diff --git a/docs/modules/servers/nav.adoc b/docs/modules/servers/nav.adoc
index b0030e78704..1e619f3879e 100644
--- a/docs/modules/servers/nav.adoc
+++ b/docs/modules/servers/nav.adoc
@@ -14,6 +14,7 @@
 *** xref:distributed/architecture/index.adoc[]
 **** xref:distributed/architecture/implemented-standards.adoc[]
 **** xref:distributed/architecture/consistency-model.adoc[]
+**** xref:distributed/architecture/blobstore.adoc[]
 **** xref:distributed/architecture/specialized-instances.adoc[]
 **** xref:distributed/architecture/data-tiering.adoc[]
 *** xref:distributed/run/index.adoc[Run]
@@ -88,6 +89,7 @@
 *** xref:postgres/architecture/index.adoc[]
 **** xref:postgres/architecture/implemented-standards.adoc[]
 **** xref:postgres/architecture/consistency-model.adoc[]
+**** xref:postgres/architecture/blobstore.adoc[]
 **** xref:postgres/architecture/specialized-instances.adoc[]
 *** xref:postgres/run/index.adoc[]
 **** xref:postgres/run/run-java.adoc[Run with Java]
diff --git a/docs/modules/servers/pages/distributed/architecture/blobstore.adoc b/docs/modules/servers/pages/distributed/architecture/blobstore.adoc
new file mode 100644
index 00000000000..ea6c3469dcf
--- /dev/null
+++ b/docs/modules/servers/pages/distributed/architecture/blobstore.adoc
@@ -0,0 +1,4 @@
+= Distributed James Server — BlobStore
+:navtitle: BlobStore
+
+include::partial$architecture/blobstore.adoc[]
diff --git a/docs/modules/servers/pages/postgres/architecture/blobstore.adoc b/docs/modules/servers/pages/postgres/architecture/blobstore.adoc
new file mode 100644
index 00000000000..db1e165838a
--- /dev/null
+++ b/docs/modules/servers/pages/postgres/architecture/blobstore.adoc
@@ -0,0 +1,4 @@
+= Postgresql James server — BlobStore
+:navtitle: BlobStore
+
+include::partial$architecture/blobstore.adoc[]
diff --git a/docs/modules/servers/partials/architecture/blobstore.adoc b/docs/modules/servers/partials/architecture/blobstore.adoc
new file mode 100644
index 00000000000..fb4ef523d1e
--- /dev/null
+++ b/docs/modules/servers/partials/architecture/blobstore.adoc
@@ -0,0 +1,108 @@
+James stores large, non-indexable binary payloads in a BlobStore. Typical examples
+are message bodies, attachments, deleted messages retained by the vault, and mail
+queue payloads.
+
+Mailbox, Mail Queue, and Deleted Messages Vault components rely on it.
+
+Server components usually depend on the higher-level `BlobStore`. `BlobStoreDAO`
+is the lower-level virtual storage abstraction implemented by concrete storage
+connectors such as memory, file, Cassandra, Postgres, and S3 compatible object
+stores. It allows James to compose storage features such as encryption or
+compression independently of the storage connector.
+
+== Abstraction layers
+
+Most James components use `BlobStore`, which is responsible for saving content
+and returning a `BlobId`. `BlobStoreDAO` is the lower-level persistence contract:
+it stores, reads, lists, and deletes blobs for a given `BucketName` and `BlobId`.
+
+A `BlobStore` exposes a default logical bucket through `getDefaultBucketName()`.
+Callers can explicitly pass another `BucketName` when they need to store data in
+another logical bucket. More advanced organization rules inside a logical bucket
+are not part of the generic `BlobStore` contract; callers need to model them
+explicitly or provide a custom `BlobStore` implementation.
+
+Cross-cutting storage features can be composed around this DAO contract. For
+example, deduplication decides blob identifiers at the `BlobStore` level, while
+wrappers such as compression or encryption can transform payloads and metadata
+before delegating to the concrete storage connector.
+
+=== BlobStore implementations
+
+James composes several behaviors at the `BlobStore` level:
+
+* `PassThroughBlobStore` is the non-deduplicating strategy. It generates a new
+  `BlobId` for each save, delegates persistence to the configured
+  `BlobStoreDAO`, and deletes blobs directly.
+* `DeDuplicationBlobStore` is the deduplicating strategy. It derives `BlobId`
