Goldfish - Press Option. It knows your work and replies like you
by•
Most AI tools make you explain the context before they can help. Goldfish already has it. It privately remembers what you’ve been working on across your Mac, then helps you write better from any app. Press Option in a text field to draft replies, summarize threads, rewrite sentences, or recall important details from your recent work without copying, pasting, or re-explaining the whole backstory.
Replies
Congrats on the launch! I have a small question, with Goldfish capturing context across so many apps, how does it filter out irrelevant noise when you're rapidly switching between unrelated projects throughout the day?
@crystalmei Great question. Goldfish uses the current app, focused field, surrounding UI, and recent activity as the strongest signals, so it does not just dump one giant memory into the prompt.
When you press Option, it ranks the context by relevance to what you are doing right now, then pulls in only the pieces that match. If you have been jumping between unrelated projects, the current window and field usually anchor it pretty well, and the noisy stuff gets ignored. It is also designed to be explicit about uncertainty rather than blending contexts together.
@crystalmei i good question. Goldfish starts from the exact thing you’re doing, the app, field, thread, and very recent context, then only pulls in broader memory when it helps. Random project switches should mostly get ignored, not blended into the reply 🐠
This is great! How big can the context get?
I also saw you using graphs, I guess it is to increase accuracy. So how is the memory layer structured?
@vugar_javadov the context can get pretty large, but we keep it scoped rather than just dumping everything in.
The rough structure is:
1. Immediate screen/app context from the current field
2. Recent activity from the last few minutes/hours
3. Longer-term memory summaries and searchable snapshots
4. A small identity/voice layer for how you usually write
The graph is mostly there to connect people, projects, apps, threads, docs, etc. so the memory retrieval is more accurate than plain keyword search. When you press Option, Goldfish pulls the relevant slice for that exact surface and intent, then ignores the rest.
@vugar_javadov yeah exactly. graphs are one part of it, but the main thing is scoped context, current surface first, recent exact activity second, and longer-term memory after that.
curious what your best use cases have been so far? 🐠
Just got set up. Onboarding was a bit clunky but overall really novel and refreshing. Love it so far. Interested to see how useful this is after a few weeks of use. The option key doesn't seem to work with Atlas browser.
@goldhaxx Thanks for flagging this, Zach! Atlas is likely blocking or exposing the focused field differently than Safari/Chrome, so we’ll take a look and add a fix if we can. Appreciate the comment - we're looking into increasing stability as we speak! 🙏🐠
@goldhaxx yeah Atlas support is useful to know, we’ll fix it asap.
what are your best use cases so far? 🐠
the recall important details from your recent work use case is the one that interests me most and also the one hardest to evaluate from a description. the value depends entirely on how reliably it surfaces the right detail at the right moment versus surfacing something adjacent that's close but wrong. an AI that confidently recalls the wrong version of a number or date from recent work is worse than no recall at all. what does a typical recall failure look like and how does the user know when to trust it
@ansari_adin totally fair question. The dangerous failure mode is exactly what you describe: not “it says I don’t know”, but confidently pulling in an adjacent detail from the wrong thread, old draft, or similarly named project.
We try to make that less likely in a few ways:
- recent visible context wins over older memory
- exact recalls use the raw captured text, not just a summary
- results are grounded in the source surface, like the app/thread/doc it came from
- if the user asks for something specific like a number, date, quote, or commitment, Goldfish is instructed to open the underlying snapshot before answering
A typical failure today is usually “right topic, wrong slice” rather than total hallucination. For example, it might find a related Product Hunt comment when you meant a LinkedIn DM about the same launch. That’s why we care a lot about showing enough provenance and keeping the answer scoped to what the user is actually doing right now.
Long term, trust comes from the product being willing to say “I don’t have enough context” instead of guessing. That is probably the most important behavior for recall.
@kar_re right topic wrong slice with visible provenance is a much more manageable failure mode than confident hallucination. the willingness to say I don't have enough context is the behavior that determines long-term trust and it sounds like you're already thinking about it the right way
@ansari_adin
@ansari_adin exactly. The trust bar for recall is much higher than for normal writing help. A slightly awkward rewrite is harmless, but a wrong date or number can actually cause damage.
That’s why we’re treating “I don’t have enough context” as a product feature, not a fallback.
@ansari_adin Totally. That’s the line we’re trying to hold: useful recall, but only when the source context is clear enough to trust. If it can’t tell which slice you mean, it should say that rather than smooth over the uncertainty.
@ansari_adin totally fair. the scary failure is when it sounds certain about the wrong nearby detail.
The way we think about it: current surface first, recent exact context second, broader memory last. And for names, dates, numbers etc, it should only say them when it can ground them in something visible.
So trust should come from predictable failure modes, not magic. If Goldfish is unsure, it should ask, stay vague, or show its source instead of bluffing 🐠
Been using Goldfish daily and it really does feel like the missing memory layer for work. The magic is when it pulls context from another app and writes the reply without me having to explain anything. Huge launch 🐠
@hugo_mardh1 Thanks Hugo, really happy to hear that. The cross-app context is definitely where it starts feeling magic 🐠
Do you use the desktop pet anything?
@hugo_mardh1 you’re one of the people who instantly got the cross-app magic. thank you for backing us 🐠
Using Goldfish to comment on Goldfish feels illegal, but also kind of the point. Finally no more copy pasting stuff between my chatbots and messaging boxes!!🐠
@ceciliatran hahaha exactly, feels very full circle 🐠
What has been your favorite use case so far (except for PH comments? :- D)
@ceciliatran haha love it! Thanks Cecilia! Let us know once you've tried it more!
I’ve been using Goldfish for a while now for drafts, summaries, prompt tuning, and a bunch of other day-to-day work.
It’s been genuinely useful, and we’ve rolled it out to our whole team too.
Strongly recommend!
@lmtr211 Thanks for the huge support Gustav, this is so nice to hear 🐠
@lmtr211 That's SICK! What made you wanna share it with the team?
Been using it for the last weeks, have to say I'm quite impressed 😎. It catches your tone quick, looking forward to seeing your guys' continued progress.
@andy_duly Thank you Andy! Really glad to hear it’s been working well for you 😎
@andy_duly thanks Andy, really appreciate it. tone is the part we obsess over the most, so this means a lot 🐠 any improvements we can do?
Repetitive context creation. Is it now a thing of the past?
The most annoying thing with AI is always having to repeat the same personas, the same context, and writing things in your own words again and again. I've only been using Goldfish for awhile, but having built-in context without needing to create custom contexts/skills for each tool that I use daily has been great.
I'm looking forward to using Goldfish more and see how much it can really handle. Glad it doesn't have a Goldfish memory, or that would have been troublesome.
@travis_isaacson Thanks Travis! And haha, very glad we dodged the Goldfish memory problem.
Would love to hear where it starts breaking once you’ve used it more. That “how much can it really handle” question is exactly what we’re trying to learn now.
@travis_isaacson thank you Travis, really appreciate it. repetitive context creation is exactly the thing we wanted to make disappear without forcing you to set up a new persona or custom skill for every tool.
curious what your best use cases have been so far? 🐠
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.
@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!