This article is a machine translation of the contents of the following URL, which I wrote in Japanese:
Introduction
There are many situations where you want an AI agent to carry the context of a long conversation forward. If you pass the entire history as it is, the input grows large. If you use a summary, details get lost.
The summary strategy of Amazon Bedrock AgentCore Memory has so far summarized a conversation by splitting it into "chunks." With the announcement on November 17, 2026, a "full session summary" that summarizes the whole session into a single record is now created automatically in addition to the chunks.
- Announcement: https://aws.amazon.com/about-aws/whats-new/2026/11/amazon-bedrock-agentcore-memory-full-session-summaries
- Documentation: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/summary-strategy.html
In this article, I explain what the full session summary is and which situations it suits, and then check how it behaves by comparing it with the previous approaches.
What Is the Full Session Summary?
The summary strategy summarizes a conversation within a single session. It now produces two kinds of output.
| Output | Scope | Use |
|---|---|---|
| summary chunks | A part of the conversation. A new chunk is created each time a new set of events is processed | Retrieve only the parts relevant to a query |
| full session summary | The whole session summarized in one record. It is updated while the session continues | Hand over the context of the whole session as a single record, without assembling chunks yourself |
The official documentation describes the nature of chunks as follows.
AgentCore Memory doesn't rewrite earlier chunks or merge topics across chunks. If the conversation returns to a topic in a later set of events, the later chunk contains a new
<topic>entry for that topic with only the new details.
In other words, chunks are not a summary organized by topic. They are a chronological log. When the conversation returns to a topic, it goes into a separate chunk. The full session summary is what you use when you want an overall summary.
The full session summary recaps the entire session in a single, running record. As new events arrive, AgentCore Memory folds each new summary chunk into the existing recap. The record always reflects the conversation so far.
Source: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/summary-strategy.html
The key points of the specification are as follows (all taken from the official documentation and the announcement).
| Item | Details |
|---|---|
| Storage format | A long-term memory record whose memoryRecordType is FULL_SESSION. It is in the same namespace as the chunks |
| Enabling | Not required. It is created automatically for any memory that has a summary strategy. It is also created for existing sessions |
| Updates | Generated by a model and updated while the session continues. To get the latest one, read it again after adding events |
| Retrieval | The existing GetMemoryRecord, ListMemoryRecords, and RetrieveMemoryRecords. No additional operations |
| Filtering | Filter with the system-generated metadata key x-amz-agentcore-memory-recordType. No index declaration is required |
| Namespace (default) | /strategy/{memoryStrategyId}/actor/{actorId}/session/{sessionId}/ |
| Regions | All Regions where AgentCore Memory is available (per the announcement) |
Here is a CLI example that retrieves only the full session summary (the command is from the official documentation).
aws bedrock-agentcore list-memory-records \
--memory-id "<memory-id>" \
--namespace "/strategy/<memory-strategy-id>/actor/<actor-id>/session/<session-id>/" \
--metadata-filters '[
{
"left": {"metadataKey": "x-amz-agentcore-memory-recordType"},
"operator": "EQUALS_TO",
"right": {"metadataValue": {"stringValue": "FULL_SESSION"}}
}
]'
Memory Strategies Other Than the Summary Strategy
AgentCore Memory lets you configure how long-term memory is created from conversation events as "memory strategies." If you specify no strategy, no long-term memory records are created. You can combine multiple strategies in a single memory.
There are four built-in strategies.
| Strategy | What it extracts | Output format (per the official documentation) |
|---|---|---|
| semantic | Facts and contextual knowledge (e.g., an order number, a project deadline) | JSON |
| user preference | The user's preferences and choices (e.g., a preferred carrier, formal or casual tone) | JSON (context, preference, categories) |
| summary | A summary of the conversation within a single session (the subject of this article) | XML (per topic) |
| episodic | Meaningful scenes (episodes), summarized and organized. It can also generate insights across episodes (reflections) | XML (an episode is broken down into situation, intent, assessment, justification, and an episode-level reflection) |
- For semantic and user preference, the default namespace is at the user (actor) level. For episodic, you can choose the namespace level (strategy level, actor level, or session level). For summary, the session ID is required because summaries are created per session.
