Shelf V1The beginning
it started with a simple idea: don't lose the things you care about.
shelf didn't start with a big vision.
it started with me saving things.
a screenshot here. a link there. a piece of text i knew i'd need later. something inspiring. something useful. something i wasn't ready to forget.
and then, inevitably, i'd forget where i'd put it.
the thing was still somewhere on my mac.
i just couldn't find it anymore.
saving wasn't the problem
there were already hundreds of ways to save things.
bookmarks.
folders.
notes.
cloud drives.
read later apps.
screenshots.
downloads.
clipboard managers.
the problem wasn't that we couldn't save enough.
the problem was that saving something usually created another place we had to remember.
where did i save it?
which folder?
which app?
what did i call it?
when did i save it?
the more things you saved, the harder that system became to maintain.
i didn't want another place to organize.
i wanted somewhere i could just put things.
so shelf became that place
the first version was deliberately simple.
see something worth keeping?
put it on the shelf.
a screenshot.
a link.
a piece of text.
a file.
an image.
an idea.
whatever.
no need to decide where it belongs first.
no folder hierarchy to think about.
no elaborate system to maintain.
just save it and keep moving.
that was the first version of the idea:
make saving effortless enough that you don't have to think about saving.
everything stayed close
i also didn't want shelf to feel like another cloud service.
your things should feel like they're actually yours.
so shelf was built local-first from the beginning.
your memories live on your mac.
the app works around your existing workflow instead of asking you to move your life into another ecosystem.
that constraint shaped the product more than i expected.
it made me think carefully about what deserved to happen locally, what needed intelligence, and what information should never need to leave your machine.
the shelf metaphor mattered
the name wasn't accidental.
a shelf is somewhere you put things you want nearby.
not necessarily organized perfectly.
not necessarily categorized.
just within reach.
you don't need to remember which box you put something in.
you just know it's somewhere on the shelf.
that was the feeling i wanted.
a place between the chaos of the desktop and the rigidity of a filing system.
somewhere things could wait until you needed them again.
the first version was about capture
v1 was mostly about one thing:
don't lose the good stuff.
capture something quickly.
keep it locally.
come back to it later.
it was intentionally not trying to solve everything.
there was no grand memory graph.
no attempt to understand every relationship between things.
no system trying to predict what you'd want tomorrow.
that would come later.
first, shelf needed to prove one simple thing:
maybe remembering could start with making saving effortless.
then i started using it
this is where the product started becoming interesting.
because once you have a place where everything you've intentionally kept ends up, patterns start appearing.
things from the same project start sitting together.
research starts accumulating.
screenshots start forming visual references.
old links become useful again.
something you saved three weeks ago suddenly becomes relevant to something you're working on today.
the shelf started becoming more than storage.
it was becoming a record of what i found interesting.
what i was working on.
what i didn't want to forget.
and that raised a much bigger question.
what if shelf could remember too?
saving something tells shelf that it mattered once.
but it doesn't necessarily tell shelf why it mattered.
and months later, i might not remember either.
that's where v2 began.
the question stopped being:
how do i save this?
and became:
how do i find this again when i only remember part of it?
that distinction changed the direction of Shelf.
v1 was the beginning
looking back, v1 was intentionally small.
it wasn't supposed to be the final answer.
it was the first layer.
a place to collect the things that matter before they disappear into the rest of your computer.
and that was enough to start.
because before a computer can remember something for you, you need somewhere to put the things worth remembering.
v1 built that place.
v2 started teaching it how to remember.
shelf v1
a place for the things you don't want to lose.
- capture
- context
- memory
v1 gave shelf a place to keep things.
v2 asked what it would mean for shelf to remember them.
Shelf V2The shift
a better way to remember what mattered.
i didn't want to build a smarter clipboard.
shelf started with a small, annoying problem.
you save something because it matters. a screenshot, a link, a voice note, a paragraph you didn't want to lose. then a few days pass and you remember the thing itself, clearly, but you have no idea where it went.
saving was never the hard part.
every app on your mac can save things.
getting them back out is where all the friction lives.
that's the part v2 goes after.
the mental model changed
most tools treat information as objects.
files go in folders. folders get tags. tags get searched.
it's a filing cabinet with better lighting.
people don't remember that way.
you remember why something mattered. what it was sitting next to. roughly when you saw it. who sent it. what you were working on at the time.
so shelf v2 stopped asking:
where did you put it?
and started asking:
what do you remember about it?
that sounds like a small reframe.
it changed almost everything downstream.
