Launching today
MeMex gives your Mac a private visual memory. It turns screen activity into an organized timeline of real work moments, then lets you ask plain-English questions and return to matching screen evidence. Daily and weekly reports summarize work you already did, while Pattern Insights spots repeated workflows worth automating. Recordings stay local in a location you control; AI analysis is optional and uses your own provider or compatible local endpoint.
the "recordings stay local, AI analysis is optional and uses your own provider" split is the interesting design decision here. if someone plugs in a remote provider (say a hosted API instead of a local endpoint) to ask a plain-english question, does the relevant screen evidence get sent up to that provider to answer it, or does it stay local and only the extracted text/metadata leaves the machine? for a tool whose whole pitch is private visual memory of everything you do, that's the detail that actually determines whether it's private by default or private-if-you-configure-it-carefully.
@galdayan Thanks ! Let me explain this. The current behavior is:
Raw recordings remain on the Mac and are never uploaded as complete video files.
If a remote AI provider is enabled, MeMex sends selected sampled frames directly to that configured endpoint when generating analyzed timeline cards. This only happens after the user reviews and accepts the data-sharing disclosure.
For a plain-English question, retrieval happens locally. The Q&A request sends the question plus text from the matched timeline cards, such as titles, summaries, observations, timestamps, and app/site context. It does not send the underlying screenshots or video again for that question.
Requests go directly from the Mac to the user’s chosen provider. MeMex does not proxy or store them on a MeMex backend.
So with a hosted provider, the accurate description is local storage with optional remote AI processing, not end-to-end local processing. Users who want the entire pipeline to remain local can connect a local OpenAI-compatible endpoint, and recording/timeline browsing still work without enabling AI at all.
This is a useful callout. Raw recordings stay local. AI processing runs through the model endpoint you choose.
This looks genuinely useful. I'm constantly trying to find a tab, file, or random thing i saw earlier and can't remember where it was.
Also really nice that everything stays local. Doessearch still feel fast once you've built up a few months of history?
@rudy_zhou Thanks Rudy! Search runs over locally indexed timeline cards instead of rescanning raw recordings, so it remains fast even you have lots of history datas.
One privacy clarification: raw recordings stay on the Mac. If a hosted AI provider is enabled, selected frames may be sent directly to that provider for analysis. A local endpoint keeps the entire AI pipeline local.
Ah got it, thanks for clarifying!
Having the option to keep the whole pipline local definitely makes this more interesting!
How does the local storage hold up over time, like does it balloon fast on a normal workday and need manual cleanup or does it handle that itself?
@abantakanancyy Good question. It doesn’t grow without bounds by default.
MeMex has separate configurable storage caps for raw recordings and processed timeline media, both set to 10 GB by default. It checks storage on launch and every hour, then automatically removes the oldest media when a limit is exceeded. Recent raw recordings are protected for three days, and the searchable timeline-card text remains available after older media is removed.
You can view usage in Settings and set each limit from 1–20 GB, or choose unlimited. So routine use shouldn’t require manual cleanup, although daily growth naturally depends on recording time and display resolution.
How does the search actually work under the hood, does it index OCR text locally or rely on screenshots only when you query?
Great question. Search is local and text-first; MeMex does not reopen or upload screenshots when you ask a question. During analysis, sampled frames are converted into structured timeline cards with titles, summaries, detailed observations, app/site context, and timestamps, and those fields are stored in the local SQLite library. When you ask something, MeMex parses the terms and time context locally, ranks matching cards and observations locally, then sends only the top matched text and metadata to the configured model to compose the answer. The current build does not maintain a separate full OCR index. Visual analysis happens when cards are generated, not at query time. If the configured provider is hosted, selected frames may be sent to it during that earlier analysis step; the Q&A request itself does not include screenshots or video.
Comes in handy when I can't remember which file I had open two hours ago, just ask it in plain English and it pulls up the right window. Love that the recordings stay local by default.
@urgenzeki80043 Thanks “Which file or window did I have open earlier?” is one of the everyday memory gaps I wanted MeMex to solve. The recordings stay on your Mac by default, while AI features remain optional and user-configured.
tried it for an afternoon and the timeline view feels genuinely useful, scrolling back to find that one design file i lost track of. like that the recordings stay on my machine by default.
Thanks Sinem! That “one design file I lost track of” moment is exactly the kind of context gap MeMex is meant to close. I wanted the timeline to feel like continuity, not another folder full of screenshots. I’m glad the local recording model felt right too. When you found the file, did you get there by scrolling the timeline or by asking the Assistant? That distinction would help me decide which retrieval path to improve first.