Context caching not working with Agent Engine (VertexAiSessionService drops cache_metadata)

🔴 Required Information

Describe the Bug:
VertexAiSessionService does not persist cache_metadata or usage_metadata fields when writing/reading Event objects to/from Agent Engine. This causes ContextCacheRequestProcessor to never find prior cache metadata, so context caching creates a new cache on every invocation instead of reusing existing ones.

Steps to Reproduce:

Install google-adk[eval]==1.25.1
Configure an agent with ContextCacheConfig in App():

     app = App(   
      name="my_agent",    
      root_agent=root_agent,    
      context_cache_config= ContextCacheConfig( min_tokens=2048,  ttl_seconds=1800,   cache_intervals=10),)

Run with in-memory sessions: adk web — send the same prompt twice in the same session. Observe cacheMetadata in the dev UI showing cache_name reuse on the second turn.
Run with Agent Engine sessions: adk web --session_service_uri=agentengine://<your_engine_id> — send the same prompt twice in the same session. Observe that every turn creates a new cache with a different fingerprint; cache_name is never reused.

Expected Behavior:
cache_metadata and usage_metadata should round-trip through Agent Engine sessions the same way they do with InMemorySessionService. The second invocation in the same session should reuse the existing cache.

Observed Behavior:
With VertexAiSessionService, every invocation creates a new cache. The cacheMetadata in events shows a new fingerprint and increasing contents_count each time, with no cache_name reuse.

Environment Details:

ADK Library Version: 1.25.1
Desktop OS: macOS
Python Version: 3.11

Model Information:

Are you using LiteLLM: Yes/No
Which model is being used: gemini-2.5-pro

🟡 Optional Information

Regression:
N/A — this has never worked with VertexAiSessionService as far as we can tell (tested 1.19.0 and 1.25.1).

Logs:
With InMemorySessionService (working):

cacheMetadata:
  cache_name: "projects/798321343024/locations/us-central1/cachedContents/6261708034305687552"
  expire_time: 1772555630.466322
  fingerprint: "e8e3e4d5226b764d"
  invocations_used: 2
  contents_count: 2
  created_at: 1772553830.466322

With VertexAiSessionService (broken — new cache every turn):

cacheMetadata:
  fingerprint: "c3f575505d7ec95c"
  contents_count: 6

Minimal Reproduction Code:

from google.adk.agents import Agent
from google.adk.app import App
from google.adk.context_cache import ContextCacheConfig

root_agent = Agent(
    name="test_agent",
    model="gemini-2.5-pro",
    instruction="You are a helpful assistant.",
)

app = App(
    name="test_app",
    root_agent=root_agent,
    context_cache_config=ContextCacheConfig(
        min_tokens=2048,
        ttl_seconds=1800,
        cache_intervals=10,
    ),
)

# Works:    adk web
# Broken:   adk web --session_service_uri=agentengine://<engine_id>

How often has this issue occurred?:

Always (100%)

@muzna-naeem

Hello @muzna-naeem ,

Thank you for bringing this to our attention. We were able to reproduce the issue using google-adk==1.26.0.

It appears the failure happens during the if llm_request.cache_metadata check (line no 74) in gemini_context_cache_manager.py. When the ADK web session_service_uri is used llm_request.cache_metadata is always None which causes the check to fail. However, when using the default Inmemorysessionservice, it correctly returns a CacheMetadata object.

We are currently investigating why this discrepancy is happening. We appreciate your patience while we look into this further.

Hello @DeanChensj , could you please have a look at this and share your inputs?

fix: preserve cache_metadata and usage_metadata in VertexAiSessionSer…

Hey, any update on this one?

Is there any update ?

fix: preserve cache_metadata and usage_metadata in VertexAiSessionSer…

Is the intended behavior for ADK ContextCacheConfig to reuse a cache only within the same session, or should it also reuse equivalent caches across sessions when static_instruction + tools docstrings are identical?

Is there any way to create one shared cached content resource for the root agent’s static instruction + tools docstrings, then reuse that same cached content across all Agent Engine sessions when multiple people are talking at the same time?

I really hope things will move forward soon...

@muzna-naeem

Hi @muzna-naeem @ssingh5bmcd @fufufukakaka , this issue should be resolved in the latest a few releases. Have you checked? We introduced blob based event storage in vertex ai session service. Could you double check?

This issue has been automatically marked as stale because it has not had recent activity for 7 days after a maintainer requested clarification. It will be closed if no further activity occurs within 7 days.

@DeanChensj

This has been fixed in 9da9dee! As @DeanChensj said, we store the raw event. If the issue persists, please create another issue.

Navigation Menu

Repository navigation

Context caching not working with Agent Engine (VertexAiSessionService drops cache_metadata) #4698

🔴 Required Information

🟡 Optional Information

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Context caching not working with Agent Engine (VertexAiSessionService drops cache_metadata) #4698

Description

🔴 Required Information

🟡 Optional Information

Activity

llalitkumarrr commented on Mar 4, 2026

muzna-naeem commented on Apr 10, 2026

ssingh5bmcd commented on Apr 16, 2026

josipTheorem258 commented on May 25, 2026

fufufukakaka commented on Jun 9, 2026

DeanChensj commented on Jun 9, 2026

adk-bot commented on Jun 16, 2026

haranrk commented on Jun 17, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions