PRIVATE BETA · ANDROID
MementosCODENAME · PUBLIC NAME TO BE ANNOUNCED
Your voice memos, searchable. Nothing ever leaves your phone.
Record in English or Hinglish, ask a hazy question later, and hear the answer in your own voice. Transcription and search run on the phone itself.
Android · test build 4 in progress
Hold the orb and speak — in English, Hinglish, or both.
WHAT IT DOES
A memory you can ask.
- 01
Record the way you talk
Hold the orb and speak, in English, in Hinglish, or switching between the two mid-sentence.
- 02
Private by design
Transcription and search run on the phone. The app has no internet permission, and audio and text are stored encrypted.
- 03
Ask the way you remember
“What was that 15% thing on Tuesday?” Hazy questions, typed or spoken, find the right memo.
- 04
Hear it in your own voice
The answer is your own recording, played from the right moment, with the text beside it. It is exactly what you said.
HOW IT UNDERSTANDS HINGLISH
Two speech models, one memo.
People in India switch languages mid-sentence. No single on-device model handled both well, so Mementos cuts each memo at pauses and sends every piece to the model that speaks its language.
Read the full write-upRouting a memo that switches language
One memo, about 50 seconds, cut at pauses into pieces of up to 12 seconds.
- English
- Hinglish
- wrong model
Parakeet— English
Hinglish-Swift— Hinglish
Step 4 of 4: ResultRouting piece by piece, instead of letting the first 10 seconds decide, cut word errors on mixed-language clips from 35.7% to 18.2%.
MEASURED, NOT PROMISED
Numbers from our test builds.
Measured on a Redmi Note 14 Pro+ (a mid-range Android phone) unless marked laptop. Small test sets; the full method is in our research write-up.
- Right memo first
- 9 of 10 test questions
- test build 1
- Right memo in the top three
- 10 of 10
- test build 1
- Time per search
- 5–13 ms
- test build 1
- Sound starts after tapping a result
- about 1 s
- test build 1
- A 55-second memo ready to search
- 12–14 s
- test build 3
- Word errors on mixed English–Hinglish clips
- 35.7% → 18.2%
- laptop CPU, before build 3
- Peak memory, English / mixed memo
- 1.30 GB / 1.66 GB
- test build 3
BUILD BY BUILD
Every build is logged.
Each test build gets a changelog entry, a git tag and measured results — including what we tried and dropped.
- 9 OCT 2026
Test build 1
The first build: record a memo, have it transcribed on the phone, ask a question, and hear the answer in your own voice.
- 10 OCT 2026
Test build 2
One fix: the memo list no longer crashes when a memo has no speech in it.
- 10 OCT 2026
Test build 3
Memos that mix English and Hinglish are transcribed piece by piece, each by the right model. A security hole found in testing was closed the same evening.
- IN PROGRESS
Test build 4
Planned: answers that play the sentence before and after the match, a quick name check after recording, tap-to-fix for wrong words, sound-alike search for misheard names, and tags.
What didn't make it: why a small language model lost to our rules.
Can I download Mementos?
Not yet. Mementos is in a private beta on Android, and test build 4 is in progress. The public name is not decided yet; Mementos is the codename. To hear when it launches, email contact@bodhamlabs.com with the subject “Mementos”.
Does Mementos upload my recordings?
No. The app has no internet permission at all, so it cannot send audio, transcripts or personal data anywhere. Recordings and text are encrypted on the phone.
Which languages does it understand?
English, and Hindi or Hinglish (Hindi mixed with English). A memo can switch between them: the app cuts it into pieces at pauses and sends each piece to the right speech model.
Which phones does it run on?
The beta needs Android 10 or newer on a 64-bit phone, about 1.5 GB of free storage, and works best with 8 GB of RAM. Our test phone is a Redmi Note 14 Pro+.
Why is the app so large?
The speech models are inside the app, so nothing has to be downloaded and nothing goes to a server. The first test build was 960 MB.