Core argument
Conversation is chronological, but inquiry is relational. A useful AI workspace should preserve both without asking people to manage a graph by hand. Cards can give local questions a place to live, a graph can preserve direction and provenance, and a continuously compiled article can turn the accumulated research into something coherent enough to read.
A research conversation often leaves its original path for a good reason. An unfamiliar term deserves a question. That answer reveals another connection. By the time the detour is resolved, the first line of inquiry is still somewhere above, buried in the transcript.
Nothing is wrong with conversation being chronological. The problem is asking one timeline to serve as transcript, research structure, and finished document at the same time. It records what happened next. It does not necessarily preserve how the ideas relate.
These are three different structures:
- Conversation happens in time and maintains immediate attention.
- Cards and a graph persist relationships, branches, paths, and sources.
- An article takes an editorial form and becomes the version intended for reading.
The timeline is not the enemy
Linear chat has real advantages. It is immediately legible, keeps two participants focused on the same moment, and gives each response an obvious local context. For a short question, a chronological thread is usually the right interface.
The trouble begins when inquiry branches. A reader pauses on a company, a technical term, or a background event, follows that question for a while, and then wants to return. Most chat products insert the detour permanently into the main transcript. The answer is useful, but the original line of inquiry loses its place.
Search can recover a sentence. Scrolling can recover a time. Neither reliably recovers the role that a question played in the investigation. As a session grows, chronology becomes a weaker proxy for meaning. The interface has confused the order in which ideas appeared with the structure they eventually formed.
Chronology is a good record of interaction. It is not a complete model of knowledge.
Branching should be a gesture, not a management task
A branch can begin with an ordinary action. The model marks a phrase that can be explored, and the user opens it. Or the user selects a passage and asks a question. That gesture already says enough: continue from here. The product can record a new node and its source without asking the user to create a folder, name a branch, or choose a parent.
The same principle applies when the detour ends. People should not have to merge context by hand. The system can decide which material belongs in the next answer and which evidence should inform the reading view. Users need clear ways to return, close, revisit, and undo. They do not need the vocabulary of version control in the middle of a conversation.
Automation still needs a stable substrate. User actions and source relationships should be appended, not silently rewritten. A model may compile a new article or select a relevant context window, while the original exchanges remain available for inspection. The graph is trustworthy only when its edges can be traced back to something that actually happened.
A Card gives local attention an address
Making every message a graph node would produce structure faster than understanding. A more useful unit is a bounded local exchange: one source question, its answer, a few follow-ups, and the anchors that can lead elsewhere. In the Lattice prototype, that unit is a Card.
The Card should still read like chat. It does not need a summary header, a temporary conclusion, or a strip of suggested reading. A short title can be generated once when the branch is created and stored as navigation metadata. The conversation itself should remain natural.
Cards also make a path spatial. A new Card sits above the context from which it opened. Closing it returns to the sheet underneath. Spreading the Deck exposes earlier positions in the current line of inquiry. The paper quality is not the point. The useful part is giving return, depth, and current path a consistent physical expression.
The Card matters because a local question becomes somewhere you can return to.
The graph is for orientation and provenance
Once an investigation has several branches, a global view becomes valuable. Where did this idea come from? Which paths remain unexplored? Did two independent lines of inquiry converge on the same event? A graph can answer these questions more faithfully than a transcript.
Convergence is one reason the structure is better described as a directed acyclic graph, or DAG, than a tree. Two branches may reach the same crisis, company, or interpretation. They should point to one shared node instead of duplicating the material and allowing two versions to drift apart.
Graphs also fail quickly when asked to do too much. Permanent labels compete with edges. Automatic regions invent a taxonomy that the conversation never produced. Dense networks can offer the feeling of total knowledge while making every individual relationship harder to read. The graph should stay secondary, keep its geometry stable, and reveal labels or connections only when they help with orientation.
Complete prose belongs in Cards and in the reading view. The graph is a map of relation and provenance, not another place to read the entire argument.
Reading needs a different projection
Rejecting a single chat timeline does not mean rejecting linear reading. The path taken during research and the order needed by a reader are different things. Inquiry may jump from childhood to corporate control and back to technical strategy. A good article must reorganize that material by chronology, theme, or argument.
The final reading view should therefore look closer to a continuous reference article than to a printed graph. A section may synthesize several Cards. One Card does not need to become one heading. As research grows, the system can recompile the current edition, but the reader should always meet a complete document rather than internal labels such as unfinished, processing, or provisional conclusion.
That article is not the new source of truth. Each passage should retain links to the Cards that informed it, allowing a reader to return to the original question and its context. The article is an editorial projection. The graph is a relational projection. The Cards preserve the conversation. Their value comes from coexisting.
A nonlinear process can still produce a deliberately linear reading experience.
From session history to durable knowledge
Chat products preserve process but often leave little that can be reused. Note-taking products preserve outcomes but usually ask people to perform a second job after the real work is done. A better system could let ordinary conversation accumulate into three connected representations: the original actions and exchanges, local Cards and an exploration graph, and a continuously compiled article.
The durable layer might be local Markdown, bidirectional links, or another portable file format. Different agent harnesses could operate on the same material. But the filesystem should be the substrate, not the interface. People should explore and read in a product designed for those activities, while retaining ownership of files that can outlive any one model or vendor.
This changes what automatic organization means. The goal is not to split every conversation into a larger pile of notes. It is to preserve the process, the relationships, and a coherent result without requiring the user to rebuild those connections after the session ends.
A useful hypothesis, with limits
Not every chat needs a graph. A simple question should remain simple. Structure should grow only when a person opens an anchor, selects a passage, revisits a historical Card, or otherwise makes a relationship meaningful through action. A graph generated in advance for visual effect is just another form of clutter.
Automatic context management can also be wrong. Models may infer a false relationship, produce a vague node title, or compress away an important qualification while rewriting the article. Provenance, undo, stable history, and a visible account of current context are product requirements, not implementation details.
This leaves a set of hypotheses to test. Do people return to the original question more reliably? Does a Deck feel like a path rather than an animation? At what scale does the graph stop helping? Does the compiled article reduce real organization work? Visual polish matters, but these are the questions that determine whether the interface earns its complexity.
Let conversation leave something worth returning to
A better interface does not ask people to think in graphs. They should still be able to ask, follow up, wander, and return without maintaining a structure by hand.
The conversation can preserve what happened. Cards and the graph can preserve how the inquiry unfolded. The article can turn the accumulated material into something coherent enough to read and reuse. None of these views needs to replace the others.
The useful promise of AI knowledge tools may be less about generating more notes, and more about letting an ordinary exchange leave behind a structure and a result worth continuing.
This essay grew out of the interaction design work behind Lattice.