memory has context
a screenshot is rarely just a screenshot.
it might be a reference for a project you're deep in. something a friend sent you at 1am. one tab out of forty from a research session. an idea you flagged to come back to.
or something you saved for a reason you couldn't have explained at the time.
which is often the most useful kind.
v2 gives shelf a richer read on the things you save.
what something is, yes.
but also what it seems to belong to.
the project around it. the time around it. the people around it. the other things that make it make sense.
search became remembering
you shouldn't have to remember a filename to find something.
or the folder.
or the exact words you used when you saved it three months ago.
you can remember approximately.
that's how memory actually works.
that screenshot about the landing page
the article i saved about ai interfaces
that thing i saw while researching the project
all three are vague.
all three are enough.
shelf searches by meaning, so it meets you where your memory already is instead of asking you to meet it at a filing system.
the interesting part is what sits around it
finding one thing is table stakes.
the more interesting question is what happens after you find it.
when you land on a memory in v2, shelf can surface the things around it.
similar things.
related things.
things from the same afternoon, the same project, the same person.
references that quietly belong together even though you never told shelf they did.
quantity was never the problem.
anyone can dump fifty results onto a screen.
the work is picking the right handful.
that's much harder to get right.
and much more useful when you do.
sometimes you don't know what to search for
this was one of the biggest ideas behind v2.
and probably the one i went back and forth on the longest.
some memories are useful precisely because you forgot about them.
you can't search for something you've stopped thinking about.
so shelf can bring an old memory back on its own, but only when it has a real reason to believe it matters right now.
no feed.
no notification spam.
no streaks, badges, or daily digests designed to pull you back in.
just:
you might want this again.
and then shelf gets out of the way.
intelligence should feel invisible
i spent a lot of time on this one.
the best version of shelf doesn't feel like an ai product.
it shouldn't keep announcing itself:
ai found this.
ai connected these.
ai thinks you might like this.
nobody asked for that.
the intelligence should stay underneath the experience.
it should simply feel like shelf understood.
a connection tells you why it's a connection.
a collection tells you why those things belong together.
a rediscovered memory shows up with a reason attached.
the reasoning stays visible.
the machinery stays underneath.
the principle behind v2
one line ended up settling most of the hard calls:
earn the right to interrupt memory.
if shelf isn't confident something is useful, it says nothing.
if two things don't actually belong together, it doesn't connect them.
if an old memory has no reason to come back today, it stays where it is.
silence is a feature.
a memory tool that talks constantly is just another thing in your way.
less database, more memory
v2 isn't really a pile of new features.
it's a change in what shelf is trying to be.
a little less database.
a little more memory.
a little less organize everything.
a little more remember what mattered.
that distinction shaped nearly every decision in this release.
including the decisions where the right answer was to build less.
what's next
i'm shipping v2 and letting it live for a while.
the next version shouldn't come from a roadmap alone.
it should come from what actually happens once this is in people's hands.
what people save.
what they search for.
what they come back to.
what they rediscover.
and what they ignore completely.
that last one usually teaches you the most.
the question i keep circling isn't:
what feature should shelf build next?
it's:
what does a person actually want their computer to remember?
shelf v2
remember less about where you put things. remember more about why they mattered.
Shelf V2.5The place
the memory became a place.
v2 taught shelf how to remember.
v2.5 is about what happens once those memories start accumulating.
a memory is no longer just something you find and open.
it can be a doorway.
ideas tell their own story
an idea used to be a quiet grouping. now it has somewhere to go: a name, how many memories sit inside it, how it grew, and the reason those memories belong together.
that last part is the one that matters.
shelf doesn't just group memories. it explains why they belong together.
because these memories belong to Atlas.
its own words.
what it's about.
where it came from.
related is not the same as similar
these two words get used interchangeably. they shouldn't be.
similar asks whether two memories resemble each other.
related asks whether a memory belongs to the context of an idea.
one is a resemblance. the other is a reason.
so only relationships shelf can actually point at show up. if there's no evidence, there's no connection. an invented one is worse than none.
exploring instead of managing
gallery, cards, timeline, ideas, connections, threads. in v2 these were useful, separate places. in v2.5 they behave like one memory space you can move around inside.
browsing should feel like walking through your own archive.
not administering a database.
memory should never become a dashboard.
longer histories reveal themselves
some things you keep for an afternoon.
some things you keep returning to for a year.
threads now notice the difference. a short span stays simple. a longer span becomes month aware, then year aware, and an idea you've been circling for months finally looks like it.
shelf doesn't write the story.
your own memories make it.
