Long chats break when the model forgets what you said ten turns ago. Oasis uses optional structured session memory (internal nickname Railroad ) to help follow-ups stay coherent.
Insight for other users: Session memory extracts structured facts across turns and re-injects them into later prompts so the assistant can remember details within your session and plan limits. You do not edit memory by hand. It updates as you chat.
What it is not: Memory is not a full backup of web pages. It is distilled text used to steer replies, not a copy of everything you browsed.
How to work with it:
Oasis Browser
Congrats on the iteration! The import flow mentions pulling passwords across from your old browser, and switching browsers, that's always the thing I look at hardest. You mentioned elsewhere that history, bookmarks and semantic indexes stay on-device while the assistant runs on cloud models. Where do imported passwords land in that split: on-device only, and encrypted how? And is the vault inside the same boundary, fully walled off from anything the trainable assistant can see?
Oasis Browser
@ferdi_sigona Big thanks for raising this important question!
Short answer: Imported site passwords stay on your device, in Firefox’s Login Manager vault (encrypted at rest). They are not part of the local history/bookmark/semantic-index layer, and the Oasis Assistant has no tool or API path to read the vault. Using the assistant still sends prompts and replies (and sometimes page text or tab URL/title) to cloud models—that’s a separate boundary from the password store.
Long answer on how everything works under the hood:
https://www.producthunt.com/p/kahana/oasis-browser-import-bookmarks-passwords-history-and-autofill-data-from-other-browser
Thanks Adam, that's the answer I was hoping for. Big congrats again!
Oasis Browser
@ferdi_sigona Awesome! Glad I could answer that. I appreciate your support 🫡
How do you give users peace of mind when using your AI instead of X or Y?
Do you support self-hosted models?
Also, I noticed the desktop navigation changes the URL, but the pages don’t actually load afterward on your website.
Oasis Browser
@gkanev Hey Gabriel, thanks for raising these points.
1. Peace of mind comes from the anonymous experience of using Oasis AI. None of your personal data is collected, only raw prompts and outputs which help us create our own "amplifier" model that seeks to improve the speed, accuracy, and quality of commands for you and the broader userbase. I explain this in detail in this video covering our privacy-first approach to training anonymously. You can also see examples of the exact type of JSON payloads that we receive when you send a prompt and learn more about this in our doc on interaction data.
2. We do not support self-hosted models yet, however, that is something we've been exploring and would be open to focusing on in the next sprints on the roadmap. We would also explore allowing you to use your own API keys for other models. Are there any local models you prefer? I've been using Qwen series off Ollama mostly.
3. I'm having trouble understanding your observation here. Which "desktop navigation" or URL are you referring to? Our main website is [https://kahana.co/] if you could point me to an exact page that is not loading, that would be helpful. Or am I missing something?
Thank you again for checking out our campaign and sharing these questions and observations :)
@adamthecreator
1. Thanks for the explanation.
Would love to see it, I’ll be waiting for it.
Youtube video - hopefully it’s already available. The issue I found is that on desktop Chrome (latest version on Mac), clicking menu items changes the URL but the page itself doesn’t load. If you force refresh the page, it works correctly, but only then.
Oasis Browser
@gkanev thank you so much for flagging the nav issue on Chrome (Mac), where the URL changed but the page didn’t update until a refresh. Thank you for going to the lengths to record a video. This is next-level feedback and support. I appreciate you so much.
We tracked it down to how the desktop dropdown menu was built (Products, Learn, About, etc.), including links like Oasis Browser in the Products menu. We’ve fixed that and improved how pages load when you click those items.
When you have a chance could you please try again on https://kahana.co/ and let us know if it’s still not working for you?
A hard refresh once (Cmd+Shift+R) is fine if you still have an old tab open.
Things to try:
From any page (e.g. Blog), open Products → click Oasis Browser
Try Learn → Blog and Pricing as well
You should see the URL and the page content change together, without needing a full reload.
Thanks again for reporting it 🫡
As a mobile game developer I spend a lot of time researching competitors and reading docs. The idea of a browser that learns your workflow is genuinely interesting — does it get better at surfacing the same types of sites you visit regularly, or is it more about how you interact with pages?
