In recent years, more and more people have become accustomed to using AI (LLMs) to handle various problems, gradually getting used to treating AI as their "second brain." It can answer questions, write, code, analyze, and offer advice. Often, its performance is even more stable and efficient than that of humans. Most of the time, we think it is omnipotent.
However, after deep usage, some strange issues begin to emerge. For example, it can appear very intelligent in specific parts but lacks consistency overall. Sometimes it grasps the key points, while other times it gets stuck on details; sometimes its judgment is clear, and other times it seems to lose direction.
And, of course, there's the unavoidable fact that it often "fabricates" information.
These problems are inherent flaws. After all, for a probabilistic model, all its outputs are, philosophically speaking, fabricated—they are guesses. But if you are a deep, AI-first user, you'll realize that these issues arise not just because of its flaws. The more fundamental problem is that it is not a true "silicon-based brain" and lacks the ability to think.
LLM-type AIs excel at processing information, but thinking is more than just processing information. The foundational prerequisite for thinking is the ability to distinguish the "importance and urgency" of matters. When humans make judgments, there is always an implicit prioritization—a pre-processing of known information to determine which information to prioritize, which to ignore, and which to defer. This prioritization is implicit.
In AI, this layer is missing. That's why we see familiar phenomena: in long conversations, key premises gradually fade, irrelevant details are repeatedly amplified, and the focus of each response drifts. This isn't just a misunderstanding; it's more like the lack of a stable priority system supporting the output.
Text is a tool for transmitting information through imagination. Therefore, if the sender lacks the ability to describe facts accurately, and the receiver lacks imagination or has an overactive imagination, the entire information transmission process becomes a disaster.
LLMs are systems that rely solely on text for reconstruction. What they receive, from the very beginning, is compressed information. What they output is a high-probability restoration based on incomplete information.
In this process, humans use experience, feedback, and real-world constraints to refine their understanding. But AI lacks these; it can only continue making inferences within the textual space. So sometimes it appears very accurate, while other times it deviates completely. This is because it is always using limited information to imagine a world. Humans use imagination to compensate for the shortcomings of language, while AI has only imagination.
When OpenClaw became a big hit recently, I didn't jump on the bandwagon, even though I've been following it since its inception. I understand very well that its most striking feature lies in the human-computer interaction layer. A concrete character image and conversational interaction give OpenClaw the potential to serve ordinary people. However, at the same time, relying entirely on conversational text interaction amplifies the flaws of the underlying LLM.
Current Agents lack a reliable memory system.
Human judgment gradually converges because memory is constantly at work. The decisions you've made, the pitfalls you've encountered, and the preferences you've formed are repeatedly invoked in subsequent judgments and gradually reinforced, eventually forming a relatively stable cognitive structure. This structure isn't perfect, but it gives your judgment continuity.
Most current popular Agents rely on markdown or so-called RAG as memory systems, but their flaws are very obvious. I won't even bother discussing RAG. The biggest problem with markdown is that it is merely a storage format, not a cognitive structure. People can use it to record information, but it cannot express relationships. Moreover, it is static; once written, the content hardly changes. A judgment that has been overturned and a conclusion that remains valid are essentially indistinguishable in the system. It isn't corrected or phased out; it just keeps accumulating.
Such memory systems only allow AI to see more information but do not enable it to form genuine judgments. What truly determines the quality of judgment is never the amount of information but how information is filtered, organized, and forgotten.
Agents lacking a reliable memory system are merely more user-friendly information manipulation tools; they don't solve the problem of how to use tools effectively.
We can see that many researchers and AI-first professionals can use AI exceptionally well. They can provide AI with ample and rigorous contextual constraints, keeping the output results within a controllable range. This shows that, purely in terms of model capability, it already meets the needs of the vast majority of people. Therefore, what we should consider is how to make AI truly serve non-professional users.
Giving AI a "brain" isn't about stuffing it with more context; it should be about equipping it with a multi-dimensional, lifecycle-aware memory system.