Couchbase AI Data Plane cuts agent token usage by up to 45%
Couchbase has announced a significant expansion of its AI Data Plane, introducing native Amazon Bedrock integration with semantic caching, a managed Agent Memory service and a managed MCP server on AWS. The company says the new capabilities are designed to reduce both the cost and complexity of running AI agents in production at enterprise scale, with all features generally available from 20 October 2026.
The platform is positioned as a converged alternative to the fragmented stacks that many enterprise teams currently operate. Barry Morris, Chief Product and Strategy Officer at Couchbase, said: "Most are still running three or four separate systems for memory, vectors, operational data, and analytics, each carrying its own sync logic, security model, and failure mode. What we're shipping now collapses that stack."
Performance claims and benchmarks
Couchbase reports that semantic caching, delivered via its native Amazon Bedrock integration, can reduce LLM calls by 40 to 60% by reusing cached responses to semantically similar queries. Its Agent Memory component, benchmarked against competing offerings on the LoCoMo standard, is said to cut token consumption by 33 to 45%. Vector-driven context retrieval, meanwhile, is reported to produce prompts 70% smaller than full-context alternatives. Early enterprise adopters have told the company that agent deployment timelines have shortened from several months to weeks.
The managed Agent Memory service persists profiles, conversation history and vectorised facts across sessions, allowing agents to resume prior context rather than starting afresh on each interaction. The managed MCP server provides a governed, open-standard data access path and has been validated with LangGraph, CrewAI and LlamaIndex, giving engineering teams flexibility over their choice of orchestration framework.
A notable real-world reference is Mavenir, the telecom network software vendor, which has run subscriber profiles and live session state on Couchbase since 2017. Brandon Larson, SVP Product Management at Mavenir, said the company is extending that deployment to underpin AI agents for subscriber services and network operations, citing the platform's reliability and scale across millions of subscribers as the key reasons for the expansion.
Market context
The economics of production AI are emerging as a critical pressure point for enterprise technology buyers. McKinsey research cited by Couchbase suggests that moving AI from pilot to production can nearly quadruple costs as usage-based model pricing, growing workloads and more complex agentic workflows compound. Separate KPMG research in the UK found that approximately a third of business leaders identify limited visibility into AI usage costs as a barrier to deploying agents at scale.
Couchbase competes in an increasingly crowded operational and vector database market. Established players such as MongoDB and Redis, specialist vector-database vendors including Weaviate and Qdrant, and hyperscaler-native services from AWS and Google Cloud all offer overlapping capabilities. Couchbase's differentiation argument centres on running a single governed platform identically across public cloud, private data centre and edge environments, which resonates with regulated industries and multi-cloud enterprises concerned about data residency and operational consistency.
The EU AI Act's provisions for general-purpose AI systems, alongside the UK Government's emerging AI assurance framework, are increasing enterprise scrutiny of the governance and auditability of AI agent infrastructure. A platform that keeps memory, caching, vector search and operational data under a single security and access-control model addresses a growing compliance concern, particularly for customers in financial services, telecoms and critical infrastructure. Couchbase has not published detailed SOC 2 or ISO 27001 certification status for the new managed services, which procurement teams in regulated sectors will typically require before sign-off.