シャドーイング練習: AI Agent Memory Architecture: Structured SQL, Semantic Vector Search & Memory Bank - 動画で英語スピーキングを学ぶ

練習モードを選ぶ

各モードは全画面の練習ルームで開きます。

173 文

文が短すぎたり長すぎる場合は、Editをタップして調整してください。

Hi, I'm Annie.

If your agent can hear, see, and take action, the next question is, what should I remember after the session ends?

In this episode, I will show you the architecture for long -term memory in a multi -model agent,

facts in structure memory, understanding in semantic memory, and raw audio or video outside of memory.

Hi there.

Hey there.

Good for that customer dinner tomorrow

or just browsing oh i want

you suggest something for customer dinner tomorrow based on my wish list for sure you have those leather ankle boots

and a classic black pump on your list either would go great with

that red cocktail dress you bought but

if you want something a bit more relaxed for dinner maybe mix it up with the plaid button -up

and classic denim jeans all right let's start with the two types of the memory when you're talking to a multimodal agent.

It feels like the agent remembers everything.

It remembers what you said five minutes ago.

It can use what it just heard.

But most of that is short -term memory.

It is like RAM.

It is useful while the session is running, but it's gone when the session ends.

So the real design question is not, can we save everything?

The real question is, what is worth carrying into the next session?

And And that memory comes in two kinds.

The first kind is facts.

Facts are exact.

You have a dinner on Thursday.

You bought the blue jacket.

Those are things you can put in rows and columns.

They have IDs.

They have dates.

They can be created, updated, and deleted.

That is structured memory.

The second kind is understanding.

Understanding is not one clean role in a table.

It is an age in learning what something means about you.

You prefer simple over flashy, you run warm, you reach for dark colors when you want to feel confident.

That is not just a fact to look up.

That is a pattern.

It is a difference between knowing your calendar and knowing you.

So, here is the first takeaway.

Facts are exact.

You look them up.

Understanding is meaning.

You recall it.

Now let's talk about the decision that makes or breaks your memory system.

Those two kinds of memory should not live in the same place.

Facts belong in a structured database.

That usually means SQL, Postgres, MySQL, Cloud SQL, and Google Cloud whatever fits your stack.

Why SQL?

Because facts need exact answers.

If I ask what is on my calendar Thursday, I do not want a similar memory.

I want an exact event.

The database should return the actual roles.

That is what structured databases are good at.

Now, understanding belongs somewhere else.

Understanding stays in a vector store.

A vector store lets you search by meaning.

So instead of asking, find a role where preference equals dark colors,

the agent can ask, what do I know about any style preference for this kind of event?

And the memory system can bring back memories that are semantically related.

And that is the job of embeddings and vector search.

Here is an important detail.

VectorStore is a category.

It is not one specific product.

You could use PG Vector inside Postgres.

You could use AlloyDB.

On Google Cloud, one managed option is agent platform memory bank.

Memory bank is useful because it does more than store vectors.

You can read past sessions, extract useful memories, and update those memories over time so

if you use a plain vector store you usually build

that extraction logic yourself if you use memory bank

that part is managed for you

but the lesson is not you must use memory bank the lesson is understanding goes in a system

that can search by meaning now here's a mistake many people make they hear semantic memory

and they put everything into the vector store.

Facts, preferences, calendar events, everything.

That sounds simple, but it creates a real problem.

A vector store is designed for similarity.

Similarity is great when the question is fuzzy, but similarity is dangerous.

when the answer must be exact.

If the user asks, what did I buy?

The agent should not guess from a similar memory.

It should curate the database.

If the user asks, what do I usually like for dinners?

Then semantic memory makes sense.

That is a meaning -based question.

So the rule is, structured facts goes in database like Cloud SQL.

Unstructured understanding goes in a vector store.

And there is one more category people often forget, raw perception.

The camera frames, the audio waveform, the live stream itself, that does not usually belong in memory.

Your agent does not need to permanently remember every pixel or every sound wave.

It needs the meaning you extract from them.

For example, she is wearing that black hat again.

Or that you just said this dinner is formal.

Those meanings can become memory.

The raw video and audio are separate storage decisions.

You might keep them for replay, audit, debugging, or training data.

But that is not the same thing as agent memory.

Agent memory should keep what matters, not every raw signal that pass through the app.

So here is a second takeaway.

100

Facts go in structured memory, like classical.

101

Understanding goes in a vector store, like agent, platform, memory, bank.

102

Role perception usually stays out of memory, unless you have a separate reason to store the media itself.

103

Alright, let's make this concrete.

104

How does the agent actually remember?

105

There are two important moments, the start of a session and the end of a session.

106

At the start, the agent loads memory before the conversation really begins.

107

It reads structured facts from the database, for example, upcoming occasions, wishlist items, profile details.

108

Then it searches the vector store for a relevant understanding, for example, style preference, comfort constraints, favorite colors from past conversations.

109

Then the agent gets both kinds of context, so when the user starts talking, the agent is not starting from zero,

110

it already has the facts and it already has a meaning.

111

That is why the code open works.

