ReMyndIf you are the sort of engineer who checks what a binary talks to before granting it screen-recording permission, this page is written for you. ReMynd is a local AI memory for your Mac that captures what you see, makes it searchable, and answers questions from it — pointed at whatever model you choose to run.
You are also, statistically, the densest concentration of people this is built for. Per the Anthropic Economic Index, Computer and Mathematical tasks account for 23.8% of global AI usage — the largest single category of AI work.
Bring your own model
The default assumption of most AI tools is that inference happens on someone else's hardware. ReMynd does not assume that.
With LM Studio. Install LM Studio, pull a model in the 7B–30B range, and start its local server (it exposes an OpenAI-compatible endpoint on localhost by default). In ReMynd, open settings, choose the local provider, and point it at that endpoint. From then on, retrieval and answering both happen on your machine — the model reads slices of your own history and responds, and nothing about the query leaves the laptop.
Why a mid-sized local model is enough here. The job is not open-ended reasoning from scratch. It is reading retrieved context and answering from it, which is a much easier task and one where local models punch well above their weight. This is the specific case where the private option is not the worse option.
When to use a cloud model anyway. Harder synthesis across weeks of history benefits from a frontier model. That should be an explicit switch you flip, not a default buried in the architecture — so it is.
What it is actually useful for
Retracing a debugging session. Three days later, you need the stack trace, the config value you tried, and the GitHub issue that turned out to be the answer. All of it was on screen; none of it was saved. Ask for it.
Finding which doc or tab answered the question. Not "search my notes" — you never took notes. The API doc with the rate-limit table, the Slack thread where someone explained the deploy order, the pricing page you compared six weeks ago.
Code review context. What did this module look like before the refactor, what did the author say about it in the call, what was the ticket actually asking for. Three sources, one timeline.
Reconstructing your own week. Standups, retros, and the honest version of where the time went, without manual logging.
The architecture, in one screen
Capture. The recorder runs as a separate process from the UI. It captures the focused window rather than everything on every display, and skips frames when nothing has changed.
OCR. Apple's Vision framework, on the Neural Engine, on your machine. This is the step people assume must be a server call; on a Mac it is not.
Index. Local SQLite. Full-text search for exact matches — an API key, an error string, a hostname — plus vector embeddings for meaning-based retrieval.
Inference. Local via LM Studio, or a cloud model when you pick one.
Storage. Your disk. Optional encrypted cloud backup, exportable to an S3 bucket you own, with your keys. The claim we are willing to make in public is the narrow, checkable one: your captured memory is never in our cloud.
Agent access. Read-only, scoped through typed tools rather than arbitrary filesystem access.
The part most pages leave out
Any real Mac app makes network calls. ReMynd signs you in and reports crashes, like everything else you have installed. That is not your memory, and you should hear it from us rather than find it in Little Snitch.
What we will not tell you is that nothing ever leaves your machine. That claim is false the moment you count authentication, and an app that makes it either has not thought carefully about the boundary or is hoping you will not check.
If the architecture is what you care about, the longer version is in private AI on your Mac. If you are comparing against the open-source option, ReMynd vs Screenpipe is deliberately generous to Screenpipe.