- Besides the built-in strategies, there are "built-in overrides," which change the prompts of the built-in strategies, and "self-managed" strategies, where you implement the extraction and consolidation processing yourself.
The full session summary in this article is generated only by the summary strategy.
Sources: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-strategies.html , https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/built-in-strategies.html
Use Cases That Suit It
First, a caveat: the only use the official documentation states explicitly is "handing over the context of a whole session as a single record." The following is my own view based on that property, and it includes items I have not verified. What I did confirm is described in the verification chapters.
| Situation | Why the full session summary suits it |
|---|---|
| Resuming a long conversation | You can restore the context by retrieving one record, without reprocessing the entire history |
| Saving context window space | You can replace old turns with the full session summary |
| Handover to a human agent | You can hand over only the current state, not the whole course of events |
| Checking the conclusions of meeting or planning sessions | You can check only what was decided |
There are also situations where it does not suit.
| Situation | Reason |
|---|---|
| Looking up a specific fact or number | In my verification, 1 of 14 facts was wrongly attached to a different topic during summarization and could not be retrieved (described later). If you need certainty, use the original events |
| An accurate record for auditing | A summary is generated by a model, so use the original events |
| You need to reflect even the most recent utterance exactly | In this measurement, the first chunk appeared about 65 seconds after the events were submitted, and the full session summary was updated 3 to 4 seconds after that. There was also a period during which chunks kept increasing |
Comparison with the Previous Approaches
I compare three approaches.
| Approach | Description |
|---|---|
| A. Full history | Put all events of the session into the context as they are |
| B. Chunk concatenation | Retrieve all chunks with ListMemoryRecords and concatenate them yourself before passing them in |
| C. Full session summary | Retrieve a single FULL_SESSION record and pass it in |
| Aspect | A. Full history | B. Chunk concatenation | C. Full session summary |
|---|---|---|---|
| Input tokens (56 events, average) | 1,226 | 1,296 | 965 |
| Implementation effort | Low | Retrieving, sorting, and concatenating are required | A single retrieval |
| Grouping of the same topic | As in the original conversation | Scattered chronologically (per the documentation). In my verification, it also split into separate topics | Grouped into one (Verification 2) |
| Correct answers to detail questions (56 events, 28 runs) | 28 | 26 | 26 |
| Correct answers to detail questions (18 events, 12 runs) | 12 | 12 | 12 |
| Correct answers to status questions (18 events, 2 questions x 3 runs) | 6 | 6 | 6 |
The input token counts include the system prompt and the question. The verification details are in the next chapter.
Verification
I ran this on November 18, 2026, in the Tokyo Region (ap-northeast-1). I called the APIs from boto3. The answering model was jp.anthropic.claude-haiku-4-5-20251001-v1:0 (temperature 0).
Verification Setup
- Scenario: a support inquiry about headphones with a defect, 18 events in this order: the defect and the replacement (topic 1), a change of delivery address (topic 2), back to topic 1, compensation points (topic 3), and back to topic 2. Verification 2 focuses on the change of delivery address (topic 2)
- Questions: 4 detail questions (order number, scheduled shipping date, points, confirmation code) and 2 status questions (what remains for the customer to do, what was agreed). Each was run 3 times
- Scoring: a script checks whether the answer contains the correct keywords that I embedded in the conversation
- Additional conversation: a 56-event conversation (a hotel reservation) containing 14 fine-grained facts. Each fact was asked about once, run 2 times
Note on the quoted text: the conversations were held in Japanese, and the memory records (chunks and full session summaries) were generated in Japanese. In this article, all conversations and records are shown as English translations of the Japanese originals. The structure, line breaks, and formatting follow the originals, but the wording is a translation. For the original Japanese text, see the Japanese article linked at the top.
