Couchbase Adds Managed Agent Memory and Semantic Caching to AI Data Plane

Couchbase expanded its AI Data Plane on Oct. 20 with managed agent memory and semantic caching, giving companies ways to reduce repeated model requests and context loading as they run AI agents.

In Couchbase's semantic caching tests, reusing stored results for similar requests cut large language model calls by 40% to 60%. Agent Memory used 33% to 45% fewer tokens than leading alternatives on the LoCoMo benchmark. Vector-based retrieval generated prompts 70% smaller than full-context alternatives.

The figures measure different parts of an agent's workload: the number of model calls, the tokens used by memory and the size of prompts assembled through retrieval. They do not represent a single estimate of operating-cost savings for customers.

Semantic caching lets an agent reuse a saved result instead of sending a similar request to a model again. Agent Memory retains profiles, conversation histories and facts between sessions, reducing the need to rebuild that context for each interaction.

The update makes Agent Memory available as a managed service on Amazon Web Services. A managed Model Context Protocol (MCP) server gives agents a standard way to access Couchbase data.

Couchbase's Model Service also gained native Amazon Bedrock integration alongside semantic caching. Bedrock model selection was already available in Capella iQ when Couchbase introduced the AI Data Plane on June 30; the October integration applies to a different part of the platform.

Couchbase also added vector search to Operational Insights, its analytics product. Teams can combine similarity searches with structured analysis in one SQL query. Operational Insights 3.0, released in October, includes a CLUSTER BY clause for grouping embeddings by semantic similarity.

The June launch included self-managed versions of Agent Memory and the MCP server. Barry Morris, Couchbase's chief product and strategy officer, described the newer release as bringing services together: “What we're shipping now collapses that stack.”

Couchbase announced the expansion at 9 a.m. ET (13:00 UTC) on Oct. 20. The new capabilities became generally available that day.