112

The agent can say, you have that dinner on Thursday from the structure database

113

and I know you like to keep it simple from semantic memory.

114

One answer, two memory systems.

115

Now let's talk about saving memory.

116

Facts should usually be saved immediately.

117

If the user says I Add that Blazor to my wish list, the agent should write that to the database right away.

118

That is a tool call.

119

It is structured, it is exact, and the user expects it to be saved now.

120

Understanding is different.

121

You usually do not want to update someone's long -term identity from one random sentence.

122

If I say, maybe I like yellow today, that does not mean yellow is now my whole style.

123

Understanding needs judgment.

124

It should come from stronger signals.

125

Maybe the user explains why something matters.

126

Maybe the same pattern repeats across sessions.

127

That is how you keep memory from getting noisy.

128

So instead of treating every sentence as a permanent preference, you let the memory series review the conversation and extract what is actually worth keeping.

129

Now, when does the extraction happen?

130

You have a choice.

131

You can extract memory during the conversation, after each turn, or at interverse.

132

Agent Platform Memory Bank supports this directly.

133

You can trigger memory generation mid -session, not only at the end

134

and extracting as you go has real advantages shorter exchanges often produce cleaner memories

135

and if the user drops off early you have already saved what they said

136

but there's a catch in a live multi -modal agent extraction

137

is heavier than a database write a model has to read the conversation

138

and infer meaning that takes time and a live voice agent is latency -sensitive,

139

you cannot let a slow extraction step block the audio.

140

So the rule is not always wait until the session ends.

141

The rule is keep extraction off the live path.

142

Run it asynchronously, after the turn or after the session, wherever it will not stall the conversation.

143

And in this demo, the structured facts are written immediately during the conversation.

144

The understanding is extracted off the hot path after the session and stored as semantic memory.

145

That gives you two tracks, fast synchronous writes for facts, slower asynchronous extraction for understanding,

146

and the next time the user comes back and the agent loads both again.

147

That is the loop.

148

Read at the start, write what changes, bring it forward next time.

149

Here's the third takeaway.

150

So both memory systems, when the session starts, save facts as soon as it happens,

151

save understanding after you have enough context to know what is actually worth remembering.

152

All right, let's recap.

153

If you want your AI agent to remember users across sessions, do not start by dumping everything into one type of database.

154

Start by asking, what kind of memory is this?

155

There are two main kinds.

156

Facts are exact.

157

It looks them up.

158

Understanding is meaning.

159

You recall it.

160

Facts belong in a structural database like classical.

161

Understanding belongs in a vector store.

162

Agent Platform Memory Bank is one managed option because it can help extract and maintain that semantic memory for you.

163

But the architecture works beyond one product and raw camera frames or raw audio do not automatically become memory.

164

You usually store the meaning you extract from them, not the raw signal itself.

165

That is what turns a forgetful agent into an agent that can actually pick up where it left off.

166

It knows the facts of your life and it understands what patterns that makes those facts useful.

167

That's it.

168

So what is the first thing you would want your agent to remember about you?

169

Don't forget to leave that in the comments.

170

In the next episode, we will give the agent eyes with real -time vision.

171

All right, I will see you in the next episode.

172

Bye.

173

Thank you.

アプリをダウンロード

流暢に話すために必要なすべて

AI発音採点文ごとに採点

IPA練習すべての音をマスター

単語学習自分の単語帳を作る

単語ゲーム遊びながら覚える

QRをスキャン

Download on theApp Store Get it onGoogle Play

同じチャンネル

FOR YOU

このレッスンの語彙とスピーキングのポイント

このスピーキングレッスンは、動画「AI Agent Memory Architecture: Structured SQL, Semantic Vector Search & Memory Bank」を教材にしています。繰り返し出てくる語は次のとおりです：memory, agent, understanding, session, vector。この動画には、シャドーイング用の文が173文、単語が1681語あります。音声の長さは12:17です。話す速さは1分あたり約137語で安定しており、シャドーイングしやすいペースです。単語の84%は英語の頻出3,000語に含まれます。残りは練習の前に確認しておきましょう。

この動画の重要語彙

動画に出てくる覚えておきたい単語15語を、発音と意味つきで紹介します。

単語	発音	意味
vector 名詞	/ˈvɛktɚ/	ベクトル
database 名詞	/ˈdeɪtəˌbeɪs/	データベース
extract 名詞	/ˈɛkstɹækt/	エキス, 越幾斯
audio 形容詞	/ˈɔ.di.oʊ/	オーディオ
belong 動詞	/bɪˈlɒŋ/	所属する, 属する
preference 名詞	/ˈpɹɛf(ə)ɹ(ə)ns/	選択
calendar 名詞	/ˈkæl.ən.də/	暦, カレンダー
similarity 名詞	/ˌsɪm.ɪˈlæɹ.ɪ.ti/	類似, 類似性
lesson 名詞	/ˈlɛs.ən/	授業, 学科
architecture 名詞	/ˈɑː.kɪˌtɛk.t͡ʃə/	建築, 建築学
recall 動詞	/ɹɪˈkɔl/	呼び返す
perception 名詞	/pɚˈsɛpʃ(ə)n/	知覚
sensitive 形容詞	/ˈsɛn.sɪt.ɪv/	敏感な
formal 形容詞	/ˈfɔɹməl/	正式, 公式
concrete 形容詞	/ˌkɑnˈkɹit/	実際の, 具体的

