InfrastructureAmazon Web Services2026-05-30AWS Bedrock Gets Runtime-Defined Agents and 90-Day Memory
AWS has launched Bedrock Inline Agents, letting developers configure agent behavior at invocation time rather than through pre-built profiles, plus a persistent memory layer that retains session context for up to 90 days. Both features are live now in us-east-1 and eu-west-1.
Original sourceAmazon Web Services has shipped two meaningful additions to its Bedrock agent platform: Inline Agents and a persistent memory layer. Inline Agents let developers pass agent instructions, tools, and configurations directly in the API call at runtime, bypassing the need to pre-register an agent profile in the console. This shifts the complexity from setup-time configuration to call-time construction, which matters for applications that need to dynamically adjust agent behavior per user, per request, or per workflow branch without maintaining a sprawling library of pre-baked agent definitions.
The persistent memory feature adds a managed context store that survives session boundaries, retaining information for up to 90 days. Rather than requiring developers to build and maintain their own memory retrieval layer on top of DynamoDB or S3, Bedrock now handles chunking, indexing, and surfacing relevant prior context automatically during agent invocations. The combination of the two features targets a specific and common pain point: stateful, multi-session agentic workflows where agent behavior needs to flex per context.
Both features are available immediately in us-east-1 and eu-west-1, with no announced timeline for additional region support. Pricing details follow Bedrock's existing token-based model, with memory storage billed separately. The release is a meaningful infrastructure step for developers building production agentic systems on AWS rather than stitching together retrieval and session logic from scratch.
Panel Takes
The Builder
Developer Perspective
“The primitive here is clean: swap pre-registered agent profiles for runtime-constructed ones, and offload session memory to a managed store instead of rolling your own retrieval pipeline. The DX bet is that putting configuration in the API call — not in the console — is where complexity belongs for dynamic workloads, and that's the right call. If the SDK surface is as described and I'm not wrestling with 4 IAM roles before hello-world, this is a genuine skip of maybe 200 lines of boilerplate I was writing myself.”
The Skeptic
Reality Check
“The 90-day memory cap is the number I'd stress immediately — real CRM-style agents need user context that doesn't expire on an arbitrary billing cycle, and 90 days is a product decision dressed up as a technical constraint. The scenario where this breaks is any multi-tenant SaaS where one noisy tenant's context bleeds retrieval quality for others, and AWS hasn't said anything about isolation guarantees at scale. What kills this in 12 months isn't a competitor — it's AWS itself shipping a better version natively in Bedrock Agents v2 that makes this an awkward transition footnote.”
The Futurist
Big Picture
“The thesis baked into Inline Agents is falsifiable: agent topology will be determined at runtime by application logic, not at deploy time by platform configuration, because real workflows are too branchy to pre-register. That bet is correct and early — most teams are still thinking in fixed agent graphs. The second-order effect that matters here is that managed memory shifts the locus of user-state ownership from application databases to the AI infrastructure layer, which has significant implications for portability, vendor lock-in, and who controls the longitudinal user model in an agentic app.”
The Founder
Business & Market
“The buyer is any AWS-committed enterprise team that's already running Bedrock and was previously paying engineers to maintain bespoke session memory and agent configuration systems — that's a real cost center and AWS is collapsing it into the platform bill. The moat isn't the feature itself, it's the switching cost: once your agent's memory layer lives in Bedrock, migrating to a competitor means migrating state, and nobody does that willingly. The risk is straightforward — if memory storage pricing isn't transparent at scale, the first large enterprise to get a surprise invoice will make it a conference talk.”