+  values from content hashes, so identical content can share the same stored
+  blob. A single delete does not remove the underlying blob immediately; garbage
+  collection is responsible for eventually removing unreferenced blobs.
+* `MetricableBlobStore` decorates another `BlobStore` with timing metrics.
+* `CachedBlobStore` decorates another `BlobStore` with a Cassandra-backed cache
+  for small, frequently read blobs.
+
+=== BlobStoreDAO implementations
+
+Concrete `BlobStoreDAO` implementations persist payloads in a storage backend,
+for example memory, file, Cassandra, Postgres, or S3 compatible object storage.
+
+Some `BlobStoreDAO` implementations are wrappers rather than final storage
+connectors:
+
+* `AESBlobStoreDAO` encrypts payload bytes before delegating writes to the
+  underlying DAO, and decrypts them transparently on reads. This protects blob
+  content at rest, especially when James stores blobs in third-party object
+  storage.
+* `ZstdBlobStoreDAO` can compress payload bytes before delegating writes to the
+  underlying DAO. When it stores compressed bytes, it records metadata such as
+  `content-encoding` and the original size. On reads, it uses this metadata to
+  transparently decompress the payload. This reduces storage usage and network
+  transfer for compressible blob content.
+
+AES and Zstd can be enabled together. In the Guice binding chain, compression
+wraps encryption: `ZstdBlobStoreDAO` delegates to `AESBlobStoreDAO`, which then
+delegates to the concrete storage DAO. Writes therefore compress first and
+encrypt afterwards; reads decrypt first and decompress afterwards. This ordering
+preserves the benefit of compression, as encrypted payloads are generally not
+compressible.
+
+== Logical buckets
+
+`BucketName` is a James logical namespace in the `BlobStoreDAO` contract. It is
+not an AWS S3 bucket name, even when the selected connector stores data in an S3
+compatible object store.
+
+Each connector maps this logical namespace to its own storage model. Depending
+on the implementation and configuration, a logical bucket can be stored as a
+directory, an object-storage bucket, a database partition, or another
+connector-specific representation. Code using `BlobStoreDAO` should only rely on
+the James `BucketName` abstraction.
+
+== Metadata
+
+Blob metadata stores side information needed to interpret a blob payload without
+changing the payload bytes or blob identifier. One use case is object store
+compression: James uses a marker such as `content-encoding` to detect a
+compressed payload and transparently decompress it when reading.
+
+James uses a hybrid metadata model:
+
+* Metadata actively interpreted by James should expose typed helpers or constants
+  in the API. For example, `BlobMetadata.contentEncoding()` reads the
+  `content-encoding` entry.
+* Other metadata stays available through the underlying map as an extension point
+  for James library users and custom storage implementations.
+  Custom storage metadata could be used by James library users or storage extensions to implement other use cases.
+
+Metadata-aware storage implementations and wrappers should preserve unknown
+metadata entries.
+
+=== Metadata names
+
+`BlobMetadataName` defines the portable metadata key convention:
+
+* names are case-insensitive and are canonicalized to lowercase;
+* names must be non-empty;
+* names must be shorter than 128 characters;
+* names can contain only ASCII letters, digits, and `-`.
diff --git a/docs/modules/servers/partials/architecture/index.adoc b/docs/modules/servers/partials/architecture/index.adoc
index 449a31c99e3..634004753c0 100644
--- a/docs/modules/servers/partials/architecture/index.adoc
+++ b/docs/modules/servers/partials/architecture/index.adoc
@@ -279,6 +279,8 @@ the same content will be stored one once.
 The downside is that deletion is more complicated, and a garbage collection needs to be run. A first implementation
 based on bloom filters can be used and triggered using the WebAdmin REST API.
 
