Couchbase adds managed AI memory and Bedrock caching to cut agent overhead

Couchbase is promising cheaper AI agents by changing the services around the model, rather than the model itself. On October 20, 2026, it made new AI Data Plane capabilities generally available, including Amazon Bedrock integration, managed memory and expanded search. Its headline reductions in model calls and token use are company claims.

Couchbase’s pitch is that customers can replace separate memory, search, analytics and operational-data services with one governed platform.

Reusing answers instead of calling the model again

The expanded Model Service gives customers serverless, pay-per-token access to hundreds of foundation models through Bedrock, plus vectorization over private networking. Couchbase says teams can run inference and generate embeddings—the numerical representations used for similarity search—without separately managing GPU instances.

Semantic caching supplies the cost-saving mechanism: when a query is similar in meaning to one already answered, the application can serve a cached response rather than invoke the model again.

There is a correctness tradeoff. In a SiliconANGLE guest column, analyst Zeus Kerravala warned that a cache can return a similar answer to a genuinely different question. He recommended tuning and monitoring similarity thresholds, and said savings depend heavily on how repetitive the workload is.

Memory moves from self-managed to managed

Agent Memory and Couchbase’s enterprise-supported MCP server launched as self-managed capabilities in June. This release makes them fully managed cloud services on AWS, shifting the work of standing up and operating those services away from customer teams.

Agent Memory retains profiles, conversation history and vectorized facts across sessions. The managed Model Context Protocol server supplies an open-standard connection through which agents and tools can access Couchbase data.

Couchbase says both services have been validated with LangGraph, CrewAI and LlamaIndex, letting teams switch or combine agent frameworks without rebuilding their memory layer.

Searching by meaning, then calculating the result

Operational Insights now combines vector search—finding information by similarity—with structured analytical queries against changing operational data. Couchbase says teams can select records by meaning, then aggregate the results in one SQL query, without a separate data-transfer pipeline or another system to keep synchronized.

The company’s example is revenue by region filtered to clusters of frustrated customers. Its write-back function sends query results into the operational data that applications and agents already read.

Couchbase also says vector-driven context retrieval produces prompts 70% smaller than full-context alternatives. That measures prompt size, not a 70% reduction in total agent operating costs.