What AI Agents Still Lack: A Personal Memory and Decision Layer

ChatGPT Feature requests
1

As AI assistants evolve from answering questions into agents that can actually take actions, I think we are approaching a different kind of problem.

The question is no longer just:

Can the AI do this task?

It is increasingly:

Should the AI be allowed to make this decision for me — and what should it remember afterward?

I think personal AI agents may need a missing infrastructure layer:

A User-Controlled Personal Agent Layer

A layer that sits between the user, the AI agent, and external tools, and manages four things:

Memory. Permissions. Delegation. Feedback.

1. Memory should become decision memory, not just conversation memory

Imagine I tell my AI:

“This drink tastes terrible. The artificial flavor is too strong. Don’t buy it again.”

I don’t want to write a prompt, fill out a preference form, or manually maintain a profile.

I just want the AI to understand the feedback naturally.

More importantly, it shouldn’t only remember:

User dislikes Product X.

It should gradually understand things like:

  • I dislike strong artificial flavors.

  • This preference may apply to similar products.

  • I am still willing to try new drinks.

  • One bad experience should not automatically ban an entire category.

The value of long-term memory is not remembering more conversations.

It is making better future decisions.

2. Users should be able to delegate decisions within explicit boundaries

For example, I might tell an AI:

“You can spend up to $50 per month trying new snacks and drinks for me.”

Within that budget, I might be comfortable letting the agent choose and purchase things without asking every time.

But I would not give it the same authority to buy:

  • a laptop;

  • expensive furniture;

  • medication;

  • something difficult to return.

So agent autonomy should not be binary.

It could depend on:

  • category;

  • amount;

  • frequency;

  • reversibility;

  • risk.

3. Agents should never expand their own long-term authority

This feels especially important.

An agent may say:

“You’ve accepted my last 20 recommendations. Would you like to increase my monthly autonomous purchase limit?”

That’s reasonable.

But the agent should not decide:

“I’ve been performing well, so I increased my own limit.”

The agent can improve its decisions.

It should not redefine the boundaries of its own authority.

Permission expansion should always remain under explicit user control.

4. We may need to optimize for affordable mistakes, not perfect decisions

This came from thinking about shopping agents, but I think it generalizes.

Sometimes users don’t need an AI to be correct 100% of the time.

For low-risk decisions, they may prefer:

less friction + controlled mistakes

over:

constant confirmation + theoretical safety

For example, if I explicitly allocate a small monthly budget for trying new products, a bad recommendation isn’t necessarily a system failure.

It may simply be part of exploration.

What matters is:

  • the loss was bounded;

  • the agent learned from it;

  • the same mistake does not keep repeating;

  • I remain in control of the boundary.

5. The best agent may be the one that knows when not to bother the user

A lot of current agent UX still looks like:

  1. AI generates options.

  2. AI explains every option.

  3. User reviews everything.

  4. User confirms.

  5. AI finally acts.

That’s useful, but in many low-risk situations it still leaves most of the cognitive work to the user.

A mature personal agent should probably learn another skill:

Which decisions are not worth interrupting the user for?

That may eventually be as important as reasoning quality itself.

The broader idea

I started thinking about this in an e-commerce context, but I don’t think it is fundamentally a commerce problem.

The same architecture could apply to:

  • scheduling;

  • travel;

  • personal productivity;

  • subscriptions;

  • recurring purchases;

  • communications;

  • many other delegated actions.

As AI systems gain more real-world agency, I suspect we will need something resembling a personal constitution for agents:

What do you know about me?
What are you allowed to do?
When must you ask me?
What can you learn automatically?
What authority can never change without my approval?

I would be very interested in how others think about this.

As agents become more capable, is user-controlled memory and permission infrastructure becoming as important as model intelligence itself?

—
He
Nankai University
AI Engineer in E-commerce

1 Like