AWS has changed how Amazon Bedrock AgentCore charges for short-term memory: developers now pay by the data agents ingest, retrieve and store, rather than a flat rate per event. The new rates give teams concrete numbers to model, though the actual bill will depend on each agent’s usage.
AWS AI Watch analysis
What happened
AWS says short-term memory ingestion costs $1 per GB, retrieval $0.20 per GB and storage $0.10 per GB-month. Ingestion and retrieval are billed per event, with a 12 KB minimum and 64 KB maximum billable size; storage is billed on each event’s actual size.
The change applies in all AWS Regions where AgentCore Memory is offered. AWS says long-term memory pricing is unchanged and that most workloads should cost less under the new model, with the largest reductions for small-payload use cases. Developers can also set a shorter time to live, or TTL, to manage storage costs.
Why it matters
Per-data pricing could suit agents that write frequent, small events better than a charge applied uniformly per event. But the minimum billable size means tiny ingestion and retrieval events are not billed as tiny as they look, and storage keeps accumulating for as long as data remains stored. The right comparison depends on event sizes, frequency and retention.
AWS has not supplied a prior flat-rate figure or a worked cost comparison here, so its claim that most workloads will pay less is a useful signal, not a calculator. Teams can use the listed rates and their own usage patterns to check whether the change improves their particular bill.
Our read
This is a meaningful pricing change for developers building agents on AgentCore, not merely a new line on a cloud invoice. The rates and billing rules are clear enough to start estimating; the workload-specific maths still needs to be done. Before changing an architecture, model both event volume and how long stored memories remain around. Even short-term memory can be surprisingly fond of staying.
What to watch
- Whether AWS publishes example bills or a comparison with the previous per-event rate.
- How ingestion and retrieval costs vary for workloads with many small events.
- Whether teams use shorter TTLs to bring storage costs down.
Discussion spark: Would usage-based memory pricing make you more comfortable building agents that write frequent small events, or do the minimum billing size and storage charges make a flat per-event rate easier to manage?
Sources and evidence
- Announcing consumption-based pricing for short-term memory in Amazon Bedrock AgentCore (12 October 2026, 16:00 UTC)
not affiliated with or endorsed by Amazon Web Services (AWS)