Oasis Browser
@jan_bremec Great use case! Thanks for sharing. Competitor research and doc spelunking is exactly the kind of workflow we had in mind.
Short answer: Today it’s more “help me find and work with what I’ve already seen” than “the browser predicts the next site you’ll want.” It gets more useful as your history, bookmarks, tabs, and hubs fill up locally, but the main lever is what you visited and what those pages were about, not deep tracking of how you scroll or click.
Two layers (worth separating)
1. On your machine — surfacing past work (no cloud model training your profile)
When you ask the assistant things like:
It can search:
Semantic history search — recent browsing (~500 pages) indexed on-device with embeddings built from title, URL, domain/path, and a short text snippet captured when the page was visited. So it’s biased toward meaning (“IAP,” “LTV,” “App Store”) not just “you opened this domain 40 times.”
Memory search — full-text search across open tabs, tab groups, bookmarks, and history (titles/URLs)
That index grows as you browse (incremental updates, persists across restarts). It does not today behave like a recommender that proactively pushes “you usually open Sensor Tower at 9am.” You invoke it via chat (“find…”, “what did I read about…”) or related tools.
2. How you interact with a page — when you’re on it
That’s a different path: tools like summarize this page or questions grounded in the active tab read visible page content and send an excerpt to the cloud LLM to answer. So for docs/API references, it’s interaction in the moment (you’re on the page + you ask), not the browser silently learning your click patterns over time.
What “trainable” means in Oasis (so expectations match)
Thumbs up/down + comments on assistant replies → product improvement (anonymous or account-linked; your choice). That’s not on-device fine-tuning of a personal model on your machine.
Optional “Personalize Oasis Assistant with my account” → ties assistant interaction logs to your account for better signed-in experience over time, still separate from uploading your whole browsing graph.
For your workflow specifically
Competitor store pages / GDC posts / SDK docs: semantic history + memory search are the wins — “surface that article again” without digging through 200 tabs.
Long doc sessions: summarize / ask-about-this-page while you’re reading.
Repeat visits to the same sites: you’ll see them more in search results because they’re in history, but we’re not (today) ranking purely on visit frequency like a dedicated “favorites brain.”
To be super transparent, we’re not yet a full “workflow OS” that learns how you research (e.g. always cross-reference App Annie → spreadsheet → Notion) and automates that sequence without you asking. We’re closer to local recall + assistant actions on tabs/bookmarks/history that compounds the more you use Oasis for that work. Though, we are very interested in expanding our system to handle more complex workflow sequences of any type.
If you decide to experiment, we'd love to know:
When you’d want a nudge (start of day, after closing many tabs, when landing on a store URL, never).
What the nudge should do (reopen group, search history, summarize, bookmark to a hub — not all of the above).
Whether it should be local-only (patterns never leave the machine) vs okay with account-linked suggestions when signed in.
One moment last week where proactive would have saved you time — and one moment where it would have annoyed you.
Thanks again for checking us out and supporting the launch! :)
A privacy-first browser you can train is a genuinely interesting model. Most teams use shared browser profiles or sync that leaks context everywhere. We've been building in the AI customer success for B2B SaaS space, and Oasis touches on something we think about a lot. How does the anonymous training actually work: is the model local only, or does any data leave the device?
Oasis Browser
@shivam_jaiswal21 Thanks for sharing your thoughts, Shivam! That shared-profile / sync-leaking-context problem is exactly the kind of thing we had in mind when we separated “your browser” from “how we improve the assistant.”
Short answer: the assistant is not local-only. When you use Oasis Assistant, prompts and replies go to cloud LLMs (routed via our backend/proxy — e.g. Gemini through Supabase Edge). That’s separate from “anonymous training,” which is about whether improvement data is tied to your account, not whether anything leaves your machine.
