DEV Community

Sarvesh Gajakosh
Sarvesh Gajakosh

Posted on

Why My Coding Agent Stopped Repeating the Same Hints

#agents #ai #learning #showdev

Why My Coding Agent Stopped Repeating the Same Hints

While testing our coding mentor, I noticed something frustrating.

The AI wasn’t wrong.

It was just repetitive.

A user fails a problem.
The mentor says: “Try breaking it into smaller parts.”

They fail again.
Same hint.

Third time?
Still the same hint.

At that point, correctness doesn’t matter. The system isn’t helping anymore.

That’s when I realized the real problem wasn’t intelligence.

It was memory.

What We Built

CodeMentor AI is a coding practice platform with an AI mentor that adapts based on user behavior.

The interface is simple:

  • Monaco Editor in the center
  • Problem list on the left
  • AI mentor on the right

The goal was not to make a smarter model.

The goal was to make the system aware of the user.

** The Core Problem**

Most AI tools are stateless.

They don’t remember:

  • previous mistakes
  • repeated failures
  • learning patterns

So even if the response is correct, it doesn’t evolve.

Every session starts from zero.

The Approach

Instead of increasing model complexity, I introduced a simple memory loop:

retain → recall → reflect
Enter fullscreen mode Exit fullscreen mode

This became the foundation of the system.

retain() — Store Meaningful Events

After every submission:

await retain(
  hindsightKey,
  bankId,
  `Failed ${problem.title} due to incorrect loop boundary`,
  {
    type: "failure",
    topic: "arrays",
    difficulty: problem.difficulty,
    timestamp: new Date().toISOString()
  }
);
Enter fullscreen mode Exit fullscreen mode

Key decision:

Do not store raw code.
Store the reason behind failure.

This keeps memory compact and useful.

recall() — Fetch Relevant Context

Before solving a new problem:

const memories = await recall(
  hindsightKey,
  bankId,
  "user mistakes in array problems",
  5
);
Enter fullscreen mode Exit fullscreen mode

This retrieves only what is relevant:

  • similar past mistakes
  • related topics
  • repeated patterns

reflect() — Convert Memory into Insight

Before generating a response:

const reflection = await reflect(
  hindsightKey,
  bankId,
  "user weaknesses in problem solving"
);
Enter fullscreen mode Exit fullscreen mode

This produces insights such as:

  • misses edge cases
  • struggles with loops
  • relies on brute force

What Changed

Before memory:

“Check your logic.”

After memory:

“You’ve missed boundary conditions in similar problems before. Check your loop limits.”

That shift makes the response specific and actionable.

What I Learned

Memory is more important than model size.

Better context leads to better responses.

Summarized insights are more useful than raw data.

Query design directly impacts output quality.

Testing across multiple sessions is necessary to see real value.

Current Behavior

The system now:

  • adjusts hints based on past attempts
  • avoids repeating the same feedback
  • builds a pattern of user weaknesses

It doesn’t feel like a stateless tool anymore.

It feels like a system that tracks progress.

Try It

https://codemento.netlify.app/

Final Thought

AI systems don’t improve just by being more powerful.

They improve by remembering.

Once memory is introduced, the system stops starting over.

And that’s when it becomes useful.

Built at HackWithIndia 2026- Sarvesh Gajakosh

Team: Krishna Hasare, Karan Patil, Priya Vhatkar, Siddhi Shinde, Aastha Sawade

Top comments (0)