動画に出てくる句動詞

単語	発音	意味
come back 動詞		戻る, 帰る
drop off 動詞		落ちる
look up 動詞	/ˌlʊk ˈʌp/	見上げる
pick up 動詞		拾う, 拾い上げる

繰り返し練習したい文

動画に出てくる短く完結した文です。日常会話でそのまま使えます。

If the user asks, what did I buy?
If the user asks, what do I usually like for dinners?
Alright, let's make this concrete.

注意したい発音

話し手はdon't, I'm, you'reなど、短縮形や弱形を3回使っています。聞こえたとおりの短い形で発音しましょう。

「sh」と「zh」の音: extraction /ɪkˈstɹækʃən/, perception /pɚˈsɛpʃ(ə)n/, occasion /əˈkeɪ̯ʒən/, structural /ˈstɹʌk(t͡)ʃəɹəl/, flashy /ˈflæʃi/
長い単語（アクセントの位置に注意）: similarity /ˌsɪm.ɪˈlæɹ.ɪ.ti/, architecture /ˈɑː.kɪˌtɛk.t͡ʃə/, permanently /ˈpɜː.mə.nənt.li/, asynchronous /eɪˈsɪŋ.kɹə.nəs/

日本語話者が苦手な音：

/r/ と /l/ の区別 — /l/ は舌先を歯茎につけ、/r/ はどこにもつけない: similarity /ˌsɪm.ɪˈlæɹ.ɪ.ti/, recall /ɹɪˈkɔl/, formal /ˈfɔɹməl/, relax /ɹɪˈlæks/, structural /ˈstɹʌk(t͡)ʃəɹəl/
/v/ — /b/ にならないように、上の歯を下唇に当てる: vector /ˈvɛktɚ/, sensitive /ˈsɛn.sɪt.ɪv/, wherever /(h)wɛɹˈɛvɚ/
/f/ — 「フ」ではなく、上の歯と下唇で出す: preference /ˈpɹɛf(ə)ɹ(ə)ns/, formal /ˈfɔɹməl/, comfort /ˈkʌm.fət/, fuzzy /ˈfʌzi/, flashy /ˈflæʃi/

この動画での練習方法

まず声を出さずに動画を最後まで聞き、知らない単語をメモします。
通常の速度で一文ずつシャドーイングし、話し手のリズムに合うまで繰り返します。
自分の声を録音して元の音声と比べます。vector, database, extractなどの単語に特に注意しましょう。

シャドーイングとは？英語上達に効果的な理由

シャドーイング（Shadowing）は、もともとプロの通訳者養成プログラムで開発された言語学習法で、多言語習得者として知られるDr. Alexander Arguelles によって広く普及されました。方法はシンプルですが非常に効果的：ネイティブスピーカーの英語を聞きながら、1〜2秒の遅延で声に出してすぐに繰り返す——まるで「影（shadow）」のように話者を追いかけます。文法ドリルや受動的なリスニングと異なり、シャドーイングは脳と口の筋肉が同時にリアルタイムで英語を処理・再現することを強制します。研究により、発音精度、抑揚、リズム、連音、リスニング力、そして会話の流暢さが大幅に向上することが確認されています。IELTSスピーキング対策や自然な英語コミュニケーションを目指す方に特におすすめです。

シャドーイングのやり方: ステップ別の完全ガイドを読む →

シャドーイング練習: AI Agent Memory Architecture: Structured SQL, Semantic Vector Search & Memory Bank - 動画で英語スピーキングを学ぶ

練習モードを選ぶ

アプリをダウンロード

同じチャンネル

The No-Code Roadmap: How I Approach Prototyping

What's new in Google Cloud in less than 1 minute

AI harness crash course

MCP vs API: The protocol every developer needs to know

おすすめ動画

Remembering and Forgetting: Crash Course Psychology #14

Agentic AI Frameworks Explained: Workflows, Multi-Agent, & Production

They JUST Accidentally Revealed the Truth: The AI Bubble

Configure storage accounts | AZ-104 | Episode 16

Jack Dorsey’s New Dessa AI Voice Sounds Scary Human

[TALK 13] Structure Prediction and Design using AlphaFold – Sami Chaaban

Master AI in 60 Minutes with Vector Databases!

3-D Brain Scans | Pathways 2

When AIs act emotional

Andrew Ng on Coding Agents and Deciding What to Build

Your AI Agent Starts Every Session Like a New Hire - Killian Carlsen-Phelan, Sonar

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

このレッスンの語彙とスピーキングのポイント

この動画の重要語彙

動画に出てくる句動詞

繰り返し練習したい文

注意したい発音

この動画での練習方法

シャドーイングとは？英語上達に効果的な理由