People interpret, compare, and doubt constantly while they work — and almost none of that reasoning reaches the AI.
Contexty lets you capture those thoughts in place, then shows the AI's context back to you as a canvas you can inspect and correct.
In a within-subjects study (N=12) it improved task awareness, thought structuring, authorship and control — and snippet-grounded answers won 78.1% of blind comparisons.
AI sees your activity, not your reasoning.
Much of what shapes a complex task happens in passing — why a finding feels significant, how two options are being weighed, which criterion just shifted. These judgments arise mid-task and disappear just as fast.
Behavioural traces and screen observation can tell an AI what you did and where you spent time, but not the why and how behind it. Saying it out loud means stopping to write a prompt — and what you do write scatters across chat turns or gets compressed into a summary you never wrote. So the AI keeps working from a picture of your task that quietly drifts from your own.
Contexty treats those in-situ thoughts as the material the AI's context is built from — and puts that context somewhere you can see it, and correct it.
One shortcut, from wherever you already are.
A global shortcut grabs whatever is on screen and asks a single question: what were you thinking just then? A short memo — an interpretation, a doubt, a criterion — travels with the capture into the AI's context. No app switch, no prompt to compose.
Passive screen observation and chat feed the same context alongside it — three channels into one shared memory.
How snippet memoing supports cognitive externalization
Ten daily LLM users ran a real sensemaking task with an early probe. It captured not only what they explicitly asked, but their interpretations and partial conclusions as they unfolded.
Expressing user-framed perspectives and forward-looking thoughts
Memos captured how people were making sense of something, not just what they were looking at — and let them save a thought before they had a question to ask.
Lower barrier to externalizing thoughts in situ
No pausing to compose a prompt, no switching applications — people annotated continuously as they worked.
Captured snippets as cognitive scaffolds
Looking back at what they had seen and thought helped people hold continuity — making their reasoning visible to themselves, not just to the system.
What it couldn't do
As snippets piled up, people lost track of what they had captured and how the AI had read it — and leaving all of it to them to manage didn't work either.
Ground AI context in users' cognitive processes.
Take in-situ thoughts as first-class inputs, rather than relying on system-inferred signals alone.
Make AI context inspectable and correctable by users.
Externalize it so users can spot mismatches and fix them — keeping authorship and control with them.
Context you can look at — and correct.
Every capture becomes an item on a canvas, and the canvas is what the AI actually reads. There is no second, hidden copy of your context. Regroup an item, rewrite a label the system got wrong, delete what isn't relevant — and you have changed what the next answer is built on.
What one item holds
- ATitleGenerated so you can scan a full canvas at a glance.
- BProvenanceWhich app, which URL, what time — read from the OS at capture.
- CCaptured contentThe snip itself, kept verbatim.
- DTagsProposed by the system. Add or delete them.
- EAI's readingWhat the system thinks you were doing. Misread? Rewrite it.
- FYour memoThe in-situ thought, pinned to the card.
New captures are scored against what's already there and grouped with what they belong to.
A later capture can change what an earlier one meant; affected items get an “Updated” badge.
Drag, regroup, edit — or just say “group these by project” and it restructures.
Seeing the context changed how people worked.
Twelve participants, two real tasks each. The baseline kept snippet memoing, observation and grounded chat — and removed only the canvas, so what moved is what seeing your context adds.
picked the snippet-grounded answer
Every query silently produced two answers — one on the full context, one with all snippet-derived context stripped out. When they differed substantively, participants chose one without knowing which was which.
Structuring improved whether people actively reorganized the canvas or only glanced at it. Control came with a cost, though: passive observation added noise to curate — even if overall workload never moved.
BibTeX
@inproceedings{kim2026contexty,
author = {Kim, Yoonsu and Park, Chanbin and Son, Kihoon and Yang, Saelyne and Kim, Juho},
title = {Contexty: Capturing and Organizing In-situ Thoughts for Context-Aware AI Support},
year = {2026},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
booktitle = {Proceedings of the 39th Annual ACM Symposium on User Interface Software and Technology},
keywords = {Human-AI Collaboration, Context Sharing, User-Inspectable AI Context, Context-Aware AI Agents},
series = {UIST '26}
}
