Think about the last time you took a note and it felt good…
By the VEKTOR team · 18 min read
Think about the last time you took a note and it felt good. Not productive. Not organised. Just good. That frictionless moment where a thought landed somewhere safe and you could move on, to another one, linking them together whilst adding depth.
For most of human history, that was the entire contract. Put pen to paper. Done. The latency was zero. The interface was invisible. The cognitive overhead was nil.
Press enter or click to view image in full size
Now count the steps between having a thought and capturing it in your current setup. How many apps are involved? How many decisions? Tag it? Title it? Which workspace? Which AI persona? Summarise now or later? The idea is already half gone by the time you’ve decided.
We did not set out to build a note app. We set out to understand why a task this simple had become this hard, and whether there was a better way.
What follows is what we found.
The Evolution Nobody Asked For
Note-taking has passed through four distinct eras, each one promising to make capture easier, each one quietly adding more complexity:
Pen and paper. Instant. Tactile. Permanent. Zero setup. The original frictionless interface.
Keyboard and mouse. Notepad.exe, then Word. We gained searchability and copy-paste. We lost portability and gained file management.
Cloud and sync. Evernote, Notion, Obsidian. We gained access anywhere, rich formatting, and databases. We gained folders inside folders inside folders, and the anxious question of whether everything is organised correctly.
AI augmentation. Every modern note app now has an AI button. Summarise. Rewrite. Extract action items. Ask your notes a question. But the prompting burden fell entirely on the user, turning capture into a workflow with preconditions.
If note-taking was supposed to get simpler with every generation of tools, why does it now require a PhD in prompt engineering to capture a stray idea on a Tuesday morning?
The answer is not that the tools are bad. The tools are technically impressive. The answer is that we optimised for feature completeness instead of cognitive lightness. Every added capability came with a new decision point. Every new decision point added friction. And friction, at the moment of capture, is the enemy of thought.
Press enter or click to view image in full size
The Behavioral Science Nobody Read
There is a law in cognitive psychology called Hick’s Law: the time it takes to make a decision increases with the number and complexity of choices available. More buttons do not make an interface more powerful for the user, they make it slower.
Research on knowledge worker productivity consistently shows the same pattern. Users ignore most of what is on screen. They develop muscle-memory paths through interfaces and rarely deviate. When a UI changes — adding a new AI panel, a new sidebar, a new context menu, productivity drops sharply while users relearn, then partially recovers as they ignore the new feature.
Press enter or click to view image in full size
We spent six months working with a wide range of AI-augmented note tools. A common pattern emerged: the technical problem of AI integration had largely been solved. The harder problem is when and how AI should enter the user’s flow, remained largely open.
Most inserted AI at the wrong moment — either too early (interrupting mid-thought with suggestions) or too late (requiring the user to trigger it manually after the fact). The result was the same in both cases: friction, context-switching, and the slow erosion of the note-taking habit itself.
This is the Clippy Paradox. Microsoft’s infamous assistant failed not because it was stupid — it was actually reasonably capable for 1997 — but because it interrupted without context and offered help the user did not ask for. The pattern keeps repeating across the industry: more AI surface area, more interruption points, more decisions handed back to the user.
The Design Problem Is Architectural
After twenty failed experiments and six months of interface iteration, we kept arriving at the same conclusion: the design problem is not aesthetic. It is architectural.
Most note apps treat AI as a layer on top of capture. You write, then you invoke AI, then AI responds, then you decide what to do with the response. This is the prompt-response loop, and it places all the synthesis burden on the user.
What if the architecture were inverted? What if the AI synthesised while you wrote — not interrupting, not demanding input, but building a parallel understanding that you could glance at, use, or ignore?
That question led us to the interface you see in VEKTOR’s JOT mode: a strict visual split between Thoughts and Synthesis.
Press enter or click to view image in full size
Thoughts — left panel
Raw capture. No formatting required. No AI interruptions. Write exactly what is in your head. The interface disappears. This side belongs entirely to the human.
Synthesis — right panel
The AI works here, quietly, 600ms after you pause. It reads what you wrote and surfaces connected insights, patterns, and implications — without asking. You can ignore it entirely or click any idea to expand it.
This split is not a UI preference. It is a statement about where human cognition ends and machine synthesis should begin. The line between them should be visible, literal, and respected.
Press enter or click to view image in full size
The Zen Constraint
Early versions of VEKTOR JOT had sixteen toolbar buttons, a sliding temperature control for AI creativity, four AI modes, and a tag management system. Users had to make eleven decisions before writing a single word.
We threw all of it away.
We kept iterating toward what we called the Bento Box principle: compartmentalised, clean, and bounded. Each element of the interface has exactly one job. Nothing overlaps. Nothing competes for attention at the moment of capture.
Write on Medium
Design Principle
The best note interface is the one that disappears. A user in flow should not be able to remember whether they used an app or a napkin. Every visible element is a cost paid against that ideal.
The toolbar reduced to a single icon row that stays hidden until you need it. The AI temperature slider became a five-position mode control: Precise, Balanced, Creative, Deep, Fast — because a label is comprehensible and a number is not. The tag system became automatic. The merge button moved from a prominent header element into the toolbar where it belongs, used only when needed.
