A customer contacts your support bot in March. The last time they talked to it was in October, about a cargo bike reservation that never happened. Does your agent know that? In most setups I have seen, the answer is: only if someone wrote a fair amount of glue code to reconstruct it.
On November 17, 2026, AWS shipped two changes to Amazon Bedrock AgentCore Memory that make this much simpler. I tried them out with a small support assistant for a made-up bike rental and wrote everything up in AgentCore Memory session summaries for AI agents. Here is what I took away from it.
Why summary chunks were never enough
AgentCore Memory has two layers. Short-term memory stores events: every customer message and every agent reply, grouped by actor and session. Long-term memory is built asynchronously from those events by strategies, which produce memory records under a namespace such as /summaries/{actorId}/{sessionId}/.
Until now, the summary strategy gave you several summary chunks per session, one per processing run. They are never merged, so topics repeat. If you wanted one readable recap, you had to sort, deduplicate and possibly send the chunks through a model once more. That is exactly the kind of code nobody wants to understand six months later.
What changed on November 17
One record per session. Every memory with a summary strategy now also maintains a single record of type FULL_SESSION that is updated whenever new events have been processed. There is nothing to switch on, and according to the announcement it covers existing sessions too. You find it via the system-provided metadata key x-amz-agentcore-memory-recordType, with the same IAM actions as any other record.
Up to ten years of raw events. The eventExpiryDuration for short-term memory now goes up to 3650 days instead of one year. The catch: retention is fixed when an event is written. If you raise it later, older events keep their original expiry.
In my example, the assistant receives the summaries of the last three sessions as dated one-liners in its system prompt, never the full old transcripts. That keeps prompts small and the cost per reply predictable. The project uses the Serverless Framework v4, TypeScript and a Bedrock model. Trying it out looks roughly like this:
serverless deploy --stage dev curl -s -X POST "$API/chat" \ -H 'content-type: application/json' -H 'x-demo-customer: customer_42' \ -d '{"sessionId":"s-2026-11-18","message":"Is the cargo bike free on Saturday?"}' # a few minutes later curl -s "$API/sessions/s-2026-11-18/recap" -H 'x-demo-customer: customer_42' Don't be surprised by a 404 on the recap at first: summaries are created asynchronously and need a moment.
Where the money actually goes
Ten years of retention sounds expensive, so I calculated it for a small shop with 2,000 conversations per month (prices from the AgentCore pricing page, US regions, as of November 18, 2026; the volumes are assumptions). The raw events turn out to be almost negligible: under half a dollar per month for ingestion, and about the same for storage even after ten years.
The real cost driver is long-term records without an expiry date. With four records per session, you would pay about 72 US dollars per month after one year and around 720 after ten years. The fix is already part of the strategy configuration: give your long-term records an expiry. For a support agent I would start with one to two years.
Privacy is the harder question
Long retention also means long deletion paths. A customer's data lives in both the events and the long-term records, and I did not find a single "delete everything for this actor" call. So plan a deletion script early. extractionMode: "SKIP" keeps sensitive messages out of summaries, and a customer managed KMS key has to be set when the memory is created.
My take: the full session summary is a small feature that removes a whole block of custom code. The ten years, on the other hand, are a business and data protection decision that a development team should not make alone.
What the full article covers
- The complete Serverless Framework project, including the memory resource and tightly scoped IAM permissions
- TypeScript code for fetching
FULL_SESSIONfor one session and across all sessions of a customer - Why ID validation matters for namespace isolation, and how the prompt builder deals with the Converse API
- Tests that run without an AWS account, common mistakes and the full cost calculation
Read the full walkthrough here: AgentCore Memory session summaries for AI agents
📖 This is the condensed version. The full article on sascha-ms.de goes deeper – with all code samples, diagrams and a short knowledge check.
👉 Next step: Working with AWS CDK? Grab my printable AWS CDK cheat sheet with commands, concepts and a stack skeleton.
📬 Stay updated: New articles on serverless, AI and leading tech teams via RSS, browser push or the weekly newsletter (in German).
Originally published on sascha-ms.de. More articles on serverless, AI in everyday development and leading tech teams: sascha-ms.de/en