The 18-event conversation used in Verifications 1 to 3 and in the first half of Verification 5 is shown below (an English translation of the Japanese original). I submitted the first 8 events and the last 10 events separately.
First part (8 events):
Customer: My wireless headphones NC-700, which I bought on the 28th of last month, produce no sound from the left side. The order number is XYZ-48213.
Agent: I'm sorry for the inconvenience. Order number XYZ-48213, NC-700, correct? Has the left side had no sound since you first paired it?
Customer: Yes, the left side has been silent from the start. I re-paired it several times, but nothing changed.
Agent: It is likely an initial defect, so we will handle it with a replacement. We will cover the shipping cost of the replacement.
Customer: Thank you. By the way, I would like to change my delivery address. Is that possible? It is currently Shinagawa Ward, Tokyo.
Agent: We can change it if it has not shipped yet. Please tell us the new delivery address.
Customer: I would like to change it to Kita Ward, Osaka City, Osaka Prefecture.
Agent: Understood. We will accept the change to Kita Ward, Osaka City, Osaka Prefecture. For identity verification, please make a note of the confirmation code 7391.
Last part (10 events):
Customer: The confirmation code is 7391, I've noted it. By the way, when will the replacement be shipped?
Agent: The scheduled shipping date of the replacement is November 20. The old unit must be returned, and we will email you the return label on November 19.
Customer: The return label will arrive by email on November 19. Understood.
Agent: To confirm the defect, could you email us a video showing the left side with no sound by November 19?
Customer: I will send the video by November 19. Also, is there any compensation for the inconvenience?
Agent: As an apology, we will grant you 800 points. The points are granted after the replacement ships.
Customer: 800 points, thank you. Lastly, is the delivery address I mentioned earlier, Kita Ward, Osaka City, Osaka Prefecture, correct?
Agent: Yes, we will send the replacement to Kita Ward, Osaka City, Osaka Prefecture. The change was made before shipping, so there is no problem.
Customer: I see. So all I need to do now is send the video, right?
Agent: Yes, the only thing needed from you is to send the video. Please contact us if you have any questions.
Verification 1: Generation of Chunks and the Full Session Summary
I created a memory with one summary strategy (it became ACTIVE in about 20 seconds) and submitted the first 8 events.
| Time (JST) | What happened |
|---|---|
| 07:08:12 | Submitted 8 events |
| 07:09:17 | One chunk was created (about 65 seconds later) |
| 07:09:20 | The full session summary was created (3 seconds after the chunk) |
The recordType in the chunk's metadata was BASE. This value is not described in the documentation; I learned it from this observation. The chunk was in XML <topic> format, and the full session summary was in Markdown (headings, bold text, and tables).
The first chunk and the full session summary at that point are shown below (English translations of the Japanese originals; the structure and line breaks follow the retrieved text).
First chunk:
<topic name="Product defect report">
The user reported a defect in the wireless headphones NC-700 (order number: XYZ-48213) purchased on the 28th of last month: only the left side has had no sound since the beginning. Re-pairing did not help. The assistant judged that an initial defect was likely, proposed a replacement, and said that the shipping cost of the replacement would be borne by the company.
</topic>
<topic name="Change of delivery address">
The user requested to change the delivery address from Shinagawa Ward, Tokyo, to Kita Ward, Osaka City, Osaka Prefecture. The assistant explained that the change is possible if the order has not shipped yet and accepted the change. A confirmation code "7391" was issued for identity verification.
</topic>
Full session summary:
## Support case summary
### Product defect report
- The user reported a defect in the **wireless headphones NC-700** (order number: **XYZ-48213**) purchased on the **28th of last month**.
- Details of the defect: **only the left side has had no sound since the beginning**.
- The user tried re-pairing, but it did not help.