+See xref:{xref-base}/architecture/blobstore.adoc[BlobStore architecture page] for more details.
+
 === Task Manager
 
 Allows to control and schedule long running tasks run by other
diff --git a/server/blob/blob-api/src/main/java/org/apache/james/blob/api/BlobStore.java b/server/blob/blob-api/src/main/java/org/apache/james/blob/api/BlobStore.java
index 8d5101e0db7..0493e27743b 100644
--- a/server/blob/blob-api/src/main/java/org/apache/james/blob/api/BlobStore.java
+++ b/server/blob/blob-api/src/main/java/org/apache/james/blob/api/BlobStore.java
@@ -27,6 +27,19 @@
 
 import reactor.util.function.Tuple2;
 
+/**
+ * High-level James blob storage abstraction.
+ *
+ * <p>A {@link BlobStore} stores binary payloads in a James logical {@link BucketName}
+ * and returns {@link BlobId} references. A configured {@link BlobStore} exposes a default
+ * logical bucket through {@link #getDefaultBucketName()}.</p>
+ *
+ * <p>A {@link BucketName} is a James-specific logical bucket. Each storage connector decides how this logical
+ * bucket is represented in its backend. It should not be conflated with an S3 bucket name and does not have to map one-to-one
+ * to a physical bucket.</p>
+ *
+ * <p>See {@code docs/modules/servers/partials/architecture/blobstore.adoc} for more details.</p>
+ */
 public interface BlobStore {
     String DEFAULT_BUCKET_NAME_QUALIFIER = "defaultBucket";
 
diff --git a/server/blob/blob-api/src/main/java/org/apache/james/blob/api/BlobStoreDAO.java b/server/blob/blob-api/src/main/java/org/apache/james/blob/api/BlobStoreDAO.java
index ed0679bd510..a527dad8259 100644
--- a/server/blob/blob-api/src/main/java/org/apache/james/blob/api/BlobStoreDAO.java
+++ b/server/blob/blob-api/src/main/java/org/apache/james/blob/api/BlobStoreDAO.java
@@ -38,6 +38,19 @@
 import com.google.common.io.ByteSource;
 import com.google.common.io.FileBackedOutputStream;
 
+/**
+ * James virtual blob store abstraction.
+ *
+ * <p>A {@link BucketName} is a James-specific logical bucket. Each storage connector decides how this logical
+ * bucket is represented in its backend. It should not be conflated with an S3 bucket name and does not have to map one-to-one
+ * to a physical bucket.</p>
+ *
+ * <p>{@link BlobMetadata} is part of the contract so wrapper DAOs and storage implementations can keep side information
+ * needed to interpret a payload, such as compression markers. Metadata actively used by James should expose typed
+ * helpers, while the underlying metadata map remains an extension point for James library users and custom implementations.</p>
+ *
+ * <p>See {@code docs/modules/servers/partials/architecture/blobstore.adoc} for more details.</p>
+ */
 public interface BlobStoreDAO {
     record BlobMetadataName(String name) {
         private static final CharMatcher CHAR_MATCHER = CharMatcher.inRange('a', 'z')
@@ -46,6 +59,7 @@ record BlobMetadataName(String name) {
             .or(CharMatcher.is('-'));
 
