diff --git a/gapic/cli/generate.py b/gapic/cli/generate.py
index e8eee1f034..212f7de4f7 100644
--- a/gapic/cli/generate.py
+++ b/gapic/cli/generate.py
@@ -23,6 +23,7 @@
 from gapic import generator
 from gapic.schema import api
 from gapic.utils import Options
+from gapic.utils.cache import generation_cache_context
 
 
 @click.command()
@@ -56,15 +57,23 @@ def generate(request: typing.BinaryIO, output: typing.BinaryIO) -> None:
         [p.package for p in req.proto_file if p.name in req.file_to_generate]
     ).rstrip(".")
 
-    # Build the API model object.
-    # This object is a frozen representation of the whole API, and is sent
-    # to each template in the rendering step.
-    api_schema = api.API.build(req.proto_file, opts=opts, package=package)
+    # Create the generation cache context.
+    # This provides the shared storage for the @cached_proto_context decorator.
+    # 1. Performance: Memoizes `with_context` calls, speeding up generation significantly.
+    # 2. Safety: The decorator uses this storage to "pin" Proto objects in memory.
+    #    This prevents Python's Garbage Collector from deleting objects created during
+    #    `API.build` while `Generator.get_response` is still using their IDs.
+    #    (See `gapic.utils.cache.cached_proto_context` for the specific pinning logic).
+    with generation_cache_context():
+        # Build the API model object.
+        # This object is a frozen representation of the whole API, and is sent
+        # to each template in the rendering step.
+        api_schema = api.API.build(req.proto_file, opts=opts, package=package)
 
-    # Translate into a protobuf CodeGeneratorResponse; this reads the
-    # individual templates and renders them.
-    # If there are issues, error out appropriately.
-    res = generator.Generator(opts).get_response(api_schema, opts)
+        # Translate into a protobuf CodeGeneratorResponse; this reads the
+        # individual templates and renders them.
+        # If there are issues, error out appropriately.
+        res = generator.Generator(opts).get_response(api_schema, opts)
 
     # Output the serialized response.
     output.write(res.SerializeToString())
diff --git a/gapic/schema/metadata.py b/gapic/schema/metadata.py
index 7df8d0291f..8defa70dc2 100644
--- a/gapic/schema/metadata.py
+++ b/gapic/schema/metadata.py
@@ -35,6 +35,7 @@
 from gapic.schema import imp
 from gapic.schema import naming
 from gapic.utils import cached_property
+from gapic.utils import cached_proto_context
 from gapic.utils import RESERVED_NAMES
 
 # This class is a minor hack to optimize Address's __eq__ method.
@@ -359,6 +360,7 @@ def resolve(self, selector: str) -> str:
             return f'{".".join(self.package)}.{selector}'
         return selector
 
+    @cached_proto_context
     def with_context(self, *, collisions: Set[str]) -> "Address":
         """Return a derivative of this address with the provided context.
 
@@ -398,6 +400,7 @@ def doc(self):
             return "\n\n".join(self.documentation.leading_detached_comments)
         return ""
 
+    @cached_proto_context
     def with_context(self, *, collisions: Set[str]) -> "Metadata":
         """Return a derivative of this metadata with the provided context.
 
diff --git a/gapic/schema/wrappers.py b/gapic/schema/wrappers.py
index 17a7832756..3c3c3d5849 100644
--- a/gapic/schema/wrappers.py
+++ b/gapic/schema/wrappers.py
@@ -67,6 +67,7 @@
 
 from gapic import utils
 from gapic.schema import metadata
+from gapic.utils import cached_proto_context
 from gapic.utils import uri_sample
 from gapic.utils import make_private
 
@@ -410,6 +411,7 @@ def type(self) -> Union["MessageType", "EnumType", "PrimitiveType"]:
             "This code should not be reachable; please file a bug."
         )
 
+    @cached_proto_context
     def with_context(
         self,
         *,
@@ -805,6 +807,7 @@ def get_field(
         # message.
         return cursor.message.get_field(*field_path[1:], collisions=collisions)
 
+    @cached_proto_context
     def with_context(
         self,
         *,
@@ -937,6 +940,7 @@ def ident(self) -> metadata.Address:
         """Return the identifier data to be used in templates."""
         return self.meta.address
 
+    @cached_proto_context
     def with_context(self, *, collisions: Set[str]) -> "EnumType":
         """Return a derivative of this enum with the provided context.
 
