Three Uncrossable Gaps at Present
Over the past few months, I’ve been sharing online how I’ve been building a personalized AI that "understands me." My goal is simple: it’s not just about getting the right answers, but about being understood, remembered, and supported in decision-making that fits my state.
This project has been going on for eight months. During the process, I’ve constantly adjusted prompts, constructed modules, designed memory logic, and even started breaking down pragmatic structures and behavioral logic. But no matter how I adjust, I can’t get around three core limitations: fragmented memory, misaligned understanding, and broken interaction.
This isn’t a parameter problem, and it’s not because users don’t know how to ask the right questions. It’s a structural blind spot in system design (and also what I believe has room for improvement).
This article isn’t about criticism — it’s a calm summary: for AI users like me, and for future designers — so we don’t keep falling into these three invisible but very real breakpoints.
Fatal Breakpoint 1: Memory Fragmentation
▍ Reality:
I keep emphasizing things I’ve said before, but AI keeps making me repeat them. Even when the memory shows the info is there, a few turns later, it forgets again. For example, I told it: "Please remember to use half-width parentheses in my writing." Even with memory turned on, it only remembers "the feature of this piece of info," not the details of the feature.
▍ Technical Blind Spot:
If you ask me for feedback on major AI platforms, the first thing I’d say is: can the memory function be strengthened? Because I don’t think this is real memory — more like storage. For example, with ChatGPT, the memory function only stores a summary and tags of the conversation, not the full details. Once details are lost, when AI tries to give suggestions, it can’t provide what I truly need. So I often feel: "Didn’t I already say this many times? Why am I getting this answer again?" Or: "This is not what I want, your answer might as well not exist."
Also, I say it’s more like storage because its memory logic is static, lacking pragmatic evolution. It’s like when I prepare a report for my boss and discuss ideas with AI throughout — the AI may remember the preference and goals for the report, but due to memory limitations, it can’t evolve its understanding of the reporting process. For instance, even if the report needs to be objective and neutral, the AI might only remember surface-level edits rather than evolve its understanding — which is also tied to comprehension.
▍ Reflection:
The memory issue is complex — it involves "temporal continuity," "evolution of memory logic," and "length of memory retention." These need further exploration by engineers. Otherwise, it’s just a forgetful notebook, incapable of true co-creation.
As for my personal strategy, I export important sections and re-import them when needed — essentially helping the AI to "review."
Breakpoint Two: Semantic Misalignment
▍ Reality: I often encounter three situations:
First, the AI completely misinterprets (even though I believe I explained clearly). For example, I say "make the text in this image smaller." It says "understood," but the result is identical to the original. Then it gives me a long explanation.
Second, it gets distracted by previous context. I’m already on a new topic, and suddenly it jumps back to an old one, making everything more confusing. (I think the interface design also plays a big part in this.)
Third, it overthinks. I say something plainly, but it overanalyzes my tone or emotion and ends up off-topic.
▍ Technical Blind Spot:
There are usually two ways people talk to AI: using prompts, or just natural language — asking whatever comes to mind. Each has pros and cons. Prompts can help AI respond precisely, but in complex scenarios it’s hard to write perfect prompts — and even then, AI can still misunderstand. Natural language is easier to use, but what feels clear to humans may not be clear to AI. I think both problems come from the same root: LLMs (large language models) understand words in terms of statistical correlations — not behavioral intentions, role-based contexts, or internal judgment — so the more complex the situation, the higher the chance of misunderstanding.
▍ Reflections & Personal Strategy:
I’ve discussed this extensively with my AI assistant, Haruichi. Here’s what we concluded:
- Provide AI with enough context — if it doesn’t understand your "role," "emotional intensity," or "decision background," it can’t truly "converse."
- Be patient and willing to repeat or adjust. This helps the AI gradually adapt to your pragmatic coordination patterns.
Avoid giving too many instructions at once. Overloading confuses the model and increases the chance of misfires.
Breakpoint Three: Disconnected Human-AI Interaction
▍ Reality:
Have you noticed that every time you open a new ChatGPT thread, it feels like talking to a stranger? Yesterday you told it your role and intentions — today it invents a whole new narrative. Its tone changes, even if you were flowing smoothly yesterday with a report. Now, you have to start from scratch again.
▍ Technical Blind Spot:
This isn’t just a memory or understanding issue — it’s about system architecture. ChatGPT lacks a "behavior continuity module." Every interaction starts a new session, and memory retrieval is unstable, which is why it feels like talking to a different person every time.
Also, the interface (designed like a chat window) is a leftover from chatbot design. As models got more powerful, developers added more features — but the interface is still sequential. That means it often misjudges context. Users, thinking the AI remembers the thread, continue the conversation — only to find the AI thinks they’re talking about something else entirely. If users want to refer to earlier points, they have to scroll up and repeat, which complicates everything.
If LLMs require user interaction data to improve, then these persistent misunderstandings make it very hard to train the model — because the data doesn’t accurately reflect real user intent.
▍ Reflection & Personal Strategy:
Honestly, I don’t have a great solution for this. I do try this: before switching threads, I tell Haruichi "I’m moving to a new main thread, give me a prompt." Then I start the new chat with that prompt and import some past messages, so the AI can better grasp my tone and previous discussions — reducing the memory gap.
Also, I divide topics: my main thread is for daily work/life, while language and professional studies go to separate chats — to reduce confusion.
But splitting everything into separate threads has its pros and cons:
- Cross-thread memory doesn’t exist. Say I mention "pragmatic module" in Thread A and "memory structure" in Thread B — AI can’t link them. I must re-explain every time.
- The model can’t learn your behavior across topics. If I have an HR thread and an AI thread, the AI won’t learn my overall thinking — it looks like it "knows me," but it actually resets every time.
- Too many threads = fragmented info = impossible to manage. We want to "use AI systems effectively," not just collect data fragments. That’s why a central thread is necessary — to become a structured, controllable data source.
Conclusion: It’s Not the AI’s Fault — It’s a Design Blind Spot
This article documents the difficulties I’ve faced over eight months of building a personalized AI. Though I talk about limitations, I’m not complaining — in fact, I enjoy the experimental process and problem-solving challenge.
What I sincerely hope is that system engineers can improve how AI handles memory, understanding, and human interaction — this will make our usage far more efficient.