	In Browser
	StumbleUpon
	del.icio.us
	Google
	Google Buzz
	reddit
	LinkedIn

	Facebook
	Twitter
	Linkedin
	E-Mail

Generative AI > Google Gemini API > ADK Episodic Memory Agent

ADK Episodic Memory Agent

Author: Venkata Sudhakar

Standard ADK agents are stateless by default - each session starts fresh with no knowledge of previous interactions. Episodic memory changes this by giving the agent access to a searchable log of past events. When a customer contacts ShopMax India again, the agent can retrieve relevant past interactions and use them to provide a personalised, context-aware response without asking the customer to repeat themselves.

The pattern stores interaction summaries in a simple key-value store keyed by user ID and timestamp. A FunctionTool retrieves the most recent N episodes for a user. The agent reads these episodes as context before responding. In production, replace the in-memory store with Firestore or Cloud Bigtable for durability and scale.

The below example shows a ShopMax India support agent that remembers a customer's previous order and complaint history.

import datetime
import google.genai as genai
from google.adk.agents import LlmAgent
from google.adk.tools import FunctionTool
from google.adk.sessions import InMemorySessionService
from google.adk.runners import Runner

# Episodic memory store: {user_id: [{ts, summary, outcome}]}
MEMORY_STORE = {
    "cust_202": [
        {
            "ts": "2026-03-01 10:23",
            "summary": "Customer ordered Samsung 55-inch TV (SM-TV-55), Rs 72000. Delivery to Pune.",
            "outcome": "Order confirmed, delivery in 3 days"
        },
        {
            "ts": "2026-03-05 14:45",
            "summary": "Customer reported TV remote not working. Raised replacement request REQ-8821.",
            "outcome": "Remote dispatched, expected in 2 days"
        }
    ]
}

def recall_episodes(user_id: str, max_episodes: int = 3) -> dict:
    # Retrieve recent interaction history for user
    episodes = MEMORY_STORE.get(user_id, [])
    recent = episodes[-max_episodes:]
    if not recent:
        return {"found": False, "episodes": []}
    return {"found": True, "episode_count": len(recent), "episodes": recent}

def save_episode(user_id: str, summary: str, outcome: str) -> dict:
    # Persist a new interaction episode
    ts = datetime.datetime.now().strftime("%Y-%m-%d %H:%M")
    episode = {"ts": ts, "summary": summary, "outcome": outcome}
    if user_id not in MEMORY_STORE:
        MEMORY_STORE[user_id] = []
    MEMORY_STORE[user_id].append(episode)
    return {"saved": True, "ts": ts}

recall_tool = FunctionTool(func=recall_episodes)
save_tool = FunctionTool(func=save_episode)

agent = LlmAgent(
    name="shopmax_memory_agent",
    model="gemini-2.0-flash",
    tools=[recall_tool, save_tool],
    instruction="""
    You are a ShopMax India support agent with episodic memory.
    At the start of every interaction:
    1. Call recall_episodes with the user_id to fetch history.
    2. Use that history to personalise your response.
    At the end of every interaction:
    3. Call save_episode to record a brief summary and outcome.
    Be warm, reference past interactions by date, and avoid asking
    customers to repeat information you already have.
    """
)

session_service = InMemorySessionService()
runner = Runner(agent=agent, app_name="shopmax_memory", session_service=session_service)
session = session_service.create_session(app_name="shopmax_memory", user_id="cust_202")

from google.genai import types
query = "user_id=cust_202 | Hi, I am Ananya. Has my remote issue been resolved?"
message = types.Content(role="user", parts=[types.Part(text=query)])
for event in runner.run(user_id="cust_202", session_id=session.id, new_message=message):
    if event.is_final_response():
        print(event.content.parts[0].text)

It gives the following output,

Hello Ananya! Welcome back to ShopMax India.

I can see from our records that on 5th March 2026, you reported that the remote
for your Samsung 55-inch TV was not working, and we raised replacement request
REQ-8821. The replacement remote was dispatched and was expected within 2 days.

Since it has been a few days, the remote should have arrived by now. Did you
receive it? If there are still any issues, I can escalate this immediately.

[Episode saved: Ananya checked remote replacement status for REQ-8821]

Episodic memory transforms a transactional support bot into a relationship-aware agent. Customers feel recognised, reducing frustration and repeat explanations. For ShopMax India, store episodes in Firestore with user_id as the document key and a subcollection of episodes sorted by timestamp. Set a TTL policy to auto-delete episodes older than 12 months to comply with data retention policies. Index the summary field for keyword search when the agent needs to find a specific past episode by topic rather than recency.

Send your comments, suggestions or queries regarding this site to [email protected].