with you. My take As someone who is (tragically) old enough to have lived through several waves of technology shift and expansion, I can't help but return to my statement from the introduction. That any shift in scale and complexity generally means a shift in practice and architecture. And in this sense, Couchbase's AI Data Plane is most interesting when viewed as a category bet that its existing practices and architecture can be adapted for the agentic era, rather than simply as a product bundle. Strip away the launch vocabulary and the company is making a relatively clear argument -- that the memory layer beneath production agents will become a specialized operating infrastructure in its own right. And Couchbase is making a play to suggest that the flexible, distributed and low-latency NoSQL capabilities it built during the cloud era give it a credible story for transitioning into that market. And so the most interesting question for me in that context is whether Couchbase is seeing the next infrastructure layer early because it has a head start -- or is simply interpreting the early signs of a new category through the biases of the tools it already has. Because historically, NoSQL databases themselves did not eliminate SQL. They took over emerging workloads that didn't exist when the older technology was developed -- and simply became part of a broader evolving data landscape. And so there's a chance that something similar may happen here. Essentially if agent memory and context lead to new challenges that diverge from the cloud-era problems NoSQL vendors already understand then it could perhaps leave space for entirely new categories -- such as context graphs -- to form around the higher-level problem of making agent context meaningful. Which is where buyers should evaluate Couchbase's story with the bigger picture of agent context in mind. Couchbase certainly offers strong narrative coherence in positioning itself as infrastructure for storing, retrieving, synchronizing and distributing agent memory -- and even has a credible sounding claim to be a key player based on what we know today. But the storage layer is not the whole system of meaning agents will need to act well and so enterprises will need to evaluate the company's narrative and capabilities within the wider evolving context. Even so, Couchbase's AI Data Plane narrative points to something real -- the new scaling crisis enterprises will have to confront as they move agents beyond pilots and into operational use. And in doing so the company offers a credible argument for why the infrastructure it built for the last scaling crisis might also be the one that helps its customers overcome the next.

Couchbase Launches the AI Data Plane, the Operational Data Foundation for the Agentic Enterprise

Unified data layer from cloud to edge gives production AI agents the governed memory, context, and performance they need to remember, reason, and act on real-time context Couchbase today announced general availability of the AI Data Plane, a unified data infrastructure layer for enterprise AI agents. The Couchbase AI Data PlaneTM gives enterprises persistent agent memory, real-time context retrieval, and consistent data access from cloud to edge and into their lakehouse architectures. By collapsing the fragmented data services that have stalled agent deployments into a single, governed layer, enterprises can move from pilots to production-grade agents that deliver more consistent decisions, richer customer experiences, and measurable efficiency gains at scale. The AI Data Plane unifies Agent Memory, an Agent Catalog for discoverable agent tooling, and an enterprise-supported, self-managed MCP server for standardized integration of model-context protocols. It consolidates previous Couchbase deployment models into a single architecture that runs across Couchbase Capella and self-managed environments, and is complemented by new Enterprise Analytics 2.2 capabilities for Apache Iceberg-based lakehouse federation in addition to a Trino adapter (currently expected to launch in Q3 calendar 2026). The enterprise-supported platform is backed by Couchbase's engineering and support organization, eliminating point solutions and giving platform teams a single operational surface for the data services their agents depend on. "Most enterprises quickly discover that moving from chat-style pilots to production-grade agentic systems is really a data problem, not just a model problem," said Devin Pratt, Research Director, AI, Automation, Data & Analytics, IDC. "IDC expects that 80% of agentic AI use cases will require real-time, contextual, and widely accessible data, so the architecture has to support that. Approaches that make agent memory and context retrieval first-class capabilities of the database itself, like Couchbase's AI Data Plane, address this directly. By unifying vectors, documents, cache, and operational data in a single distributed platform, from cloud to edge, Couchbase reduces the integration tax that has been slowing down real-world agent deployments and gives organizations a more governable, scalable foundation for the next wave of AI-powered applications." Persistent Agent Memory CIOs shaping their AI infrastructure strategies need a unified data platform that governs memory, context, and retrieval across the full agent lifecycle - not another point solution to integrate and maintain. Couchbase Agent Memory delivers this by providing a unified persistence layer as a single service within the operational data platform instead of forcing teams to stitch together separate caching, vector, and document stores. Because it is framework-agnostic and validated with LangGraph, CrewAI, and LlamaIndex, engineering teams can switch or combine orchestration frameworks without rebuilding their memory layer. As enterprises move from prototypes to production agents, the gap between what agents can reason about and what they can remember across sessions has become a critical bottleneck. Simple agents can succeed with vector search, which Couchbase offers at billion scale, but production-grade agents require the ability 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. "What matters most for enterprise-grade conversational AI agents is that data retrieval is fast, consistent, and seamless. When you're running human-to-AI agent interactions, everything behind the scenes needs to be predictable and consistent to provide natural interaction," said Patrick Ferriter, SVP of Product at Agora. "That's what we're solving together with Couchbase, and it's why we chose them as a partner for the data layer for our conversational AI platform. Every one of our conversational AI use cases requires efficient data retrieval to feed the pipeline for AI agents, whether that's outbound sales, customer service, physical AI, or something entirely new. We've had a multi-year relationship with Couchbase, and as we've scaled into agentic workloads, this was a natural extension to our partnership." Built for Agentic AI at the Edge The shift from single-prompt AI applications to multi-step, autonomous agent architectures has exposed a fundamental mismatch between how agents work and how most data infrastructure is built, especially at the edge. Agents operate across sessions, accumulate context over time, and must act on structured operational data and unstructured embeddings simultaneously across cloud, edge, and devices. The Couchbase AI Data Plane is architected to meet this full set of data requirements. Throughput is equally critical, since agentic workloads can demand orders of magnitude more from the data layer than traditional applications do. Every agent action triggers context retrieval, memory writes, and state synchronization in rapid succession, often across thousands of concurrent sessions. The AI Data Plane is engineered for this scale, leveraging Couchbase's proven scale-out memory-first architecture, which already supports tens of millions of transactions per second with sub-millisecond latency for some of the world's most demanding enterprises. The AI Data Plane builds on Couchbase's distributed multi-model architecture, which supports JSON documents, key-value, SQL for JSON queries, full-text search, eventing, and vector search in a single distributed system. Agent Memory extends this foundation with session persistence and context retrieval, while the MCP server and Agent Catalog provide the integration and observability layers required for production agent deployments. Enterprise Analytics 2.2: Lakehouse Federation with Iceberg and Trino Couchbase also announced Enterprise Analytics 2.2, a major expansion of its analytics capabilities that opens operational data in Couchbase to the broader lakehouse ecosystem while strengthening the query engine itself. Enterprise Analytics 2.2 introduces Apache Iceberg lakehouse federation, allowing teams to query real-time operational analytics from Couchbase alongside existing open Iceberg-based lakehouse tables without complex ETL or data duplication. This lets enterprises adopting Iceberg for its open governance, performance, and ecosystem benefits derive more value from those investments by integrating Iceberg tables into the same platform that serves their agentic workloads. Core analytics enhancements include 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 corresponding SDK updates across Java, .NET, Python, JavaScript, and Go. These enhancements give platform teams faster, more governed analytics across their existing tools and languages, so they can build and optimize AI and data workloads without adding infrastructure complexity. A new Trino adapter, expected in Q3, will provide in-place SQL access to Couchbase operational data from Trino-based platforms including AWS Athena, Amazon EMR, Google Dataproc, and Starburst. This eliminates the need for enterprises to extract and replicate live data into separate analytical stores before querying when building AI and analytics workflows that span operational and lakehouse environments. Capella iQ Enhancements Capella iQ, the platform's natural-language query assistant, now supports multi-model provider selection with AWS Bedrock and OpenAI, governed by organization-level policies. Administrators can control which models are available to which teams, so inference costs and data residency requirements stay within organizational guardrails without slowing down individual developers. This gives teams the flexibility to choose the right model for each workload while administrators keep inference costs, compliance, and data residency tightly controlled through central policies. Edge, Mobile, and Distributed Application Updates As AI agents become part of the operational workforce, their data needs to follow the work, which increasingly happens on devices, in the field, and at the network edge rather than behind a desk or in a data center. Couchbase extends the AI Data Plane to the edge, so agents running in mobile and edge environments can access replicated data and perform local vector search, even with intermittent or no connectivity. New innovations include: * Couchbase Lite (CBL) 4.1: Native, out-of-the-box peer-to-peer sync over Bluetooth with automatic switching to Wi-Fi for more reliable collaboration in disconnected edge environments. A modernized Android API with native Kotlin @Serializable support removes boilerplate data mapping and enables efficient, reactive UI updates through direct serialization and delta-based change detection, while new C++ API bindings simplify building high-performance embedded applications. * Edge Server 1.1: Client-level access control for fine-grained local permissions, CORS support for browser-based edge apps, simplified credential rotation for distributed device fleets, and expanded platform support for Windows and ARM architectures. * React Native 1.1: Enterprise-grade support with Turbo Module integration gives cross-platform mobile teams direct access to Couchbase Lite performance without bridging overhead. * Sync Gateway 4.1: Cloud-to-edge data synchronization gateway for non-disruptive rolling upgrades and concurrent distributed resync for high-volume workloads, reducing operational overhead for enterprise deployments. Available as a managed service through App Services. "The database layer is where agentic AI either scales or stalls, and most of the industry is still treating agent memory as an afterthought," said Barry Morris, Chief Product and Strategy Officer at Couchbase. "We built the AI Data Plane because our customers told us that stitching together separate vector, caching, and document stores for every agent was the single biggest drag on their production timelines. Agent Memory gives them a unified, framework-agnostic persistence layer that operates identically in cloud and self-managed environments from cloud to edge, and runs at the latency their agents actually need. That's what it takes to move from pilot to production -- and the vendors who understand this will define the infrastructure category for the next decade of AI."