@@ -1058,6 +1062,7 @@ class ExtendedOperationInfo:
     request_type: MessageType
     operation_type: MessageType
 
+    @cached_proto_context
     def with_context(
         self,
         *,
@@ -1127,6 +1132,7 @@ class OperationInfo:
     response_type: MessageType
     metadata_type: MessageType
 
+    @cached_proto_context
     def with_context(
         self,
         *,
@@ -1937,6 +1943,7 @@ def void(self) -> bool:
         """Return True if this method has no return value, False otherwise."""
         return self.output.ident.proto == "google.protobuf.Empty"
 
+    @cached_proto_context
     def with_context(
         self,
         *,
@@ -2357,6 +2364,7 @@ def operation_polling_method(self) -> Optional[Method]:
     def is_internal(self) -> bool:
         return any(m.is_internal for m in self.methods.values())
 
+    @cached_proto_context
     def with_context(
         self,
         *,
diff --git a/gapic/utils/__init__.py b/gapic/utils/__init__.py
index 8b48801730..23c5739156 100644
--- a/gapic/utils/__init__.py
+++ b/gapic/utils/__init__.py
@@ -13,6 +13,7 @@
 # limitations under the License.
 
 from gapic.utils.cache import cached_property
+from gapic.utils.cache import cached_proto_context
 from gapic.utils.case import to_snake_case
 from gapic.utils.case import to_camel_case
 from gapic.utils.checks import is_msg_field_pb
@@ -34,6 +35,7 @@
 
 __all__ = (
     "cached_property",
+    "cached_proto_context",
     "convert_uri_fieldnames",
     "doc",
     "empty",
diff --git a/gapic/utils/cache.py b/gapic/utils/cache.py
index f9c4d703f5..5f9b0f658d 100644
--- a/gapic/utils/cache.py
+++ b/gapic/utils/cache.py
@@ -13,6 +13,8 @@
 # limitations under the License.
 
 import functools
+import contextlib
+import threading
 
 
 def cached_property(fx):
@@ -43,3 +45,91 @@ def inner(self):
         return self._cached_values[fx.__name__]
 