- The assistant judged that an **initial defect was likely** and proposed a **replacement**. The shipping cost of the replacement is **borne by the company**.
### Change of delivery address
- The user requested to change the delivery address from **Shinagawa Ward, Tokyo** to **Kita Ward, Osaka City, Osaka Prefecture**.
- The assistant explained that the **change is possible if the order has not shipped yet** and accepted the change.
- The confirmation code **"7391"** has been issued for identity verification.
Verification 2: How a Returning Topic Is Grouped
When I submitted the last 10 events, a second chunk was created. The second chunk added the same delivery-address topic as in the first part as a new topic named "Confirmation of delivery address," separate from "Change of delivery address" in the first chunk. As the official documentation states, topics are not merged.
The full session summary, on the other hand, grouped them into a single "Change and confirmation of delivery address."
The second chunk and the updated full session summary are shown below. Pay attention to how the delivery-address topic is handled in the chunks and in the full session summary.
Second chunk:
<topic name="Shipping and return procedure for the replacement">
The scheduled shipping date of the replacement is **November 20**. The old unit must be returned, and the return label will be sent by email on **November 19**. The user acknowledged receiving the return label.
</topic>
<topic name="Sending a video to confirm the defect">
The assistant asked the user to email a video showing the left side with no sound by **November 19**. The user confirmed that they would send the video by November 19. The only action required from the user is sending the video.
</topic>
<topic name="Apology compensation points">
**800 points** will be granted as an apology for the inconvenience. The points will be granted after the replacement ships.
</topic>
<topic name="Confirmation of delivery address">
The user confirmed whether the delivery address (Kita Ward, Osaka City, Osaka Prefecture) was correct, and the assistant confirmed again that the change is fine because the order has not shipped yet.
</topic>
Updated full session summary:
## Support case summary
### Product defect report
- The user reported a defect in the **wireless headphones NC-700** (order number: **XYZ-48213**) purchased on the **28th of last month**.
- Details of the defect: **only the left side has had no sound since the beginning**.
- The user tried re-pairing, but it did not help.
- The assistant judged that an **initial defect was likely** and proposed a **replacement**. The shipping cost of the replacement is **borne by the company**.
### Change and confirmation of delivery address
- The user requested to change the delivery address from **Shinagawa Ward, Tokyo** to **Kita Ward, Osaka City, Osaka Prefecture**.
- The assistant explained that the **change is possible if the order has not shipped yet** and accepted the change. It has been reconfirmed that the new address (Kita Ward, Osaka City, Osaka Prefecture) is correct.
- The confirmation code **"7391"** has been issued for identity verification.
### Shipping and return procedure for the replacement
- The **scheduled shipping date of the replacement is November 20**.
- The old unit must be returned, and the **return label is scheduled to be sent by email on November 19**. The user has agreed to receive it.
### Sending a video to confirm the defect
- The assistant asked the user to email a video showing the left side with no sound by **November 19**.
- The user has confirmed that they will send it by the deadline. The only action required from the user is **sending the video**.
### Apology compensation points
- **800 points** will be granted as an apology for the inconvenience.
- The points will be granted **after the replacement ships**.
Verification 3: Update After Adding Events
| Time (JST) | What happened |
|---|---|
| 07:10:30 | Submitted the last 10 events |
| 07:11:36 | The second chunk was created (about 66 seconds later) |
| 07:11:40 | The full session summary was updated (4 seconds later, with the same record ID) |
The full session summary was updated in place as the existing record, not created as a new one. In the 56-event conversation, chunks kept increasing for a while after all events had been submitted. When I ran the comparison about 47 seconds after submission (4 chunks), the results were poor, and they improved when I ran it again after the fifth chunk had appeared (described later). Note that the results can change depending on when you retrieve the records.
