Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Short-term memory costs $1.00/GB to ingest, $0.20/GB to retrieve and $0.10/GB-month to store. Storage uses actual event size and retention time, prorated hourly.
- Ingestion and each retrieval use a billable size between 12KB and 64KB per event. The 64KB figure is not the actual CreateEvent payload limit.
- In AWS's 3KB illustration, 100,000 writes, one retrieval per event and 30 days of storage produce a $1.47 short-term memory subtotal. The $18.97 example total also includes long-term memory.
- Changing eventExpiryDuration affects only events written afterward. Existing events keep their original expiry, and expired content cannot be recovered.
Short-term memory has changed; long-term rates have not
Amazon Bedrock AgentCore short-term memory (STM) stores raw events from a session so an agent can keep using its current context. It now charges $1.00/GB for ingestion, $0.20/GB for retrieval and $0.10/GB-month for storage. The announcement covers AWS Regions where AgentCore Memory is already offered; it does not establish availability in every AWS Region.
AWS dated the announcement October 12, 2026, while its pricing table labels the new rates as effective October 6. The old ingestion rate of $0.25 per 1,000 new events is marked as applying through October 5. Keep the announcement date separate from the pricing table's effective date, and do not treat this article's date as the launch date.
Long-term memory (LTM) turns raw events into knowledge that can persist across sessions. It retains its own record and retrieval charges: AWS explicitly says long-term memory pricing is unchanged. An application using both types of memory needs to budget for them separately.
A 3KB event is billed at 12KB; 64KB is not a request limit
Short-term memory ingestion calculates billable size for each event, with a 12KB minimum and a 64KB maximum. Retrieval applies the same range to each event retrieved on each read. A write or one retrieval of a 3KB event is therefore billed at 12KB, rather than simply multiplying its actual 3KB size by the read or write rate.
Those figures describe the size used for billing. They do not truncate the actual event to 64KB or mean that a request can submit only 64KB. Storage uses the event's actual size, so the 64KB billing cap must not be applied to storage charges.
Actual submission limits are a separate matter. The current AgentCore Memory quota table lists a 10MB maximum event size for CreateEvent, a 100KB maximum per message and up to 100 messages per operation. Meeting one limit does not satisfy the others, and these CreateEvent limits should not be generalized to every Memory API.
Why AWS's 3KB example has a $1.47 STM subtotal
The following is an illustration from AWS's pricing page, not a bill we measured: 100,000 short-term events are ingested in a month, each actually 3KB, each retrieved once and, under the example's assumptions, all stored for a full 30-day month. Reads and writes are billed at 12KB; storage still uses the actual 3KB size.
The page calculates 100,000 events × 12KB as 1.2GB, yielding $1.20 for ingestion and $0.24 to retrieve each event once. Actual stored data is 100,000 × 3KB, or 0.3GB, costing $0.03 when held for one month. These three short-term memory charges add up to $1.47.
The $18.97 total shown in the same example also includes $7.50 for 10,000 long-term memory records and $10.00 for 20,000 long-term retrievals. It is not a short-term memory rate. Nor should a real workload whose events arrive throughout the month be treated as identical to an illustration that holds all data for a full month.
Choose a TTL around the context you need to keep
TTL is the retention period before a raw event expires. Storage uses the event's actual size and retention time, prorated hourly. A shorter retention period for new events can reduce their storage time, but it also removes that raw context sooner.
The current Memory quota table lists an EventExpirationDuration range of 7–365 days for CreateEvent operations, and the creation guide says events can be retained for up to 365 days. Check the applicable configuration in the documentation for the operation you use.
When creating a resource in the console, the documented path is Memory → Create memory, then set the number of days under Short-term memory (raw event) expiration. Choose a duration based on whether your application needs to revisit a session, troubleshoot it or retrieve its raw events again.
eventExpiryDuration is applied when an event is written. Updating it affects only events created after the change. Stored events keep their original expiry, cannot have it extended and cannot be recovered after expiration. Lowering the TTL today does not imply that old events expire immediately or that their storage charges immediately fall.
Count repeat reads before comparing retention periods
Repeated retrievals produce additional retrieval charges: fetching a 3KB event once uses a 12KB billable size, and fetching it again is another billable read. List your event size distribution, writes, actual retrievals per event and retention period. That better reflects the rules than multiplying an average payload size by the new rate.
Compare old and new costs for the same events and reading requirements. Comparing only the old ingestion rate with the new ingestion price leaves out retrieval and storage. AWS says it expects lower costs for most workloads, but your sizes, read frequency and retention needs may differ. Savings are not guaranteed, and the illustration does not verify your account's bill.
Check credits and model charges separately
The AgentCore pricing page says new AWS customers can receive up to $200 in Free Tier credits. That does not make short-term memory free or establish eligibility for existing accounts or every new account. Check the AWS rules that apply to your account for eligibility, eligible charges and remaining credits. Regional availability also does not establish account permissions.
If you also enable long-term memory, include its separate charges. The pricing page additionally notes that extraction strategies with overrides may incur model usage charges in your account. This illustration separates only the listed Memory charges; it does not determine taxes, other AWS service charges or the amount payable after account discounts.
Sources
Frequently Asked Questions
Must I enable long-term extraction if I only need session context?
No. AWS's creation guide says that when no memory strategy is defined, only short-term memory containing raw events is stored. Choose whether to extract knowledge across sessions based on your application. An illustration using both types of memory does not make a long-term strategy mandatory.
How are tiny monthly charges rounded, and is there a minimum GB charge?
The verified rules specify a 12KB minimum per event for ingestion and retrieval, and hourly proration for storage. The official sections and illustration used here do not specify a separate minimum GB charge or final invoice rounding policy. The example's decimal arithmetic should not be treated as a complete description of account billing rules.