M9-B — Memory, Planning, and Execution

How to read this note

An agent without memory forgets. Without planning it jumps randomly. Without execution it only talks. This note explains all three — simply.

🧠 Part 1 — Memory

Memory = what the agent carries from earlier steps.

Three memory types

Type	Lasts how long	Example
Short-term	Current task / chat	"User asked for Mumbai weather"
Working	Current multi-step job	"I already searched — now summarising"
Long-term	Across days/sessions	User preferences in a `.md` file

Open notebook analogy

Short-term  = what you hold in your head during one conversation
Working     = sticky notes on desk for today's project
Long-term   = diary on shelf you open tomorrow

Tools like OpenClaw use markdown files (MEMORY.md, SOUL.md) as long-term memory — the agent reads them before every reply.

User message arrives
      ↓
Agent loads: MEMORY.md + recent chat + task file
      ↓
Bigger prompt → smarter continuity

For students

Your Obsidian vault is agent memory if you design it that way — structured notes the AI reads first.

🗺️ Part 2 — Planning

Planning = turning a big goal into ordered steps before or while acting.

Without planning

Goal: "Prepare me for Module 7 exam"
Agent: writes random paragraph about AI history ❌

With planning

Goal: "Prepare me for Module 7 exam"

Plan:
  Step 1 — List all Module 7 topics from uploaded syllabus
  Step 2 — For each topic, generate 3 MCQs
  Step 3 — Ask student to answer
  Step 4 — Explain wrong answers using notes

Two planning styles

ReAct style	Plan-first style
Think → act → think → act	Write full plan → execute step by step
Good for exploration	Good for predictable workflows
Like fixing bike while riding	Like following recipe card

Backend (simplified):

LLM output: "Thought: I need syllabus first"
LLM output: "Action: read_file(syllabus.pdf)"
Tool returns: [file text]
LLM output: "Thought: now I can list topics…"

⚡ Part 3 — Execution

Execution = actually calling tools and using results — not just describing what you would do.

PLAN says: "Search web for today's news"
EXECUTION: HTTP request → search API → JSON results back
PLAN says: "Save to notes.md"
EXECUTION: write_file("notes.md", content)

Talk only	Execute
"I would search Google"	Search runs, results appear
"You should save this"	File saved on disk
Stops at advice	Changes something in the world

The real insight

Execution is what makes agents dangerous and useful. Dangerous if unchecked. Useful if tools + limits are designed well.

🔗 How memory + plan + execution connect

┌──────────┐
│  MEMORY  │──┐
└──────────┘  │
              ▼
         ┌─────────┐      ┌───────────┐
 GOAL →  │  PLAN   │ ──→  │ EXECUTE   │ ──→ tool results
         └─────────┘      └───────────┘
              ↑                    │
              └──── update memory ─┘

Each tool result updates memory → next plan step gets smarter.

🛑 Safety rails (must teach)

Rail	Why
Max steps (example: 10)	Stops infinite loops
Allowed tools list	No random email sending
Human approval gate	"Draft ready — send? Y/N"
Log every action	Debug when agent goes wrong

From Module 6: human-in-the-loop still applies to agents.

✅ Mini exercise

Goal: "Organise my desktop study PDFs list"

Write on paper:

What memory does agent need?
What 3-step plan?
What tools (file list, rename, move)?
Where does human approve?

Previous: M9-A — What Are AI Agents
Next: M9-C — Autonomous Workflows
Bonus read: Open claw — real system using markdown memory files