AWS recently announced Amazon S3 Annotations, a feature that lets teams attach rich, searchable context such as summaries, classifications, compliance data, or AI-generated insights directly to S3 objects. Annotations can be updated independently of the object and queried across datasets, reducing the need for separate metadata systems.
Written in JSON, XML, or YAML, annotations provide AI agents and analytics tools with the context they need to find and use objects in S3. While the object storage already supported tags and both system and user-defined metadata at the object level, Daniel Abib, senior specialist solutions architect at AWS, explains:
While these capabilities work well for their intended purposes, they have limitations when you need to attach much richer context without building and maintaining separate metadata systems. Annotations address these needs by providing metadata capabilities at a fundamentally different scale and flexibility, offering mutable, queryable context per object compared to 10 immutable tags or 2 KB of headers.
S3 Annotations significantly expand S3's metadata model by allowing up to 1000 mutable annotations per object, with a combined capacity of 1 GB, compared to 2 KB for user-defined metadata and just 10 tags per object. The extended flexibility makes S3 Annotations suitable for storing rich, structured business context rather than simple attributes or lifecycle metadata. In the article "Context intelligence for your data and AI agents at scale," Mai-Lan Tomsen Bukovec, technology VP at AWS, adds:
Annotations become queryable through S3 Metadata. When you enable annotation tables on a bucket, every annotation flows automatically into a fully managed Iceberg table. You can query across all your objects with Amazon Athena, Amazon Redshift or any Iceberg-compatible engine, and agents can discover annotations in natural language through the S3 Tables MCP server.
Editing object metadata on S3 has been a long-awaited request from the community. The reaction on Reddit has been overwhelmingly positive, with user ReturnOfNogginboink writing:
The really important part here is that annotations can be modified. Unlike object metadata, which requires you to read the full object out of s3, and rewrite it to S3 with new metadata. This is the big deal here. This is going to unlock all kinds of new workflow possibilities.
Expected use cases for annotations include attaching operational, analytical, and compliance context directly to data assets in media, financial services, and life sciences, making metadata easier to manage, discover, and query without separate systems. Corey Quinn, chief cloud economist at The Duckbill Group, writes in his newsletter:
S3 gets an object store in the object store, and now you can bolt a full gigabyte of "context" onto each object, because the four existing metadata mechanisms weren't confusing enough. Every announcement these days whispers from behind you "agentic workflows," which is Seattle for "your AI will generate the data, then pay Athena to read it back." Vertical integration is great, but now it's on your own bill.
S3 Annotations are stored and billed at S3 Standard rates regardless of the underlying object's storage tier. Annotations are also replicated when objects are copied, with each annotation copy counted and billed as a separate PUT request. S3 Annotations are generally available in all regions.