Your AI agent is forgetful. Here’s how to give it a brain.

0:00 MARTIN OMANDER: Hi, agent.

0:01 I'm planning a two-day trip to Tokyo, and I love historic sites.

0:06 ANNIE WANG: Great.

0:07 I'll plan a visit to the imperial palace.

0:09 MARTIN OMANDER: Perfect.

0:10 Can you send me the itinerary, agent?

0:13 ANNIE WANG: Hi.

0:14 How can I help you plan a trip today?

0:15 MARTIN OMANDER: What?

0:16 [CHUCKLING] ANNIE WANG: Yeah.

0:19 This is a goldfish memory problem.

0:22 So even if our AI agent is brilliant,

0:25 it will look dumb if it doesn't remember anything.

0:28 And let's see how we can fix that.

0:30 [MUSIC PLAYING] MARTIN OMANDER: Welcome back to the show, Annie.

0:41 What's your story at Google?

0:43 ANNIE WANG: Oh, thanks, Martin.

0:44 I spent years as a software engineer here,

0:47 but I realized that my real passion is AI education.

0:51 So now, I'm over on the other side, helping developers, building AI agents.

0:55 MARTIN OMANDER: Yeah.

0:56 And you've been deep in the world of AI agents, right?

0:59 ANNIE WANG: Yeah.

1:00 I've noticed a pattern.

1:01 Everyone thinks AI intelligence is just about model and tooling.

1:06 And they're forgetting the other important piece that is memory.

1:10 MARTIN OMANDER: Guilty.

1:11 I definitely focus on the model first when I'm building AI apps.

1:14 ANNIE WANG: Yeah.

1:15 Most people do.

1:16 But if you want to build an app that actually impresses people,

1:19 you need an agent that uses a full brain.

1:23 And here, we have six patterns for adding agent memories.

1:26 Let's talk about each one and when you should use it.

1:30 I also built a Codelab for this so that you can try it out yourself.

1:33 MARTIN OMANDER: OK.

1:34 We'll add a link to your Codelab, Annie.

1:36 Let's dive in.

1:37 ANNIE WANG: All right.

1:38 So when you talk with a person,

1:40 it will be weird if they forgot what you just told them a minute ago, right?

1:44 So the same goes for agents.

1:47 That's why your agent needs sessions to hold all the conversation history.

1:52 MARTIN OMANDER: Makes sense.

1:53 And how do you implement that?

1:55 ANNIE WANG: Yeah.

1:55 If you're using Google Agent Development Kit, ADK, it's fairly easy.

2:00 Just create and use a session object.

2:04 MARTIN OMANDER: So now the agent will remember where the user wants to go?

2:07 ANNIE WANG: Yeah.

2:08 And let's try it.

2:09 As you can see, I will enter that I

2:12 am going to Tokyo and that I like historical sites.

2:18 And then, I will ask for an itinerary.

2:22 And you can see that the agent will give a plan for each day.

2:27 And after I confirm for each day,

2:29 it will eventually give me the three-days plan.

2:33 MARTIN OMANDER: And it worked.

2:34 ANNIE WANG: Yay.

2:34 MARTIN OMANDER: The agent does not have the memory of a goldfish anymore.

2:38 ANNIE WANG: Yeah.

2:38 So session is the most basic type of agent memory.

2:42 But in complex apps, you often have a team of agents working together.

2:47 MARTIN OMANDER: So the agents, they need to share memory between them somehow?

2:51 ANNIE WANG: Yeah, exactly.

2:52 So here is we use state to share context among agents.

2:58 And that state is basically a shared digital folder for the entire session.

3:04 And here is an example.

3:06 I will ask for the best sushi in Palo Alto.

3:10 And behind the scene, there is a foodie agent that finds a restaurant

3:15 and a navigation agent that gets direction to it.

3:20 You can see the state value on the ADK web UI.

3:24 MARTIN OMANDER: And what does the code look like for this?

3:26 ANNIE WANG: Sure.

3:27 Here is the definition of the foodie agent.

3:31 And here's the definition of the transportation agent.

3:35 Martin, do you see the destination string in both of them?

3:38 MARTIN OMANDER: Mm-hmm, I do.

3:39 ANNIE WANG: Yeah?

3:40 So the foodie agent saves its result to the key called destination.

3:45 And then, the transportation agent uses

3:47 that key in its prompt with curly brackets.

3:49 MARTIN OMANDER: OK.

3:50 So we have two agents.

3:52 How do we now put them together?

3:54 ANNIE WANG: So we create a route agent of type, sequential agent,

3:57 that calls the foodie agent first and the navigation agent second.

4:01 MARTIN OMANDER: OK.

4:02 So our agents are now talking to each other.

4:04 But what happens if I close the app or the server reboots?

4:08 ANNIE WANG: Oh, that would be total memory loss.

4:11 Everything we've done so far has been in-memory.

4:14 If the script stops, the agent forgets that you ever existed.

4:18 MARTIN OMANDER: Oh, no!

4:20 And how do we make memory stick between restarts?

4:23 ANNIE WANG: So we swap the in-memory session service for a persistent database.

4:27 By using a database session service,

4:30 the agent can look back at the conversations you had days,

4:34 weeks, or even months ago.

4:35 MARTIN OMANDER: And I'm guessing that's how we get that personal assistant feel?

4:40 It remembers my preferences over time?

4:43 ANNIE WANG: Exactly.

4:44 And here is how to implement it.

4:46 So first, we retrieve the old session from the session service.

4:51 And then, we build a previous context from the session object.

4:56 And now that we have the previous context, we can add it to the query.

5:01 MARTIN OMANDER: Very nice.

5:02 ANNIE WANG: Yeah.

5:03 And those were the first three ways of giving your agent memory.

5:08 Let's talk about the other three in our next video.

5:11 MARTIN OMANDER: Excellent, Annie.

5:12 Can you give us a quick recap, please?

5:15 ANNIE WANG: Of course, Martin.

5:16 So first, we use session memory to hold a conversation without forgetting.

5:21 And secondly, we use multi-agent state to share notes between agents.

5:27 And lastly, we use persistence to remember things even after a system reboots.

5:32 MARTIN OMANDER: That's a great list.

5:34 Thank you for sharing this with us, Annie.

5:36 And I guess we'll see each other in the next video.

5:38 ANNIE WANG: Yeah.

5:39 Thanks for having me, Martin.

5:40 MARTIN OMANDER: And thank you, everyone, for watching.

5:43 If you have any questions for Annie or me, please let us know in the comments.

5:47 Also, do let me know what you thought of today's episode.

5:51 I read every single comment.

5:54 Until next time.

5:55 [MUSIC PLAYING]