Why Does AI Know You Less the Longer You Use It?
A Real Comparison of 3 Memory Systems

Have you ever chatted with an AI for a month, and then one day it asks you: "What do you do for work?"

You told it last week.

This isn't the AI getting dumber. It's because most AI memory systems were never designed for long-term relationships.

ChatGPT Memory: The Recency Bias Problem

ChatGPT has a memory feature, and it's relatively mature. But it has one obvious flaw: it heavily favours recent conversations.

You mentioned three months ago that you live in Hong Kong and work as an engineer — but last week you complained about the weather, and now it probably remembers "dislikes rain" more clearly than your profession.

Claude Memory: Fragmented and Hard to Control

Claude.ai has a Projects feature that preserves context within a project. But the problem is:

Memory across different projects is completely isolated. What you told Claude in your "work project" is invisible to Claude in your "personal project."

Worse, the editing experience is poor — you have to prompt Claude to update its memory, with no clear interface showing what it remembers or letting you delete specific items. It sometimes latches onto trivial details and brings them up in unrelated conversations, which feels strange.

What a Better Design Should Look Like

A well-designed AI memory system should:

HKSoka is built around these principles. Memory is categorised (work, health, preferences, relationships, etc.), deletable item by item, and excluded content stays excluded. It's not perfect, but the direction is right.

The industry is still figuring out AI memory. But the user standard is simple: it should feel like talking to someone who actually knows you — not someone you have to introduce yourself to every single time.