What stays on your device (by default)
Browsing history, bookmarks, and the local memory / semantic search index (including on-device history embeddings)
Saved passwords in Firefox’s encrypted Login Manager — not in assistant memory or training payloads
Day-to-day browsing profile — we’re not building a “shared browser profile” that syncs your full context to ads or a central browsing graph
What leaves the device when you use the assistant
Inference: Your message, the model’s reply, and whatever context the agent attaches (e.g. active tab URL/title, tool results, and sometimes page text if you use summarize / page-grounded tools) — that goes to the remote model so the assistant can answer.
Product improvement (optional logging): With “Personalize Oasis Assistant with my account” turned off (the default for new profiles), we still log assistant interactions to improve the product, but user_id is null — no email, no account block in the payload. “Anonymous” here means not linked to your identity in our DB, not “nothing leaves the device.”
“Anonymous training” specifically
That’s the thumbs up/down + comment flow on replies. You can submit training as anonymous (feedback_events without your user id) or personalized (linked to your account). That’s a second knob from the Privacy setting above — same words, different tables. Either way, you’re giving us signal on what worked; anonymous mode just keeps that row off your account.
Other references: https://www.producthunt.com/p/kahana/oasis-browser-technical-and-interaction-data
Docs: https://kahana.co/docs/technical-and-interaction-data
5-minute YouTube video explaining data that is sent: https://youtu.be/8C3FucA95Lg
What we’re not doing
We’re not claiming a fully on-device model for chat today (perhaps in the near future!)
We’re not claiming zero data when the assistant runs — cloud inference + (by default) anonymous interaction logs.
We’re not uploading your whole browsing history as a training corpus; improvement signal is assistant-shaped (prompts, responses, tab context when you’re in the assistant, tool traces).
For B2B CS workflows: Oasis is closer to “your machine holds the messy browser state; the assistant only sees what you invoke and what tools pull in for that turn” than to “train a shared synced profile.” If you need air-gapped / no cloud at all, we’re not there yet — we’d say that upfront.
Happy to go deeper on any layer (telemetry JSON shape, user-initiated training vs telemetry data from general llm usage, or what’s local for search vs chat).
Just curious, what other aspects are you and your team thinking about? I love this space and these conversations :)
Very cool product!
Can't wait to see more.
How is it about the use of resources? Is it not too heavy?
Oasis Browser
@fberrez1 Short answer. Yes, it's not too heavy from what we've seen so far!
Testing with 125 internal beta testers so far, we've received 1 mentions so far of Oasis slowing down performance on a device, but we haven't yet pinpointed whether that's an issue with the application itself or the device/user. We’re still early and tuning performance. If you try it and it feels heavy on your machine, we’d love to know your OS and whether the assistant was open. We’re optimizing for “powerful when you need it, quiet when you don’t.”
This version of Oasis is built on Firefox, so day-to-day browsing experience is in the same ballpark as Firefox.
A bit more detail:
Browsing & privacy: History, bookmarks, and semantic search indexes stay on device.
Assistant: Replies are powered by cloud models when you ask Oasis something, so you get capability without keeping a huge local LLM loaded all the time. Though, we have received questions already about supporting local models too
In practice: If the assistant is idle, resource use should feel close to Firefox. When you’re chatting, using voice, or running browser actions through the assistant, you’ll see the usual spikes (CPU/network) you’d expect from AI + automation.
What’s interesting to me here is that most AI products optimise for answers, while this feels like it’s optimising for continuity. Curious what behaviours surprised you most in beta users, are people treating Oasis more like a browser, a memory system, or almost a second brain?
Oasis Browser
@surabhi_minocha Hey Surabhi! We're still learning all the time right now. The behaviors are developing in real-time. I think we're seeing people enjoy treating it like "Jarvis" from the Iron Man movies. I found myself up late half-watching TV and snacking while I was doing some vibe coding. Then I needed to check on a deployment in Github. Rather than click around, I just realized I could say "Hey Oasis can you open Github workflow tab?"
And then it did it.
After that, I said "Can you summarize the results of the workflow?"
And it did that!
For me and others, we are finding that the ability to use natural language and voice to control the browser lets us stay focused more and reduce mental context switching and cognitive load.
I'm not sure what this would qualify as, I suppose a combination of all of it!