Verification 4: Retrieval with the FULL_SESSION Filter
| Operation | Result |
|---|---|
list-memory-records (FULL_SESSION filter) |
Only the full session summary (1 record) |
RetrieveMemoryRecords (with the filter) |
Only the full session summary (1 record) |
RetrieveMemoryRecords (without the filter, searching for "confirmation code") |
3 records (the full session summary first, then 2 chunks) |
As the documentation says, I could filter by the system-generated key.
Verification 5: Comparison of the Three Approaches
Results for the short conversation (18 events).
| Approach | Characters | Average input tokens | Detail | Status |
|---|---|---|---|---|
| A. Full history | 867 | 874 | 12/12 | 6/6 |
| B. Chunk concatenation (2 chunks) | 734 | 716 | 12/12 | 6/6 |
| C. Full session summary | 659 | 696 | 12/12 | 6/6 |
Results for the longer conversation (56 events, 14 fine-grained facts).
| Approach | Characters | Average input tokens | Correct answers | Question answered incorrectly |
|---|---|---|---|---|
| A. Full history | 1,197 | 1,226 | 28/28 | None |
| B. Chunk concatenation (5 chunks) | 1,437 | 1,296 | 26/28 | Where to receive luggage in advance |
| C. Full session summary | 939 | 965 | 26/28 | Same as above |
- At this scale, chunk concatenation was larger than the full history in both characters and tokens. I think this is because the chunks contain overlapping content, but I did not isolate the cause
- The one missed question was caused by the fourth chunk recording the advance luggage reception (at the front desk) as a matter about invoices. The full session summary folded in this chunk, so it inherited the same error
- A chunk boundary sometimes fell between a question and its answer. That chunk recorded "no clear answer from the assistant has been confirmed," and a later chunk filled it in as "answers to previously unconfirmed questions." The full session summary ended up in the filled-in state
- In the first comparison (when there were 4 chunks), both B and C scored 20/28. After the fifth chunk appeared, I ran it again and they scored 26/28
An Example of an Error Being Inherited
This is how the error occurred for the question asking where to receive luggage in advance (in the 56-event conversation).
The original conversation (only the relevant part):
Customer: I would like to send my luggage in advance.
Agent: You can receive it at the front desk in advance. Please write the guest's name and the stay dates on the address label.
The full original conversation (56 events)
Customer: I'm looking for a hotel for my business trip to Osaka next month. My budget is within 15,000 yen per night.
Agent: Understood. Your budget is within 15,000 yen per night.
Customer: Thank you.
Agent: You're welcome.
Customer: I plan to stay for 2 nights from December 3.
Agent: Check-in on December 3 for 2 nights. Understood.
Customer: Let me think about it.
Agent: Please take your time.
Customer: I would like a place near Shin-Osaka Station.
Agent: I will search around Shin-Osaka Station. Hotel Sunrise Shin-Osaka is a candidate.
Customer: Thank you.
Agent: You're welcome.
Customer: Does it include breakfast?
Agent: There is a plan with breakfast, and breakfast is served from 6:45.
Customer: Let me think about it.
Agent: Please take your time.
Customer: I would like a non-smoking room.
Agent: I will arrange a non-smoking double room. We cannot specify the room number.
Customer: Thank you.
Agent: You're welcome.
Customer: I will pay with my company card, and please address the invoice to Yamamoto in the accounting department.
Agent: We will address the invoice to Yamamoto in the accounting department.
Customer: Let me think about it.
Agent: Please take your time.
Customer: Could you tell me the reservation confirmation number?
Agent: The reservation confirmation number is HK-90417.
Customer: Thank you.
Agent: You're welcome.
Customer: Is there a shuttle from the station?
Agent: There is no shuttle, but it is a 6-minute walk from Shin-Osaka Station.
Customer: Let me think about it.
Agent: Please take your time.
Customer: What time is check-out?
Agent: Check-out is at 11:00 a.m.
Customer: Thank you.
Agent: You're welcome.
Customer: I would like to send my luggage in advance.
Agent: You can receive it at the front desk in advance. Please write the guest's name and the stay dates on the address label.
