PAHF: Personalized AI Feedback

I kept re-injecting the same feedback to my AI. Every single session.

Same tone correction. Same structure note. Same "stop explaining what you're about to do." I write a lot — creative and technical — collaboratively with AI. And every new session, I was starting from scratch.

I assumed this was a memory problem. It wasn't.

Adding semantic memory helped partially. The agent could retrieve that I preferred concise writing. But it still guessed wrong constantly — and worse, it stored what it inferred from my corrections, not what I actually approved. So the memory got polluted fast.

The Cupcake Problem

Think of it this way.

You ask an AI for a cake. It makes a cupcake. You correct it — birthday cake. It updates its memory. But it already spent tokens on the wrong output, and it stored "birthday cake" from an inferred correction, not a confirmed one.

Next time you ask for something sweet? Confident wrong answer again.

What Actually Fixed It

The fix I built: the agent asks before it acts.

"I'm about to make you a cupcake — is that right?"

You say no. It locks in birthday cake. Then it acts. Memory only updates after you confirm the final output. And if you later want a wedding cake instead, you update that one preference — no full re-teach, no session replay, no wasted tokens.

The Pattern Has a Name

Turns out Meta AI named this exact pattern in early 2026: PAHF — Personalized Agents from Human Feedback.

The core idea: a three-step inference-time loop.

1. Pre-action clarification — eliminate assumption errors before generating
2. Grounding — retrieve approved preferences before generating
3. Post-action feedback — only write back what you confirmed

Semantic memory learns from what the agent did.
PAHF learns from what you approved.

That distinction is the whole thing.

I built this as a sidecar service — any frontend can plug in.

What feedback do you find yourself re-injecting every session?