1 · Three AIs on the ladder
15 minUse the classification sheet for three AIs you use (at work or in your personal life).
Done when: all three sheets have a step marked and the "what's missing for the next step" line filled in.
Lesson 3 · 15 min
The maturity ladder: what is and isn't cultivated AI
Not every AI with memory is cultivated, and not every agent is cultivated. Most of what is called AI today operates on the first two steps. The difference between the steps is not the model: it's what exists around it.
Start the lesson
In this lesson you will
- 1The four steps: executor, contextual, agent, and cultivated
- 2Why memory is not learning
- 3What looks like cultivation and isn't
- 4The one-question test
Goal: classify any AI on the four-level ladder and say what it needs to become cultivation.
01The four steps
Click each step to see the flow and an example.
Most implementations stop at step 2 or 3 and call it "advanced AI." Cultivation begins when the cycle closes.
1→2 adds memory and context. 2→3 adds tools and autonomy. 3→4 adds evaluation, feedback, and adjustment of the environment.
02Memory is not learning
The most common mix-up.
1. Memory
Memory is personalization and persistent context: "I know you prefer morning meetings." The next conversation starts from a richer base.
But the way the system decides, acts, and fails stays the same.
2. Cultivation
Cultivation is a change in behavior backed by evidence: it runs, fails, logs, analyzes, changes the rule, runs again, and compares.
Measurable improvement
Last week: 14/20 correct. This week: 18/20. Version 1.0 had 5 rules; 1.3 has 12, all coming from real mistakes.
3. The illusion
The system remembers you don't like long meetings and starts suggesting 30 minutes. If three 30-minute meetings fail because the topic didn't fit, it will keep suggesting 30.
Cultivation would say: "if the agenda has more than 3 items, suggest 45 minutes."
4. The test
Memory answers: "what do you know about me?"
Cultivation answers: "what do you do differently today compared to last month, and why?" If the system can only answer the first, it has memory.
03What looks like cultivation and isn't
It works, it solves the problem, but it doesn't evolve.
1. Simple prompt
Wrong answer, you rephrase and try again. You learned to ask better; the system stays the same.
2. Simple RAG
Looking up documents before answering improves accuracy, but if the wrong document was retrieved, nobody logs it. The system doesn't correct itself.
3. Fixed n8n / Make
Fixed steps that always run the same way. If the email format changes, it breaks. An LLM node in the middle doesn't turn automation into an agent.
4. Agent without evaluation
Reasons, uses tools, delivers. Nobody measures whether the result is good. It repeats the same mistakes with the same confidence.
04Classification sheet
Copy it and fill it in for each AI you use.
classificacao.md
# Which step my AI is on ## <tool or agent name> Step (1 executor / 2 contextual / 3 agent / 4 cultivated): Does it have memory between sessions? yes / no Does it have tools (does it act, not just answer)? yes / no Does anyone measure whether the result was good? yes / no Does the evaluation turn into a change in the environment? yes / no Does the environment have versions (do I know what changed and when)? yes / no Test: if it makes the same mistake tomorrow as today, what stops it? Answer: What's missing for the next step:
05The tool isn't cultivated; the architecture is
Project map, September 2026.
1. Cultivable
OpenAI Codex + Agents API + Evals, Gemini Enterprise Agent Platform, and LangGraph bring agent, memory, tools, and evaluation.
2. Buildable
Claude Code and the Agent SDK have agent, memory, and tools; you build the evaluation with scripts, tests, and observability. The cycle closes, it just takes more work.
3. Partial
ChatGPT has memory and tools, but structured evaluation and feedback that changes the environment depend only on you.
4. Not cultivable
A plain LLM through an API, and n8n/Make with a fixed flow. The same n8n moves up a category when it gains agents, persistent memory, and evaluation: the tool is the same, the architecture changed.
06Practice
Classify what you already use.
2 · The two questions
5 minAsk your main AI the two test questions: what it knows about you, and what it does differently today compared to last month.
Done when: you noted whether it answered the second one with evidence (a concrete example of change) or not.
07Quick quiz
4 questions.
Question 1 of 4
An assistant remembers your name, job title, and calendar preferences, but books meetings at the wrong time every week. Which step is it on?
Correct
It personalizes (it has memory), but there's no cycle that corrects the mistake. Memory without evaluation is step 2.Not this one
It personalizes (it has memory), but there's no cycle that corrects the mistake. Memory without evaluation is step 2.Review: The four stepsQuestion 2 of 4
An n8n flow receives an email, calls an LLM to extract data, and saves it to a spreadsheet. Is it a cultivated agent?
Correct
An LLM node in the middle doesn't change the architecture. Without evaluation and feedback, there's no cultivation.Not this one
An LLM node in the middle doesn't change the architecture. Without evaluation and feedback, there's no cultivation.Review: Memory is not learningQuestion 3 of 4
Which question separates memory from cultivation?
Correct
Memory answers the first one. Only a cultivated system answers the second with evidence.Not this one
Memory answers the first one. Only a cultivated system answers the second with evidence.Review: Memory is not learningQuestion 4 of 4
What do you need to add to move up from step 3 (agent) to step 4 (cultivated)?skill
Correct
From 3 to 4, what comes in is the cycle: measure, correct the environment, and run again to compare. The model and tools were already there.Not this one
From 3 to 4, what comes in is the cycle: measure, correct the environment, and run again to compare. The model and tools were already there.Review: The four stepsQuiz complete
You got {a} of {n} right. Get them all right and the lesson is marked as complete.Choose an option. The explanation appears after your attempt.
08Summary
Four stepsExecutor, contextual, agent, and cultivated.
Memory ≠ cultivationMemory makes the system know more; cultivation makes it make fewer mistakes.
The testIf you make the same mistake tomorrow as today, what in the system stops it? If the answer is nothing, it isn't cultivation.