Customer: Let me think about it.
Agent: Please take your time.
Customer: I don't need parking.
Agent: Understood. For your reference, parking is 1,500 yen per night.
Customer: Thank you.
Agent: You're welcome.
Customer: Can I cancel if my plans change?
Agent: You can cancel free of charge up to 3 days before the stay date.
Customer: Let me think about it.
Agent: Please take your time.
Customer: I have a companion. Can two people stay?
Agent: The double room can be used by up to 2 guests. There is 1 companion.
Customer: Thank you.
Agent: You're welcome.
Customer: Lastly, can I receive the receipt by email after the stay?
Agent: Yes, we will send it to your registered email address after check-out.
Customer: Let me think about it.
Agent: Please take your time.
The fourth chunk (relevant lines):
- About the invoice: it can be received in advance at the front desk. It was explained that the guest's name and the stay dates must be written on the address label.
- Sending luggage in advance: the user said they want to send their luggage first, but no clear answer from the assistant has been confirmed.
The full session summary (relevant line):
- The invoice can be received in advance at the front desk. **The guest's name and the stay dates must be written** on the address label.
The answer about the luggage was recorded as being about the invoice. The error that arose in the chunk remains as it is in the full session summary.
The full text of what I retrieved is below (English translations of the Japanese originals; the structure and line breaks follow the retrieved text).
The full session summary of the 56-event conversation (full text, 939 characters):
## Hotel search and reservation overview
### Search conditions
The user is looking for a hotel for a business trip to Osaka with the following conditions.
| Item | Details |
|------|------|
| Check-in | December 3 |
| Nights | 2 nights |
| Budget | Within 15,000 yen per night |
| Preferred area | Near Shin-Osaka Station |
| Meals | Breakfast plan requested |
### Assistant's responses
- Confirmed the stay conditions (check-in on December 3, 2 nights, budget within 15,000 yen).
- Informed that a plan with breakfast is available.
- Told that breakfast is served from **6:45**.
### Reservation details
- **Hotel name**: Hotel Sunrise Shin-Osaka (around Shin-Osaka Station)
- **Room type**: Non-smoking double room (the room number cannot be specified)
- **Reservation confirmation number**: HK-90417
### Payment and invoice
- Payment is specified to be made with a **company card**.
- The user requested that the invoice be addressed to **Yamamoto in the accounting department**. The assistant confirmed this.
- The invoice can be received in advance at the front desk. **The guest's name and the stay dates must be written** on the address label.
### Additional inquiries and answers
| Question | Answer |
|----------|------|
| Shuttle from the station | No shuttle. **6-minute walk** from Shin-Osaka Station |
| Check-out time | **11:00 a.m.** |
| Parking | **1,500 yen** per night (the user said it is not needed) |
| Cancellation | Free cancellation up to **3 days before** the stay date |
| Two guests staying | The double room is available for **up to 2 guests**. 1 companion |
| Luggage and receipt sent by email | Sent to the **registered email address** after check-out |
| Whether the receipt can be received by email | The user asked at the end, but no clear answer from the assistant has been confirmed |
### Flow of the conversation (supplement)
- The user said "Let me think about it" once, and the assistant replied "Please take your time."
- After that, the user asked again about the receipt.
Chunk 1 (full text):
<topic name="Hotel search conditions">
The user is looking for a hotel for a business trip to Osaka. The main conditions are as follows:
- Check-in: December 3
- Nights: 2 nights
- Budget: within 15,000 yen per night
- Preferred area: near Shin-Osaka Station
- Breakfast plan requested
</topic>
<topic name="Assistant's responses">
The assistant confirmed the stay conditions (check-in on December 3, 2 nights, budget within 15,000 yen). It also informed that a plan with breakfast exists and that breakfast is served from 6:45.
