Transcription was the easy part.
A compact statement of what Sonal Labs is working on and why. Memory is the product; overlap is the research bet; trust is the thing we're measuring. The rest follows.
Your brain forgets most of this day.
Not because you're careless. Because you're human. Ebbinghaus' forgetting curve, first published in 1885 and replicated by Murre & Dros in 2015, is still the cleanest description of how quickly context vanishes without reinforcement.
50%
Half of what you hear today is gone within the hour.
70%
By tomorrow, most of today's context has already vanished.
90%
Within a month, only a fraction of the original memory remains.
Every system before Sonal
relies on you.
Physical records
Clay tablets, parchment, notebooks, pen & paper. Heavy, fragile, easy to lose.
Electronic records
Phonographs, tape, digital recorders. More persistent — but every minute of recording costs a minute of replay.
Apps and AI note-takers
Transcribe meetings and scheduled calls. Limited to structured sessions; the rest of your life stays undocumented.
Each generation offloaded more of the work. None of them remove the requirement that you decide in advancewhat to remember. Sonal is the first layer built to catch the conversation you didn't know mattered until it did.
Everyday conversations vanish.
The thing your partner asked you to pick up.
The follow-up you meant to send.
The commitment your client definitely made.
The brilliant idea someone mentioned in passing.
Overlap is the trust bottleneck.
If we wanted to automate memory, we could have done it a decade ago — by writing more notes. The actually hard part turned out to be trust. Modern foundation models have commoditised transcription. You can get words out of audio for essentially free. What you cannot get for free is an answer to the question that actually matters: who said it.
Overlap is where that question becomes unanswerable. Two people talk at once; conventional pipelines make a forced choice; the transcript that comes out looks fine. It isn't. It's wrong in the specific way that assigns commitments to the wrong person — and every downstream action inherits the mistake.
A memory system that occasionally gets words wrong is imperfect. A memory system that misattributes is adversarial. That's why we start where we start.
- DR. PATEL
- Any chest pain in the last week —
- PATIENT
- — yes, on Tuesday, but —
- DR. PATEL
- — and any shortness of breath?
- PATIENT
- — it went away quickly.
“Patient denies chest pain.”
A cardiac symptom goes unrecorded.
Attribution → memory → action.
Attribution
Presence AI — our overlap-native model — resolves who said what, frame by frame, audio first with optional visual context.
Inside Presence AI →
Memory
Sonal OS turns attributed audio into structured, evidence-grounded memory objects — searchable, editable, revocable.
See the system →
Action
Reports, reminders, follow-up drafts, API-level workflow triggers. Downstream automation you can trust because the record can be trusted.
Applied programmes →
“Overlap is not an edge case in natural conversation. It is where conversation happens.”