Ask your own notes. Nothing leaves your machine.
Years of session notes, supervision records and reading — searchable at last. Every answer is checked against your own material before you see it, and you're told when it doesn't hold up.
What it does that a general chat tool doesn't
Your whole archive, not one document at a time
Local AI tools let you attach a file to a conversation and ask about it. Reverim indexes a folder you already have — Word documents, PDFs, plain text, an existing Obsidian vault — and answers across all of it. It keeps itself up to date as you write, so a note you made this morning is searchable this afternoon without pressing anything.
Answers are checked, not just cited
Any tool can show you a citation. Reverim grades each answer against the passages it actually drew on, before you see it — flagging an invented source, wording that has drifted from your notes, or an answer that should have said "that isn't in here". You can open the check and read what it found.
It compares wording against the retrieved passages, so it catches fabrication and drift. It won't catch a plausible misreading of a passage it quotes correctly. The citations are there so you can check the ones that matter.
It stays descriptive
It locates, groups, summarises and drafts from what you have already written. It will not diagnose, advise on management, or produce a formulation of its own, and it will report an interpretation you recorded as yours. A second check watches for an answer sliding into interpretation or advice and makes it rewrite — the boundary is enforced, not merely requested.
Privacy
- No cloud. The model runs on your computer. Sending anything outward is off by default and takes a deliberate act to switch on.
- No account, no telemetry. Nothing to sign up for, no usage data, no crash reporting, no update pings.
- Your notes stay yours. Plain files in a folder you chose, readable without this app, and never rewritten unless you ask.
- Every action logged, locally. An activity log on your own machine shows what the app did and when.
Questions
Does anything I write leave my computer?
No. The language model runs locally, the search index is a folder on your disk, and the app has no account and sends no usage data. There is an optional setting that would allow a cloud model to receive a small number of retrieved passages; it ships off, no cloud provider is configured, and turning it on takes a deliberate act. Until you do that, there is no code path by which note content goes anywhere.
What kinds of file can it read?
Markdown, plain text, Word documents (.docx) and PDFs. Text is
extracted at index time and the original file is never modified or converted.
Scanned documents are the exception: an image of a letter contains no text to extract. Those are skipped and reported by name, so you know what wasn't read rather than finding a silent gap later.
What do I need to run it?
Windows 10 or 11, 64-bit, and about 6 GB of free disk space once the model files are installed. It runs on the processor — no graphics card is required. 8 GB of RAM works; 16 GB is more comfortable if you keep other things open.
Answers take roughly a minute on the built-in model, and longer for drafts. That's the cost of everything running on your own machine rather than in a data centre.
Do I have to reorganise my notes first?
No. Point it at a folder you already keep and it reads what's there. An existing Obsidian vault works as-is, including frontmatter, tags and wiki-links. There is no import step and no proprietary format — if you stop using Reverim tomorrow, your notes are exactly the files they were.
Can it change or delete my notes?
Writing to your vault is switched off by default. With it off, the writing tools aren't offered to the model at all — it cannot call what it cannot see.
There is no delete function anywhere in the app. Retiring a note means moving it, which you can undo. The one thing the app can delete is its own saved conversations, because those are the only records it creates rather than reads.
Will it give me clinical advice?
No, and that is deliberate rather than a limitation. It will not diagnose, advise on treatment or management, or produce a formulation of its own. It will report an interpretation you recorded, attributed to you.
A check on every answer looks for it sliding into interpretation or advice and makes it rewrite. Reverim is not a medical device and does not provide clinical decision support.
What happens when the answer isn't in my notes?
It says so, rather than filling the gap from the model's general training — and that is treated as a correct answer, not a failure. It will usually name what your notes do cover nearby, and suggest wording that might find it.
Why does Windows warn me about the download?
Because this build isn't code-signed yet. SmartScreen warns about any installer it hasn't seen signed by a known publisher, which at this stage includes ours. Choose More info → Run anyway.
If you'd rather verify first, the SHA-256 of the installer is published at the top of this page. That matters more than usual for an unsigned build, where the operating system won't vouch for it on your behalf.
Why are there two downloads?
The application is 118 MB. The language model it runs on is another 2.46 GiB, which is more than a Windows installer can carry in one file. So the models come separately and install from a button inside the app — you don't need to find a folder or unzip anything.
Is there a Mac version?
Not yet offered. A macOS build exists in the release pipeline but has not been tested on real hardware, and shipping an untested installer to clinicians isn't something we're prepared to do. If you want one, say so — it moves up the list.
What happens if my licence lapses?
Reverim becomes read-only. Your vault still opens, search still works, questions still get answered — only writing to your notes is withheld. You are never locked out of your own records by a licence check, because they are plain files on your disk and were yours before this app existed.
Does it work offline?
Yes. After installation nothing requires a connection — not the model, not the search index, not the licence, which is verified on your machine rather than against a server. You can disconnect entirely and it behaves identically.
Honestly, where it is
- Early. It works and it's in daily use, but it is a young product and it is not code-signed yet.
- Windows for now. A macOS build exists in the pipeline but has not been tested on real hardware, so it isn't offered here.
- No MSI. It installs per-user from an .exe. Organisations that deploy only via MSI can't yet.
- Scanned documents are skipped. An image of a letter has no text to extract; those are reported by name rather than quietly indexed as empty.