After Week 2 you understand how LLMs work. Now the honest problem: a bare LLM is severely limited. Without tools and memory it cannot do most of the things that make agents genuinely useful.
📅
Frozen in time
Knowledge ends at training cutoff. Ask it about yesterday's news and it either says it does not know — or makes something up.
🔢
Cannot compute reliably
LLMs are unreliable at math. Arithmetic is not pattern matching. It will often give a wrong answer stated with full confidence.
💾
Forgets everything
Every conversation starts fresh. Close the chat, open a new one — no memory of who you are or what you decided before.
This is why tools exist. Each tool solves exactly one of these limitations. Together they turn a limited text predictor into a genuinely capable agent.
Part 2
Tool calling: how agents use the outside world
Tool calling is the ability of an agent to invoke an external function and use the result in its response. When an agent searches the web, it is calling a search tool. When it runs code, it is calling a code interpreter.
1
Agent decides it needs a tool
During THINK: "I need current information. I should use the search tool."
2
Agent calls the tool with parameters
It generates: search("KEDA autoscaling best practices 2025")
3
Tool returns a result
The search API returns 10 results with titles, URLs, and snippets.
4
Agent uses the result to continue reasoning
The result is injected back into the context. Now it reasons with real current information.
Common agent tools
Web search — real-time info · Calculator — reliable math · Code interpreter — run and verify code · Database — query structured data · Browser — navigate websites
Part 3
Memory: short-term, long-term, and RAG
Agents have two kinds of memory. Understanding the difference is critical for building agents that work across multiple conversations.
Short-Term (Context Window)
Long-Term (External Storage)
Everything the agent can see right now
Stored in a database outside the LLM
Lost when the conversation ends
Persists across conversations
Limited to ~100k-200k tokens
Effectively unlimited
Always available, no retrieval needed
Must be retrieved when relevant
The open-book test analogy for RAG
A bare LLM is like a closed-book test — it can only use what it memorized during training. RAG (Retrieval-Augmented Generation) is like an open-book test — you look up the right pages before answering. The LLM still does the reasoning, but now with access to real specific information.
✏ Activity A — See RAG Work (no code)
Ground an AI in a real document — and watch it retrieve
RAG is easier to see than to explain. Use a free no-code RAG tool to feed in a document and watch the AI answer from that source, with citations. Pick one tool: NotebookLM (notebooklm.google.com) — answers cite the exact source passage — or Perplexity (perplexity.ai) — shows the pages it retrieved.
1
Add one shared document
Use a public, non-identifying document everyone has — the school handbook, a class syllabus, or a public article. (See the access note below before using student data.)
2
Ask 3 questions it answers + 1 it can’t
e.g. “What’s the attendance policy?” Then ask one question the document does not cover.
3
Watch where answers come from
For answerable questions, follow the citation back to the source. For the unanswerable one: did it say “not in the source,” or invent something?
🌐 Access note
No logins or tools allowed at your school? Two fallbacks: (1) Teacher-projected mode — the teacher runs one document on the projector and the class directs it together. (2) Paper RAG — groups get a one-page source and may only answer using highlighted lines, citing each one; a question the page can’t answer must be marked “not in the source.” That is RAG, by hand.
✏ Activity B — Memory Bot Experiment (optional)
Test what an AI actually remembers
Run this real experiment using Claude.ai or ChatGPT. Follow the four steps and document what you find.
1
Share 4 personal facts
Tell the AI your name, favourite subject, a hobby, and something specific you are working on this week.
2
Ask it what it knows about you
Type: "What do you know about me so far?" Check if it recalls all 4 facts correctly.
3
Close and open a brand new conversation
Do not continue the same chat. Start completely fresh.
4
Ask what it remembers about you
Type: "Do you remember anything about me?" Document what happens.
Go Deeper — Optional
Week 3 Python Notebook
Run in Google Colab — no setup, no install. Interactive Python exercises that reinforce every concept from this lesson.