         public BlobMetadataName {
+            Preconditions.checkArgument(!name.isEmpty(), "Metadata name cannot be empty");
             Preconditions.checkArgument(CHAR_MATCHER.matchesAllOf(name), "Invalid char in metadata name. Must be a-z,A-Z,0-9 or - got " + name);
             Preconditions.checkArgument(name.length() < 128, "Metadata name is too long. Size exceed 128 chars");
             name = name.toLowerCase(Locale.US);
diff --git a/server/blob/blob-api/src/test/java/org/apache/james/blob/api/BlobMetadataTest.java b/server/blob/blob-api/src/test/java/org/apache/james/blob/api/BlobMetadataTest.java
index 3cabf6991a5..4f36eea2a52 100644
--- a/server/blob/blob-api/src/test/java/org/apache/james/blob/api/BlobMetadataTest.java
+++ b/server/blob/blob-api/src/test/java/org/apache/james/blob/api/BlobMetadataTest.java
@@ -20,10 +20,17 @@
 package org.apache.james.blob.api;
 
 import static org.assertj.core.api.Assertions.assertThat;
+import static org.assertj.core.api.Assertions.assertThatThrownBy;
 
 import org.junit.jupiter.api.Test;
+import org.junit.jupiter.params.ParameterizedTest;
+import org.junit.jupiter.params.provider.CsvSource;
+import org.junit.jupiter.params.provider.ValueSource;
 
 class BlobMetadataTest {
+    private static final String ONE_HUNDRED_TWENTY_SEVEN_CHARS_METADATA_NAME = "a".repeat(127);
+    private static final String ONE_HUNDRED_TWENTY_EIGHT_CHARS_METADATA_NAME = "a".repeat(128);
+
     @Test
     void blobMetadataNameShouldBeCaseInsensitive() {
         assertThat(new BlobStoreDAO.BlobMetadataName("X-Test").name())
@@ -31,4 +38,41 @@ void blobMetadataNameShouldBeCaseInsensitive() {
         assertThat(new BlobStoreDAO.BlobMetadataName("X-Test"))
             .isEqualTo(new BlobStoreDAO.BlobMetadataName("x-test"));
     }
+
+    @ParameterizedTest
+    @CsvSource({
+        "metadata, metadata",
+        "CONTENT-ENCODING, content-encoding",
+        "x-test-123, x-test-123",
+        "A1-B2-C3, a1-b2-c3"
+    })
+    void blobMetadataNameShouldAcceptLettersDigitsAndDash(String rawName, String expectedName) {
+        assertThat(new BlobStoreDAO.BlobMetadataName(rawName).name())
+            .isEqualTo(expectedName);
+    }
+
+    @ParameterizedTest
+    @ValueSource(strings = {"metadata_name", "metadata.name", "metadata name", "metadata/name", "metadata:name", "metadata#name"})
+    void blobMetadataNameShouldRejectUnsupportedCharacters(String rawName) {
+        assertThatThrownBy(() -> new BlobStoreDAO.BlobMetadataName(rawName))
+            .isInstanceOf(IllegalArgumentException.class);
+    }
+
+    @Test
+    void blobMetadataNameShouldRejectEmptyName() {
+        assertThatThrownBy(() -> new BlobStoreDAO.BlobMetadataName(""))
+            .isInstanceOf(IllegalArgumentException.class);
+    }
+
+    @Test
+    void blobMetadataNameShouldAcceptNameBelowOneHundredTwentyEightCharacters() {
+        assertThat(new BlobStoreDAO.BlobMetadataName(ONE_HUNDRED_TWENTY_SEVEN_CHARS_METADATA_NAME).name())
+            .isEqualTo(ONE_HUNDRED_TWENTY_SEVEN_CHARS_METADATA_NAME);
+    }
+
+    @Test
+    void blobMetadataNameShouldRejectNameOfOneHundredTwentyEightCharacters() {
+        assertThatThrownBy(() -> new BlobStoreDAO.BlobMetadataName(ONE_HUNDRED_TWENTY_EIGHT_CHARS_METADATA_NAME))
+            .isInstanceOf(IllegalArgumentException.class);
+    }
 }
\ No newline at end of file