This article comes from a conversation with GPT, but it is not a random flash of inspiration. More accurately, it is a "merge commit" completed in the background of my brain after long-term reading, learning, and repeated chewing. Those seemingly unrelated topics become implicit information for me to pose new questions at this moment.
1. The Beginning of the Question: Linear Chat Cannot Accommodate Divergent Thinking#
Last night, while chatting with AI, I suddenly thought of a topic: commonly used AI tools are based on a linear conversational approach. However, human thinking is jumpy, so when my thoughts and the topic I want to discuss change, should I start a new conversation or continue chatting in this window? Because I might jump out and then return to the original topic. I want to discuss new content, but I am also worried that the new content will "pollute" the context of the current window. It’s somewhat like git branches, conflicts, and merges, and a bit like writing thoughts on a whiteboard from top to bottom, suddenly leading to a line going elsewhere, starting to think about new questions, perhaps a pending issue in my mind; or a topic similar/opposite to the current one; or even completely unrelated new content. As the conversation progresses, my thoughts might shift to another new idea or return to the original topic.
However, current AI tools seem unable to accommodate this kind of thought leap. I wonder what the difference is between humans and AI?
I discussed this question with GPT.
AI: The conflict you feel is structural: human thinking resembles a DAG (Directed Acyclic Graph), while the chat window is designed as a log (linear log).
From a practical standpoint, if the topic has truly changed (the cognitive context is different), you can open a new window; if it’s just a diversion but you want to return, stay in the same window, but explicitly mark the branch (for example, "let's open a branch: A, and come back to main later") to keep AI in sync.
But this is just a stopgap. What you really want is an interactive method that can branch, backtrack, and merge within the same cognitive space.
Directed Acyclic Graph—a very apt description. So what drives us to jump from one node to another? Meaning (I think). I recall an article I read a few days ago that recorded a story: a person who underwent partial brain tissue removal surgery fell into "decision paralysis" because the removed part controlled the connection between emotions and choices. If any choice is devoid of "bias," it means that everything loses its meaning, and ultimately, his life became a mess.
Time flows linearly, but tasks and thoughts are not linear—they can branch, backtrack, and merge. Here I recall some thoughts I previously wrote about task management. Tasks can be divided into calendar items with time attributes and to-do tasks without time attributes. Different task events are not necessarily processed sequentially by time. Just like thinking, it can branch, backtrack, and merge.
We can preliminarily draw some conclusions:
- Linear chat ≠ true human thinking
- Thinking is inherently branching, backtracking, and reconstructing
- Collaboration between humans and AI should not be limited to "temporal order"
GPT told me what I need is:
A shared space that can converse, serve as a cognitive map, and allow AI to understand structure
What I said: What I need is a "shadow" that can perceive everything about me.
2. What Does This Have to Do with "Context Engineering/Agents/Memories"?#
After seeing AI's response to my first question, I suddenly thought about how my thinking relates to this year's popular "context engineering," "agents," and "large model memory functions." Are current agents addressing these issues? What problems does the memory function solve?
GPT told me:
Context is one-dimensional, but cognition is multi-dimensional.
Context Engineering: How to retain the most important content within a limited conversation window.
Human thought can suddenly jump, but still wants to return to the original theme, and hopes AI can remember the "structure," which is a problem of context selection and organization. Large models compress high-dimensional thinking into "one-dimensional," like the "paper tape" originally used by computers. They express "polygons" or even "three-dimensional figures" on a single line.
Agents: Adding limbs to the large model brain, so it is no longer a "brain in a vat."
Agents also conceive different ideas and plans when handling tasks, but do not maintain all thought branches.
Memory: Context is ultimately limited; where do we put content that cannot fit? Which content will we need later?
Memory records content, not the live scene of thought.
These seem to not answer: how to truly model "the structure of thought itself."
The next stage of AI collaboration, should the interaction with AI be closer to human cognitive structure?
How to preserve the shape of thought?
3. What Should Ideal Context and Memory Look Like?#
The question GPT left me at the end of the last topic was: "If we really want to create the tool you mentioned, what should its context and memory look like?"
I said: What do you think its context and memory should look like? Do you see any connection with Luhmann's card note method? He used coding to create a network of thought.
While chatting, I suddenly thought of "Luhmann" and his creation of "card notes." Luhmann provided a scaffold for the brain, allowing thoughts to remain growable even after leaving the brain.
I think context should be like "card notes," possessing atomicity, and linking to other "contexts" through "coding." It is a series of "thought maps" that can reference, link, and combine with each other, forming a "network of thought."
When people think, it might be like this: "I am thinking of this now because I just thought of xx. It may also relate to xx that I encountered a long time ago." These thoughts have directional significance cognitively. This is not something that vector databases, memories, or chat records can solve.
Coding = an infinitely expandable semantic coordinate system. It seems to be what I want: "branch anytime, return anytime." So what role can AI play?
Luhmann's system is dynamic human brain + static cards. Humans are responsible for building the "card network."
With the emergence of AI, this system has become dynamic human brain + dynamic second brain + dynamic cards. This network no longer has a single maintainer. Thoughts can grow together with AI.
4. Differences in "Thinking Mechanisms" Between Humans and AI#
At this point, although I have been learning about large model training recently, I am still curious about the differences in thinking between humans and AI.
Human vs AI: The difference in thinking is not in "smartness," but in "mechanism."
Let me paraphrase GPT's response in my own understanding.
Human: Meaning-driven wandering
- A thought carries a wealth of implicit background information
- Unresolved issues linger in the "background"
- Able to return to the "scene of thought," not just looking for similarities
Luhmann's card notes preserve the "scene of thought."
AI: Probability-driven activation
- Context is one-time
- Associations rely on similarity rather than "structure"
- Does not remember unfinished tasks unless prompted
So when the "second thinker" is no longer paper, but AI, what changes will occur in the structure of thought? Can AI serve human meaning leaps?
To be continued!
5. As a Side Note#
This thought did not come suddenly; it is also catalyzed by the content I have been reading recently, from Susan's "Entropy Control Theory," to the Notion CEO's "Looking at the World Through the Rearview Mirror," and to the "Luhmann Card Note Method" and "Time-Task Processing Model" that I learned about years ago. There is also the current popular "context engineering," "memory," and "RAG" in the AI field (interestingly, this text was not written in order, but was the third segment I wrote).