what does this belong to?
any memory can now show you the ideas and context around it, in plain language.
connected to 3 of these.
part of this idea.
saved while you were on this.
no graph vocabulary. no scores. no machinery on display.
the graph stays underneath. the memory stays in front.
rediscovery stays quiet
nothing about this got louder in v2.5.
no feed. no streaks. no notifications built to pull you back.
an old memory returns only when shelf has a real reason, and it brings the reason with it.
silence is still a feature.
you don't lose your place
following a connection used to mean starting over. idea, memory, related memory, and then wondering how to get back.
now you can go and return.
curiosity shouldn't cost you your position.
exploring should feel continuous, not fragmented.
one system, one language
cards, gallery, timeline, inspector, ideas, connections. same card language, same metadata, same tone.
they're not separate features anymore.
they're different views of the same memory.
what v2.5 is really about
v2.5 adds less intelligence than you'd expect.
most of the work went into making the intelligence shelf already had feel like one coherent product.
less organization.
more understanding.
less database.
more memory.
accumulation should make shelf more valuable, not heavier.
what's next
v2.5 is about making memory easier to explore.
the next question is what happens when shelf has years of it.
shelf v2.5
the memories you save become a place you can explore.
Shelf V3THE ERA
shelf rebuilds what happens after you save something.
shelf v3 is the biggest version of shelf so far.
v2 made saving things feel lighter.
v2.5 made the product more coherent, reliable, and understandable.
v3 rebuilds what happens after you save something.
shelf can now read more of what you keep, understand it, find relationships between memories, and bring things back when they become useful again.
the goal hasn't changed:
you save. shelf remembers.
a new memory architecture
v3 rebuilds shelf around memories rather than files.
images, screenshots, links, pdfs, documents, notes, and other things you save now share a more consistent underlying model.
shelf understands more about:
what something is
where it came from
what it contains
when you saved it
what makes it relevant
what it resembles
what it may be connected to
this foundation powers most of what is new in v3.
natural search
search the way you remember things.
you no longer need to remember the exact filename, title, url, or words inside something.
try things like:
that screenshot about typography
the article about memory
dark website inspiration
pdf about consumer psychology
image with a red chair
things i saved about onboarding
shelf searches across the information it can understand about your memories rather than relying only on filenames.
search inside
shelf can now look deeper inside what you save.
text can be extracted and indexed from supported:
screenshots
images
pdfs
documents
webpages
that means something doesn't need a useful filename to become searchable.
if the information exists inside the memory, shelf can help you find it.
on-device ocr
text inside images and screenshots can now become part of your memory.
shelf uses on-device processing to recognize visible text without sending your images to an external ocr service.
a screenshot of a paragraph, interface, receipt, presentation, tweet, diagram, or note can become searchable alongside everything else.
better pdfs
pdfs are no longer treated like opaque files.
shelf can understand and index their text while preserving the original document.
pdf memories now have richer information for:
search
summaries
previews
related memories
retrieval
large documents are handled without needing to turn the entire pdf into a giant block of visible text.
smarter bookmarks
links now become richer memories.
shelf can capture and preserve useful information about a webpage so a bookmark isn't just a url you eventually forget.
depending on what is available, shelf can understand things such as:
title
source
description
page information
preview
searchable content
context around the bookmark
your links become part of the same memory system as everything else.
rich memory previews
saved things now tell you more before you open them.
memory cards have been rebuilt around a shared system with richer previews for different kinds of content.
images still feel like images.
pdfs feel like documents.
bookmarks feel like pages.
files retain their identity.
the interface adapts without turning every memory into the same generic card.
automatic summaries
shelf can create compact representations of supported memories so you can understand what something contains without reopening it every time.
summaries are treated as additional context, not replacements for the original memory.
the source stays intact.
similar
sometimes you don't remember what you're looking for.
you just remember that it looked or felt similar to something else.
v3 introduces similarity as another way to move through your memory.
shelf can use visual and textual signals to surface things that resemble the memory you're looking at.
one memory can become the doorway to another.
visual understanding
images are becoming more than thumbnails inside shelf.
v3 introduces a local visual representation layer that allows shelf to reason about visual similarity without uploading your library somewhere else.
this creates the foundation for things like:
visually related memories
image discovery
reference exploration
visual search
better resurfacing
connections
saved things rarely exist alone.
an article may relate to a screenshot. a screenshot may relate to a pdf. a reference image may connect to something you saved months earlier.
v3 starts making these relationships visible.
connections are generated from evidence shelf already has about your memories rather than asking you to manually build a knowledge graph.
