diff --git a/core/src/commonMain/kotlin/com/google/adk/kt/models/Gemini.kt b/core/src/commonMain/kotlin/com/google/adk/kt/models/Gemini.kt
index 3b04d9c2..aedc7aa9 100644
--- a/core/src/commonMain/kotlin/com/google/adk/kt/models/Gemini.kt
+++ b/core/src/commonMain/kotlin/com/google/adk/kt/models/Gemini.kt
@@ -137,7 +137,7 @@ class Gemini(
     // existing cache (dropping the cached prefix) and returns the metadata to attach to responses.
     val cacheManager =
       preparedRequest.cacheConfig?.let {
-        GeminiContextCacheManager(name, GenaiCacheClient(client.caches))
+        GeminiContextCacheManager(name, GenaiCacheClient(client.caches), cacheScopeOf(client))
       }
     val cacheResult = cacheManager?.handleContextCaching(preparedRequest)
     val finalRequest = cacheResult?.request ?: preparedRequest
@@ -232,6 +232,20 @@ class Gemini(
   }
 }
 
+/**
+ * The backend namespace that owns explicit cache resources, folded into the cache fingerprint so a
+ * cache is never reused across backends or projects. Includes the backend type and, for the
+ * enterprise backend, the project and location. The server base URL is not included because the SDK
+ * [Client] does not expose it.
+ */
+internal fun cacheScopeOf(client: Client): Map<String, String> = buildMap {
+  this["backend"] = if (client.enterprise) "vertex" else "gemini"
+  if (client.enterprise) {
+    client.project?.let { this["project"] = it }
+    client.location?.let { this["location"] = it }
+  }
+}
+
 /**
  * Prepares an [LlmRequest] for the GenerateContent API.
  *
diff --git a/core/src/commonMain/kotlin/com/google/adk/kt/models/GeminiContextCacheManager.kt b/core/src/commonMain/kotlin/com/google/adk/kt/models/GeminiContextCacheManager.kt
index 6c22d905..5c4fe393 100644
--- a/core/src/commonMain/kotlin/com/google/adk/kt/models/GeminiContextCacheManager.kt
+++ b/core/src/commonMain/kotlin/com/google/adk/kt/models/GeminiContextCacheManager.kt
@@ -46,10 +46,14 @@ import kotlinx.serialization.json.encodeToJsonElement
  *
  * @param modelName The model resource name used when creating caches.
  * @param cacheClient The client used to create and delete Gemini caches.
+ * @param cacheScope The backend namespace that owns cache resources (e.g. backend type and, for the
+ *   enterprise backend, project and location). Folded into the fingerprint so a cache is never
+ *   reused across backends or projects.
  */
 internal class GeminiContextCacheManager(
   private val modelName: String,
   private val cacheClient: CacheClient,
+  private val cacheScope: Map<String, String> = emptyMap(),
 ) {
 
   /**
@@ -174,17 +178,19 @@ internal class GeminiContextCacheManager(
   }
 
   /**
-   * Generates a 16-character fingerprint over the cacheable state: model, system instruction,
-   * tools, tool config, and the first [cacheContentsCount] contents. Serializes them into a single
-   * canonical JSON object with every object's keys recursively sorted, so the fingerprint depends
-   * only on content, not on map insertion or property order, and has no field-boundary ambiguity.
-   * The model is included because explicit caches are model-specific.
+   * Generates a 16-character fingerprint over the cacheable state: model, cache scope, system
+   * instruction, tools, tool config, and the first [cacheContentsCount] contents. Serializes them
+   * into a single canonical JSON object with every object's keys recursively sorted, so the
+   * fingerprint depends only on content, not on map insertion or property order, and has no
+   * field-boundary ambiguity. The model and cache scope are included because explicit caches are
+   * model- and backend-specific.
    */
   @OptIn(FrameworkInternalApi::class)
   private fun generateFingerprint(request: LlmRequest, cacheContentsCount: Int): String {
     val fields =
       buildMap<String, JsonElement> {
         this["model"] = JsonPrimitive(modelName)
+        this["cache_scope"] = adkJson.encodeToJsonElement(cacheScope)
         request.config.systemInstruction?.let {
           this["system_instruction"] = adkJson.encodeToJsonElement(it)
         }
diff --git a/core/src/commonTest/kotlin/com/google/adk/kt/models/GeminiContextCacheManagerTest.kt b/core/src/commonTest/kotlin/com/google/adk/kt/models/GeminiContextCacheManagerTest.kt
index ed41c24a..22bd66b7 100644
--- a/core/src/commonTest/kotlin/com/google/adk/kt/models/GeminiContextCacheManagerTest.kt
+++ b/core/src/commonTest/kotlin/com/google/adk/kt/models/GeminiContextCacheManagerTest.kt
@@ -243,6 +243,37 @@ class GeminiContextCacheManagerTest {
     assertNull(result.cacheMetadata?.cacheName)
   }
 
+  @Test
+  fun handleContextCaching_differentBackend_doesNotReuseCache() = runTest {
+    val request = baseRequest()
+    // A cache fingerprinted under the Gemini backend must not validate under Vertex, because the
+    // cache scope (backend/project/location) is part of the fingerprint.
+    val geminiManager =
+      GeminiContextCacheManager("gemini-2.0-flash", FakeCacheClient(), mapOf("backend" to "gemini"))
+    val activeMetadata =
+      CacheMetadata(
+        fingerprint = fingerprintFor(geminiManager, request),
+        contentsCount = 1,
+        cacheName = "cache/gemini",
+        expireTime = Clock.System.now().toEpochMilliseconds() + 100_000,
+        invocationsUsed = 1,
+      )
+
+    val vertexFake = FakeCacheClient()
+    val vertexManager =
+      GeminiContextCacheManager(
+        "gemini-2.0-flash",
+        vertexFake,
+        mapOf("backend" to "vertex", "project" to "p", "location" to "us-central1"),
+      )
+    val result = vertexManager.handleContextCaching(request.copy(cacheMetadata = activeMetadata))
+
+    assertNull(result.request.config.cachedContent)
+    assertEquals(1, vertexFake.deleteCount)
+    assertEquals("cache/gemini", vertexFake.lastDeletedName)
+    assertNull(result.cacheMetadata?.cacheName)
+  }
+
   @Test
   fun handleContextCaching_validCache_stripsToolConfig() = runTest {
     val fake = FakeCacheClient()