My AI Mentor Remembered Every Bug I Kept Repeating — And Wouldn’t Let Me Ship With It Again.

thejus bs — Mon, 23 Mar 2026 17:31:32 +0000

“I think you’re doing that loop thing again.”

That wasn’t me. That was my app.

At some point, CodeMentor stopped reacting to my code and started recognizing how I write it. That shift—from response to pattern recognition—is what this project is really about.

What This System Actually Does

CodeMentor is a React app that combines:

Groq (LLaMA 3.3 70B) for fast feedback
Hindsight for long-term memory

Instead of treating every submission as new, it builds a memory of your behavior over time.

Each interaction follows:

Recall → Analyze → Retain
📸 Screenshot:
Memory as Context, Not Storage
Most people think memory means saving more data.
What matters is using it at the right moment.

Before every review, the system recalls past patterns:

const r = await hs.recall(bankId, query);
const mems = r.results || [];

That memory gets injected into the prompt.

Now the model doesn’t just see your code—it sees your history.

📸 Screenshot:

The Shift: From Feedback to Pattern Detection

After a few sessions, something interesting happens.

Instead of generic advice, the system starts saying things like:

“You’ve done this before”
“This looks similar to your last mistake”

That’s not intelligence—it’s memory applied correctly.

Closing the Loop

Every interaction becomes future context:

await hs.retain(bankId, Mistakes: ${mistakes.join(", ")});
That’s the entire loop.
Simple—but powerful.

📸 Screenshot:

Why This Matters

Without memory:

You improve randomly

With memory:

You improve systematically
Final Thought
This isn’t about smarter models.
It’s about giving models memory of you.

Forem: thejus bs

My AI Mentor Remembered Every Bug I Kept Repeating — And Wouldn’t Let Me Ship With It Again.