Amazon Bedrock just launched OpenAI-compatible Projects API through its Mantle inference engine — making it trivial to migrate OpenAI workloads to AWS while gaining enterprise features like IAM-based access control and per-project cost tracking. Mantle is Bedrock's distributed inference engine for large-scale model serving, now supporting both the Responses API (with server-side conversation state) and Chat Completions API. The Projects API enables teams to create isolated environments with individual access controls and cost attribution — no additional charge, pay only for inference.

The Mantle engine represents AWS's answer to the multi-model enterprise challenge. Organizations running workloads across OpenAI, Anthropic, and open-weight models need consistent infrastructure, unified billing, and centralized governance. Mantle provides this through OpenAI SDK compatibility — switch your base URL and your existing code works on Bedrock with all the AWS security, compliance, and operational tooling your enterprise already uses.

But Mantle's project isolation, while excellent for cost control and access management, creates a knowledge fragmentation problem: insights generated in Project A are invisible to Project B, and the Responses API's conversation state doesn't persist across project boundaries.

Why OpenAI Compatibility Matters

The developer ecosystem built around the OpenAI API is massive. Thousands of tools, frameworks, and applications use the Chat Completions API format. Mantle's OpenAI compatibility means these tools work on Bedrock without code changes — developers swap an endpoint URL and gain access to AWS's full model catalog plus enterprise infrastructure. For enterprises locked into AWS for compliance or operational reasons, this removes the last barrier to accessing the broader AI ecosystem.

The Responses API adds server-side state management — conversations persist without the client passing full history on every request. This simplifies agent architectures where conversations span multiple interactions. Background processing support enables long-running tasks that complete asynchronously. Combined with IAM-based access control and cost tagging, Mantle makes enterprise AI deployment operationally manageable.

But operational management and knowledge management are different challenges. Mantle solves how to run AI workloads across models and teams. It doesn't solve how to share the knowledge those workloads generate.

Project Isolation vs. Knowledge Sharing

The Projects API creates clean boundaries: each project has its own access controls, cost tracking, and conversation state. This is the correct architecture for operational isolation — the marketing team's AI costs shouldn't blend with engineering's, and the finance team's data shouldn't leak to external contractors. But operational isolation shouldn't mean knowledge isolation.

Consider an enterprise running separate Bedrock projects for customer support, sales, and product development. The customer support project discovers that users are confused by a specific feature. The product development project would benefit from knowing this — but the discovery exists only within the support project's conversation state. The sales project identifies that competitors are positioning around a specific capability. Product development needs this intelligence, but it's trapped in the sales project's interactions.

The Responses API maintains conversation state within a project. MemU maintains knowledge state across projects. Together, they enable project isolation for operations while sharing intelligence for strategy.

How MemU Adds Cross-Project Memory to Bedrock

MemU integrates with Bedrock Mantle as a memory layer that spans projects while respecting access controls. Each project writes discoveries to MemU with appropriate permission tags. Other projects retrieve relevant memories only if they have the required permissions. The operational isolation remains intact while knowledge flows across organizational boundaries.

For enterprises running multiple AI initiatives on Bedrock, MemU provides the institutional memory that turns isolated projects into a connected intelligence network. Discoveries compound across teams. Patterns identified in one domain inform work in another. The enterprise AI deployment becomes greater than the sum of its projects.

AWS built the multi-model infrastructure. MemU provides the cross-project memory. Together, they deliver enterprise AI that learns as an organization.

Get Started

Add cross-project memory to your Amazon Bedrock deployment. Explore MemU at memu.pro and on GitHub.