smart collections
folders depend on you remembering where something belongs.
shelf can now form useful groups around what it understands about your library.
collections can emerge from things such as:
topic
source
content
similarity
type
context
patterns in what you've saved
you can still organize things yourself.
you just don't have to organize everything before it becomes useful.
rediscover
some memories shouldn't require a search.
v3 gives shelf a way to quietly bring older things back when they may be worth seeing again.
rediscover is intentionally restrained.
no engagement feed.
no attention machine.
no constant notifications.
shelf should earn the right to interrupt you.
infinite canvas
some memories make more sense together than they do in a list.
v3 introduces a spatial canvas for exploring and arranging things visually.
bring images, references, screenshots, documents, links, and other memories into one space.
use it for:
moodboards
research
visual thinking
collecting references
exploring ideas
connecting related memories
your canvas lives alongside your library rather than becoming another separate place to manage.
better capture
saving something should be the easiest part.
v3 continues the work of creating one consistent capture system across different kinds of memories.
shelf takes care of more of the processing after capture so you can spend less time preparing things before saving them.
save first. understand later.
a more useful memory view
opening a memory now gives shelf more room to explain what it knows.
depending on the memory, you can move through its:
original content
metadata
extracted information
summary
related memories
similar memories
source
context
the memory itself remains the centre of the experience.
intelligence stays around it.
better evidence
shelf shouldn't pretend to know something it doesn't.
v3 strengthens the distinction between the original memory and information shelf derives from it.
where possible, the product preserves evidence about why something was found, understood, summarized, or connected.
the goal is not to make shelf sound intelligent.
the goal is to make shelf trustworthy.
local-first intelligence
a major constraint shaped v3:
shelf should not require your personal memory library to live in somebody else's cloud.
core v3 processing remains local-first.
shelf makes use of local system capabilities and on-device processing wherever practical.
no shelf account is required.
no shelf cloud stores your library.
no analytics are used to build a behavioural profile of you.
your memories stay yours.
no cloud dependency
v3 does not turn shelf into a cloud service disguised as a mac app.
your library remains on your mac.
core functionality isn't designed around:
cloud storage
remote memory databases
mandatory accounts
paid per-request infrastructure
uploading your library to shelf servers
this constraint made some parts of v3 harder to build.
it also made shelf much closer to the product i wanted from the beginning.
performance and reliability
a lot of v3 exists underneath the interface.
the memory pipeline has been rebuilt so additional understanding can happen without every new capability becoming its own disconnected system.
work across v3 included improvements to:
memory ingestion
indexing
metadata
text extraction
ocr
bookmark processing
pdf processing
visual representations
search
memory relationships
background processing
persistence
failure handling
the visible features are only one part of this release.
a quieter interface
more intelligence does not need to mean more interface.
v3 continues simplifying shelf around the memory itself.
actions appear where they are useful.
secondary information stays secondary.
automation should remove work instead of creating dashboards that need to be managed.
the product is intentionally quiet.
better mac support
shelf continues adapting itself to the mac it is running on.
the experience accounts for both notched and non-notched macs rather than assuming one hardware configuration.
shelf should feel native to the machine instead of sitting awkwardly on top of it.
privacy
shelf remains built around a simple idea:
your memory is personal.
v3 continues to avoid:
mandatory accounts
advertising
behavioural analytics
telemetry-driven profiling
cloud storage of your shelf library
where shelf can do something locally, local is the default.
what changed philosophically
v3 isn't about turning shelf into a more complicated bookmark manager.
it moves the product further away from organization software entirely.
most tools ask:
where should this go?
shelf is increasingly interested in a different question:
how will you find this again when you've forgotten where it came from?
that changes almost every product decision.
folders become less important.
perfect naming becomes less important.
manual tagging becomes less important.
memory becomes the interface.
the direction
shelf began as a small place to keep things.
v3 establishes the foundation for something larger:
a private memory layer for your mac.
a place that doesn't just store what you save, but gradually becomes better at helping you remember it.
not another second brain.
not another productivity system you have to maintain.
just somewhere you can put something down and trust that it won't disappear from your life.
saving isn't remembering.
shelf is trying to close that gap.
shelf v3
a private memory layer for your mac.
What's next
the question is getting bigger.
shelf started as a place to keep things.
then it became a way to find them again.
now the interesting question is what happens when your computer starts understanding the memories you've intentionally kept.
the next version shouldn't simply add more features.
it should make shelf better at one thing:
helping you remember what mattered, when it matters.