     return property(inner)
+
+
+# Thread-local storage for the simple cache dictionary.
+# This ensures that parallel generation tasks (if any) do not corrupt each other's cache.
+_thread_local = threading.local()
+
+
+@contextlib.contextmanager
+def generation_cache_context():
+    """Context manager to explicitly manage the lifecycle of the generation cache.
+
+    This manager initializes a fresh dictionary in thread-local storage when entering
+    the context and strictly deletes it when exiting.
+
+    **Memory Management:**
+    The cache stores strong references to Proto objects to "pin" them in memory
+    (see `cached_proto_context`). It is critical that this context manager deletes
+    the dictionary in the `finally` block. Deleting the dictionary breaks the
+    reference chain, allowing Python's Garbage Collector to finally free all the
+    large Proto objects that were pinned during generation.
+    """
+    # Initialize the cache as a standard dictionary.
+    _thread_local.cache = {}
+    try:
+        yield
+    finally:
+        # Delete the dictionary to free all memory and pinned objects.
+        # This is essential to prevent memory leaks in long-running processes.
+        del _thread_local.cache
+
+
+def cached_proto_context(func):
+    """Decorator to memoize `with_context` calls based on object identity and collisions.
+
+    This mechanism provides a significant performance boost by preventing
+    redundant recalculations of naming collisions during template rendering.
+
+    Since the Proto wrapper objects are unhashable (mutable), we use `id(self)` as
+    the primary cache key. Normally, this is dangerous: if the object is garbage
+    collected, Python might reuse its memory address for a *new* object, leading to
+    a cache collision (the "Zombie ID" bug).
+
+    To prevent this, this decorator stores the value as a tuple: `(result, self)`.
+    By keeping a reference to `self` in the cache value, we "pin" the object in
+    memory. This forces the Garbage Collector to keep the object alive, guaranteeing
+    that `id(self)` remains unique for the entire lifespan of the `generation_cache_context`.
+
+    Args:
+        func (Callable): The function to decorate (usually `with_context`).
+
+    Returns:
+        Callable: The wrapped function with caching and pinning logic.
+    """
+
+    @functools.wraps(func)
+    def wrapper(self, *, collisions, **kwargs):
+
+        # 1. Check for active cache (returns None if context is not active)
+        context_cache = getattr(_thread_local, "cache", None)
+
+        # If we are not inside a generation_cache_context (e.g. unit tests),
+        # bypass the cache entirely.
+        if context_cache is None:
+            return func(self, collisions=collisions, **kwargs)
+
+        # 2. Create the cache key
+        # We use frozenset for collisions to make it hashable.
+        # We use id(self) because 'self' is not hashable.
+        collisions_key = frozenset(collisions) if collisions else None
+        key = (id(self), collisions_key)
+
+        # 3. Check Cache
+        if key in context_cache:
+            # The cache stores (result, pinned_object). We return just the result.
+            return context_cache[key][0]
+
+        # 4. Execute the actual function
+        # We ensure context_cache is passed down to the recursive calls
+        result = func(self, collisions=collisions, **kwargs)
+
+        # 5. Update Cache & Pin Object
+        # We store (result, self). The reference to 'self' prevents garbage collection,
+        # ensuring that 'id(self)' cannot be reused for a new object while this
+        # cache entry exists.
+        context_cache[key] = (result, self)
+        return result
+
+    return wrapper
diff --git a/tests/unit/utils/test_cache.py b/tests/unit/utils/test_cache.py
index eee6865e6d..ec7f82438b 100644
--- a/tests/unit/utils/test_cache.py
+++ b/tests/unit/utils/test_cache.py
@@ -31,3 +31,43 @@ def bar(self):
     assert foo.call_count == 1
     assert foo.bar == 42
     assert foo.call_count == 1
+
+
+def test_cached_proto_context():
+    class Foo:
+        def __init__(self):
+            self.call_count = 0
+
+        # We define a signature that matches the real Proto.with_context
+        # to ensure arguments are propagated correctly.
+        @cache.cached_proto_context
+        def with_context(self, collisions, *, skip_fields=False, visited_messages=None):
+            self.call_count += 1
+            return f"val-{self.call_count}"
+
+    foo = Foo()
+
+    # 1. Test Bypass (No Context)
+    # The cache is not active, so every call increments the counter.
+    assert foo.with_context(collisions={"a"}) == "val-1"
+    assert foo.with_context(collisions={"a"}) == "val-2"
+
+    # 2. Test Context Activation
+    with cache.generation_cache_context():
+        # Reset counter to make tracking easier
+        foo.call_count = 0
+
+        # A. Basic Cache Hit
+        assert foo.with_context(collisions={"a"}) == "val-1", "a"
+        assert foo.with_context(collisions={"a"}) == "val-1"  # Hit
+        assert foo.call_count == 1
+
+        # B. Collision Difference
+        # Changing collisions creates a new key
+        assert foo.with_context(collisions={"b"}) == "val-2"
+        assert foo.call_count == 2
+
+    # 3. Context Cleared
+    # Everything should be forgotten now.
+    assert getattr(cache._thread_local, "cache", None) is None
+    assert foo.with_context(collisions={"a"}) == "val-3"