Each removal made the tool feel more at ease, calm. This runs counter to how most product teams think about features. It is worth reflecting on.
Press enter or click to view image in full size
Making the AI Actually Proactive
The hardest problem was synthesis timing. We wanted the AI to surface ideas before the user asked for them — but not in the Clippy way. The difference between helpful and annoying is almost entirely a function of timing, relevance, and interruptiveness.
Our first implementation debounced synthesis at 1,800ms and sent the entire note document to the model on every trigger. This meant:
The user paused, waited nearly two seconds, then waited again for the model response
The synthesis panel updated all at once with a jarring flash of new content
Sending the whole document on every call was slow and expensive
A previous slow response could arrive and overwrite a newer one
Future revisions will tailor this even further, the exact moment and amount if data given, micro llm calls.
None of this felt proactive. It felt like an invoice arriving after you’d already forgotten the purchase.
The solution required three architectural changes working together:
AbortController on every request. Each new keystroke burst cancels the previous in-flight synthesis call. No stale responses. No overwrites. The model is always working on the current state of the document.
Tiered micro-prompts. The prompt scales with what’s been written. Under 20 words: one sharp insight, one sentence. 20–70 words: three key points. 70+: numbered synthesis with bold titles, but using only the last 700 characters — the most recent context — not the whole document, saving on tokens.
Streaming render. Rather than waiting for the full model response, the synthesis panel updates as tokens arrive. Words appear progressively. A blinking cursor signals live generation. The user sees the AI thinking in real time, not a sudden page-refresh of completed text.
The Result
Debounce dropped from 1,800ms to 600ms. The synthesis panel feels responsive rather than lagged. Ideas appear in the right panel while the thought is still warm. And because it never interrupts the left panel, the user’s flow is unbroken.
The numbered synthesis items are themselves clickable. Tap any circle and the AI expands that idea inline — three to five sentences of additional depth, examples, and implications — with a micro-prompt that takes under two seconds. The interface becomes a thinking partner rather than a results page.
Press enter or click to view image in full size
The Technical Layer Underneath
None of this would be possible without a persistent memory layer underneath the interface. This is where VEKTOR’s architecture diverges fundamentally from every other note app we tested.
Most note apps store text. VEKTOR stores understanding. Every note ingested passes through an embedding pipeline that encodes its semantic meaning into a local vector index. Every think query runs associative recall across that index before generating a response. The AI is not answering in a vacuum — it is answering in the context of everything you have ever stored.
MCP- DXT VEKTOR MemorySQLite + Vectors Local Embeddings Skill Files Associative Recall
MCP (Model Context Protocol) is the nervous system. Standardised by Anthropic, it is the universal connection layer between AI agents and the tools and data sources they need. VEKTOR exposes its memory graph via MCP, which means any MCP-compatible client — Claude Desktop, Cursor, Windsurf, your own agent — can query your memory without extra configuration.
DXT (Desktop Extensions) is the delivery mechanism. It packages VEKTOR’s tools into a one-click installable bundle that eliminates the environment setup, dependency management, and configuration hell that stops most developers from using local AI tools at all.
Together, these two technologies allow VEKTOR to operate as what we call a Persistent Intelligence Layer: a background system that every tool you use can query for context, history, and synthesised understanding, without you having to manually provide it.
Press enter or click to view image in full size
Have We Actually Solved It?
Honest answer: partially.
The core thesis — that proactive synthesis at low friction is better than reactive AI triggered by user commands — holds up. The split interface reduces decision overhead significantly. The streaming synthesis feels alive in a way that batch responses do not. Users who have tried JOT report that it is the first AI note tool where the AI helps rather than interrupts.
But the problem runs deeper than any single interface can solve. The real tension is not between capture and synthesis. It is between the human desire to just think, and the machine’s need for structure in order to retrieve and connect. Every system that helps you store also creates a new retrieval problem. Every synthesis creates a new organisation problem.
The best version of AI note-taking is not one that does more. It is one that makes you feel like you are doing less — while quietly doing significantly more underneath.
That is the standard we are building toward. Not a prettier notebook. Not a smarter prompt box. A system that accumulates understanding over time, surfaces the right context at the right moment, and stays invisible until you actually need it.
The Clippy failure was not a failure of intelligence. It was a failure of timing, relevance, and restraint. Those three constraints are harder to engineer than any language model. They require knowing not just what is helpful, but when help becomes intrusion.
Press enter or click to view image in full size
What Comes Next
Version 1.5.2 of the VEKTOR Slipstream SDK ships with the JOT split interface, streaming synthesis, tiered micro-prompts, and the clickable synthesis expansion system described in this article. The follow-up expander in DESK mode — where each AI-suggested question opens an inline knowledge panel rather than firing a full new query — ships in the same release.
On the horizon: cross-session synthesis briefings (a daily digest of what your memory graph has connected overnight), ambient signal surfacing (relevant notes appearing proactively as you type, not in response to a command), and deeper MCP integration so that third-party agents can pull synthesis directly from your memory without a context window overhead.
The goal remains unchanged from the first line of code: let your AI fetch its own context. Stop prompting. Start building a persistent mind.
Top comments (0)