</topic>
Chunk 2 (full text):
<topic name="Hotel reservation confirmed">
The assistant proposed "Hotel Sunrise Shin-Osaka" around Shin-Osaka Station as a candidate and arranged a non-smoking double room. The room number cannot be specified. The reservation confirmation number is "HK-90417".
</topic>
<topic name="Payment and invoice instructions">
The user specified payment with a company card and requested that the invoice be addressed to Yamamoto in the accounting department. The assistant confirmed this.
</topic>
Chunk 3 (full text):
<topic name="Additional inquiries and answers">
The user asked the following additional questions, and the assistant answered:
- Shuttle from the station: there is no shuttle, but it was explained that it is a 6-minute walk from Shin-Osaka Station.
- Check-out time: answered that it is 11:00 a.m.
- The user also asked about the reservation confirmation number, but no clear answer from the assistant has been confirmed.
</topic>
Chunk 4 (full text):
<topic name="Additional inquiries (continued)">
The following additional exchanges were confirmed:
- About the invoice: it can be received in advance at the front desk. It was explained that the guest's name and the stay dates must be written on the address label.
- Sending luggage in advance: the user said they want to send their luggage first, but no clear answer from the assistant has been confirmed.
- Whether there is a companion: the user said there is a companion and asked whether two people can stay, but no clear answer from the assistant has been confirmed.
- Cancellation: the user asked about cancellation if plans change, but no clear answer from the assistant has been confirmed.
- Parking: the assistant explained that it is 1,500 yen per night, but the user answered that parking is not needed.
</topic>
Chunk 5 (full text):
<topic name="Answers to previously unconfirmed questions">
The assistant answered the following previously unanswered questions:
- Cancellation: free cancellation is possible up to **3 days before** the stay date.
- Two guests staying: the double room is available for up to 2 guests. 1 companion.
- Luggage and receipt by email: it was explained that they will be sent to the registered email address after check-out.
</topic>
<topic name="Additional inquiries (receipt)">
The user asked at the end, "Can I receive the receipt by email after the stay?" No clear answer from the assistant has been confirmed.
</topic>
<topic name="Flow of the conversation">
The user said "Let me think about it" once, and the assistant replied "Please take your time." After that, the user asked again about the receipt.
</topic>
Permissions and Encryption
- The full session summary is controlled by the same IAM actions (
GetMemoryRecord,ListMemoryRecords,RetrieveMemoryRecords) and the same memory resource ARN as other memory records. You cannot allow or deny access to the full session summary alone - To restrict by record type, filter on
x-amz-agentcore-memory-recordTypein your application - If you have configured a customer managed KMS key (
encryptionKeyArn), that key also encrypts the full session summary
Conclusion
The summary strategy of AgentCore Memory now automatically creates a full session summary that puts the whole session into a single record. When I tried it in the Tokyo Region, it was generated without any additional configuration, and the same record was updated 3 to 4 seconds after a new chunk was created. If you filter by FULL_SESSION, you can retrieve it in a single call.
Within the scope of this verification, this is how I think about choosing between the approaches.
| Purpose | Suitable approach |
|---|---|
| Handing over an overview or the current state of a conversation | Full session summary |
| Checking facts where errors are unacceptable, such as numbers, dates, and amounts | The original events |
Because the full session summary is built by folding in chunks, a fact that was attached to a different topic at the chunk stage was inherited as it was. For facts that need to be checked, I think it is preferable to design the system so that you also use the original events.
References
- Amazon Bedrock AgentCore Memory now provides full session summaries (2026-11-17): https://aws.amazon.com/about-aws/whats-new/2026/11/amazon-bedrock-agentcore-memory-full-session-summaries
- Summary strategy - Amazon Bedrock AgentCore: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/summary-strategy.html
- Memory strategies - Amazon Bedrock AgentCore: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/memory-strategies.html
- Built-in strategies - Amazon Bedrock AgentCore: https://docs.aws.amazon.com/bedrock-agentcore/latest/devguide/built-in-strategies.html

Top comments (0)