Skip to content
AI and operationsOctober 12, 20262 min · updated November 7, 2026

AgentCore moves short-term memory to consumption pricing

Daillac research
AWS: consumption-based short-term memory pricing, October 12, 2026
D
Written by
Daillac

Editorial and software engineering team

Photographic illustration of different storage drives on a workbench beside a digital scale
On October 12, 2026, AWS announced consumption-based pricing for short-term memory in Amazon Bedrock AgentCore. It replaces a flat event rate with dimensions tied to the data ingested, retrieved and stored. Long-term memory already existed, and its pricing remains unchanged under this announcement.

What changes in the calculation

The official announcement states that the pricing is available in all regions where AgentCore Memory is offered. Before revising a budget, check the service region and applicable units on the AgentCore pricing page.
For ingestion and retrieval, AWS applies a minimum billable size of 12 KB and a maximum billable size of 64 KB per event. Storage uses the event’s actual size. The configured time to live, or TTL, helps manage retention and therefore part of the storage cost. A billable-size boundary should not be read as a universal technical limit on content.

Measure usage before claiming savings

The proposed approach for a small business is to separate writes, reads and retained volume. Two agents handling the same number of conversations can have different profiles if one repeatedly retrieves history or keeps more information. Counting user calls alone therefore provides an incomplete basis for a useful cost estimate.
Prepare a representative sample with fictional data and record event size, frequency and retries. Compare retention policies that still meet the business purpose and applicable requirements. Earlier deletion may reduce storage while changing the information available later in the process. Record that trade-off so the operational owner can approve it alongside the budget.
Also distinguish this memory from a knowledge base and from the business system. In CRM automation, decide which information belongs in the customer record and which information temporarily supports the conversation. Copying every exchange into every layer makes access and retention harder to track.
A reasonable next step is to revise the cost model for one limited process, then compare observed usage and billing before expanding. For a custom web application, attribute those measurements to the relevant process and keep other costs separate. The announcement does not establish savings for a particular customer, nor does it supply test results for your own workload.
Sources & method

Synthesis of consulted official sources and original operational analysis; no product testing.

Share