Goldfish is basically the “wait, what was I just doing?” layer for my workday. Perfect for people with mid-quarterly-crisis dementia, that are busy, lazy, context-switching too much, or just someone prompt AI tools a million times a day and expects it to remember the plot. :)
The best part is that it understands the messy in-between state of work, not just neat files and folders. Stuff that simply connecting integrations never would pick up on. It can recap what I was doing, help rewrite a message, or pull up the random detail I know I saw somewhere 20 minutes ago.
The thing I use most is the option function. I can write something sloppy in Slack, email, LinkedIn, or a prompt, and Goldfish turns it into something clear and useful without me having to spend extra cognitive energy on phrasing. Slightly dangerous long term because I am clearly outsourcing both creativity and execution, but extremely useful day to day. Simply excellent. Kudos.
Love the idea of pressing Option to draft replies in your own tone. How far back does Goldfish remember your work context?
Goldfish
@doganakbulut It keeps a short buffer of the very recent context, which is usually enough for the reply/summarize moment. If it needs more, it can search longer-term local memory for the relevant project, person, or thread.
All of that memory stays on your Mac, and you can exclude apps/domains or pause capture whenever you want.
@doganakbulut we keep a recent working memory for the stuff you’ve actively had on screen, plus longer-term summaries/wiki pages for projects, people and docs you touch repeatedly.
so in practice: it can answer “what was I just working on?” from the last few hours, and also remember broader context like ongoing projects, recurring contacts, launch docs, etc. the goal is enough memory to write the right reply without you re-explaining the thread every time 🐠
But it stores exact text of everything, so goldfish could technically recall an exact quote from months ago. Do you think it works well for you so far?
this is very very cool, just tried it! :)
is it able to change the tone of your reply to say your grandmother vs your best friend?
Goldfish
@armaan_nagra Happy to hear that you like it, Armaan! :)
Yes, it knows the tone of you depending on in what context and who you're responding to. One user told us it understood even a certain dialect of a language where she only wrote in that dialect with family members, so hopefully it understands enough.
@armaan_nagra yes, it also stores memory about how your tone changes between different apps and sites! What apps do you think you'll use it most for?
@joel_edholm linkedin haha, i need to clear my inbox
Well done! Though where I'd get nervous is recall precision:) Once a month of my screen sits in that vector DB, what stops Option from grabbing the wrong thread when two clients have near identical projects open? Overall, solid work!
@artstavenka1 totally fair concern. the short answer is: Goldfish should not be “semantic search over your whole screen history and hope for the best”.
we’re treating context as scoped and ranked, with recency, active app/thread, focused field, and visible surface mattering a lot. if two client threads look similar, it should prefer the one you’re actually in, not some old near-match.
still early, and recall precision is exactly one of the things we’re obsessing over. thank you for calling it out 🐠 how do you think it works, if you've tried it?
Goldfish
@artstavenka1 super good question! We do use chronological relevance too, so if Goldfish gets any timing signal, we use that as a heuristic. Under the hood it has three representations: recent context, long-term memory, and chronological memory. Would love for you to try it and let me know how it goes!
SocialBu
Congrats on the launch, Joel! The 'starting from zero every time' problem with AI is incredibly real—I waste so much time re-pasting Slack threads and briefs. Keeping the memory layer local via SQLite/vector DB is a massive win for trust. Quick question: How does it impact Mac performance or battery life when it's actively indexing active apps?
Goldfish
@bilal_niaz I feel you, we had the same problem before goldfish existed! Yes indeed - local first is a winner in our opinion :)!
It doesn't affect performance too much, we utilize the already existing accessibility APIs to get the text from open windows, which is super lightweight. We also run an optimized model for doing embeddings as not to tax the computer unnecessarily much.
Have you given goldfish a try yet? Best way to assess performance also! :- D
@bilal_niaz thanks Bilal! We designed indexing to be lightweight and opportunistic: In normal use it should feel invisible, and we’re continuing to tune CPU/battery as more people run it day to day 🐠 But it is surprisingly light!!
Love the Option key shortcut, feels so friction less. How does it handle privacy with remembering work across apps? That balance of helpful and private is the hard part to get right. Excited to try it.
Goldfish
@shubham4real Thanks Shubham! Privacy is the main thing we designed around here.
Goldfish stores memory locally on your Mac, not in a cloud backend we can browse. You can pause capture whenever you want, exclude specific apps or domains, and delete stored history too. AI calls also have zero data retention.
That balance is definitely the hard part, so would love to hear where it feels too much or not enough once you try it 🐠
@shubham4real al thank you Shubham! The shortcut is only useful if people can actually trust it.
The way we think about it is: context should be helpful in the moment, controllable by the user, and easy to turn off when something is sensitive 🐠
wow this product is gold. well done!
I do have a few questions on the infrastructure:
Is a local model used to generate the replies?
How does It formulate a reply in the language (writing style) of the user?
Does it only work with what is given on the screen in realtime?
Goldfish
@ibrahimkakson thank you!! not local right now, so we use azure hosted models for the replies etc. Goldfish uses the context it can see on your screen in realtime, plus your private memory from previous work, if you have it enabled.
For style, it learns from how you actually write across apps and uses the current surface as context, so a Slack reply, email, Product Hunt comment, etc. come out differently! :- D
You can also control what it remembers/ignores, so it is not just blindly saving everything.
@ibrahimkakson thanks Ibrahim! short version:
No local model for generation yet. Goldfish reads the current surface first, then pulls in scoped recent context and your learned writing patterns, so the reply is grounded in what you’re actually doing and still sounds like you.
It’s not only realtime screen text, but the current screen always wins. That’s how we avoid random old context leaking into a reply 🐠
Local models will be the future for sure and we're experimenting with it
The Option-key workflow feels smart because it keeps AI inside the place you’re already writing. Curious how Goldfish decides which recent context should matter and which old context should fade out.
Goldfish
@farrukh_butt1 Very good question! We keep a buffer of the most recent-recent context which gets embedded with each option-click, but when it makes sense Goldfish can also search way back into the long term memory, but often the most recent memory is enough.
Have you tried it yet? What do you think? :)
@farrukh_butt1 butt1 thanks Farrukh! The short version is: current surface wins, recent context is weighted heavily, and older memory only comes in when it’s clearly relevant. We’re trying hard to make it feel predictable, not like a random search through everything you’ve ever seen 🐠 But it's pretty good at recalling exact quotes form several weeks ago!