Couchbase adds AI tools to cut agent costs & complexity

Couchbase has expanded its AI Data Plane with new services to run AI agents in production, aiming to cut operating costs and simplify deployment.

The additions include native Amazon Bedrock integration with semantic caching, managed Agent Memory, a managed MCP server on AWS, and vector search in Couchbase Operational Insights.

The services bring memory, vector search, operational data and analytics together on one platform. This reduces the need to move and synchronise data across separate systems used to build and run AI agents.

The expansion comes as businesses face rising costs when AI projects move beyond the pilot stage. Couchbase cited industry research showing that business leaders remain concerned about unclear AI agent usage costs and the impact of usage-based pricing as workloads grow.

Cost control

Couchbase said semantic caching can reduce large language model calls by 40% to 60%. Benchmarking against leading alternatives using the LoCoMo standard showed that Agent Memory cut token usage by 33% to 45%, while vector-driven context retrieval produced prompts 70% smaller than full-context approaches.

More than half of deployments that consolidated fragmented data services onto a single platform reported total cost of ownership reductions exceeding 50%, according to Couchbase. Early adopters also reduced AI agent deployment times from several months to weeks.

Barry Morris, Chief Product and Strategy Officer at Couchbase, said architectural complexity remained a challenge for enterprise AI projects.

"Customers building production AI agents at scale need a converged data layer, but 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," Morris said.

"What we're shipping now collapses that stack. The new capabilities all sit on one governed platform that runs identically in the public cloud, a private data centre or at the edge. As a result, Couchbase customers have fewer systems to connect and maintain, which helps them lower running costs and get their agents into production sooner."

AWS integration

The Amazon Bedrock integration provides pay-per-token access to foundation models and uses semantic caching to reuse responses to similar queries. This can reduce repeated model calls and response times.

Managed Agent Memory stores profiles, conversation histories and vectorised facts across sessions, allowing agents to retain context between interactions. The managed MCP server on AWS lets applications and agents access Couchbase data through an open standard without requiring customers to manage the underlying infrastructure.

Both services are framework-agnostic and have been validated with LangGraph, CrewAI and LlamaIndex. Engineering teams can then switch between orchestration frameworks or use several together without rebuilding the memory layer.

Telecoms use case

Mavenir has used Couchbase for subscriber profiles and live session state since 2017. The telecoms group is now applying the platform to AI agents supporting subscriber services and network operations.

Brandon Larson, Senior Vice President of Product Management at Mavenir, said reliable access to subscriber and session data was essential to maintaining service continuity.

"Couchbase is a big reason a server failure doesn't drop a subscriber's call," Larson said.

"Now we're building AI agents that act for subscribers and help run the network, and those agents are only as good as the context they can reach. We're building that context layer on Couchbase for the same reasons we built our packet core on it. It's fast, it stays up, and it scales to millions of subscribers."

Couchbase has also added vector search to Operational Insights, alongside structured queries on operational data. Teams can combine semantic analysis with standard SQL-based aggregation in a single query, without moving data through a separate extract, transform and load (ETL) pipeline or maintaining another synchronised system.

This allows businesses to analyse measures such as revenue while filtering results by themes or intent in unstructured data. Query results can also be written back to the operational database for use by applications and AI agents.