Picking up a thought where you left off
Several ideas can click into a single shape, and that shape can fall apart quickly. Writing down the words afterward doesn't keep the connections between them. Drawing on about ten years of my own trial and error, this note asks what you need to keep in order to pick a thought back up.
- Knowledge
- Memory
- Interfaces
Sometimes several concepts, and the links between them, come together all at once into a single shape. That shape can fall apart quickly. A quick note keeps only the words — “A, B, C” — while how they were connected is lost. Reading the note later doesn’t bring the understanding back; the thinking has to be rebuilt almost from scratch.
This note collects about ten years of my own trial and error with this problem, and what it has shown so far. It is an observation based on one person’s experience, not something established as generally true.
Keep the relationships, not just the words
Frame this as a shortage of information or of memory, and the answer seems to be more notes. But what goes missing isn’t the words. It’s the state of the relationships the words formed at that moment.
Prose, hierarchical notes, diagrams, voice memos, and a Zettelkasten are all useful methods. In the trials so far, though, none was a good fit for keeping that state of relationships.
Nodes on a desk
Of the methods tried, the best fit for keeping relationships was a physical one. Concepts are written one per round label — a physical node — and the labels are moved around a desk while thinking. As of September 2026, about 200 nodes are in use.
Moving closer, moving apart, grouping, placing at the center, shifting to another area: distance, center and edge, and clusters all carry meaning. The arrangement isn’t a diagram that tidies up thinking already done. Rearranging is the thinking.
Something not yet understood starts as a single flat node. As it is researched and thought through, concepts stack up on it. The stack becomes physical evidence of what was once understood. Even when the details have been forgotten, laying the arrangement out again offers a way back into the earlier thinking.
The desk has limits, though: it runs out of room, and clearing it erases the arrangement. Those limits, and a prototype on the iPhone, are covered in the note “Thinking with physical nodes.”
What the system should do
The goal isn’t a system that only stores knowledge. It’s one that gets thoughts out of your head, keeps the relationships between them, returns to an earlier line of thinking, rebuilds the understanding, and carries it through into action.
Success isn’t measured by whether something was saved. It’s measured by whether the understanding comes back when you return to it, and whether you can move on to the next thought or action.
To do that, it needs to handle at least the following:
- Putting away, combining, laying out again. Keeping a thought that isn’t in use without letting it fall apart; combining elements into a usable structure; laying that structure out again later so the thinking can continue.
- What’s been put away, and what hasn’t been touched. A part that was thought through and is now resting, and a part that hasn’t yet been explored or built. They are two sides of the same thing, and are handled separately.
- A center to return to. Something that brings you back to what all those concepts were about in the first place.
- How ready a thought is to act on. Whether a structure of thought has matured far enough to act on in the real world. How to measure this hasn’t been settled.
Being able to return, and wanting to
There is another distinction worth keeping: between the drive to start something and the catalyst to come back to it after a break.
The drive to begin can be there while the catalyst is missing, so a thought that has gone cold never gets picked up again. Once thinking is moving, friction and open questions show up on their own. What’s missing is a smaller nudge that comes before any question forms.
Being able to restore an arrangement is not the same as wanting to engage with it again. So keeping, restoring, rekindling, and maturing toward action are treated here as separate questions. One hypothesis is that simply laying out nodes at random might be enough to get thinking moving again. It hasn’t been tested yet.
Voice and AI
Speaking instead of typing lowers the load of holding a thought while getting it down, so it comes out faster. An AI hands back a structure, and the writer judges it: close, not quite, change this part.
What helps isn’t the AI settling on the right answer in one go. It’s the back-and-forth: getting a rough version out quickly, then having the writer correct it. Having something outside your head that you can correct becomes the way in to keeping a thought and giving it form.
Related work
This idea is about one person picking up a thought where they left off. Our paper “Organizational memory for interchangeable agents” is about the memory that lets an agent pick up a task.
Moss, the multi-agent system we are building at Our World, extends practical work and execution; this idea extends thinking and strategy. The goals connect, but they are separate things.
Open questions
- The unit of thought. How do you describe not a single piece of information, but the temporary shape formed by concepts, relationships, arrangement, and state?
- What counts as success. A record that survives, a record you can understand on rereading, a thought you can pick back up, a thought you can act on — these aren’t the same. Where is something lost, and which methods support which?
- What moving does. How do you tell the experience of thinking by moving nodes apart from a diagram of thinking already done? Which parts of an arrangement need to be kept to bring an understanding back?
- What rekindles a thought. Even with the drive, a thought can be hard to return to after a break. When do a center to return to, an earlier arrangement, or random connections act as a nudge that comes before the question?
Version history
- v0.1Oct 7, 2026First version.
This document’s URL will not change. When citing it, please include the version number.