A memory-hard code the size of a SeedQR: one small secret under a passphrase, with Argon2id.
Krux's KEF does the same job with PBKDF2 on microcontrollers, whose few megabytes of memory leave no room for a memory-hard function, and it is a sound choice there. Where a device has the memory, standard practice is a memory-hard function, with Argon2id first among them (RFC 9106; OWASP). AQRV brings that to a code of the same size for computers and Raspberry Pi: a next step for QR vaults on such devices, beside KEF and not in its place.
It is an open proposal seeking outside review. It is not in Faraday's app, has not been reviewed and is not a standard. It is for testing only.
A person can print or hand-copy a secret's encrypted form and keep it where a finder could see it. A finder holds nothing until they guess the passphrase, and under Argon2id each guess costs memory as well as time. Krux's KEF format does this in a similar space under PBKDF2, which needs no memory; AQRV is the same idea under a memory-hard function.
Code
Envelope
What it holds
25 × 25
32 bytes, carried as text.
A 12-word seed, or up to 15 bytes of text
29 × 29
53 bytes
Up to 32 bytes: a seed of up to 24 words, a codex32 string up to 192 bits, a passphrase or note, or a raw key
33 × 33
62 bytes
Up to 37 bytes: the same kinds, with a longer tag and room for a codex32 string up to 256 bits
What it holds, and what it is not
What it holds
A BIP-39 seed's entropy, as 12, 15, 18, 21 or 24 words.
A codex32 string.
A BIP-39 passphrase.
A short note.
A raw 32-byte key.
The kind is not stored, so a code does not show which it holds, within a size. Not every kind fits every size.
What it is not
Not in Faraday's app.
Not reviewed. There is a Rust implementation by the same authors (a library, a command line and a desktop test app); no outside one.
Not a standard.
Not a way to make a weak passphrase safe: a photographed code can be attacked offline.
Not for devices without 64 MiB of memory.
How it compares
Against Krux's KEF.
AQRV
Krux's KEF
Slow to guess
Yes: Argon2id, 64 MiB to 1 GiB
No memory cost: PBKDF2
Random salt
Yes: new for each code
No: the id, by default the wallet's fingerprint
Nothing names the wallet
Yes
No: the fingerprint is in the clear by default
KEF runs on microcontrollers and is in use, which AQRV is not. The full comparison is in AQRV.md, section 14. The vault page has all the forms side by side.
How long a passphrase holds
The time to try every passphrase of EFF's long word list, half that on average, for an attacker the size of a large data centre: a million CPU cores, about 16,000 servers of 64 cores, each core filling about 1 GiB of memory a second for Argon2, or doing about 225 SHA-256 compressions a second for KEF at its default 100,000 iterations. Holding the memory takes 64 TiB at 64 MiB and 1 PiB at 1 GiB. The times scale with the attacker: one with 1,000 cores, about 16 such servers, takes a thousand times as long, and one desktop computer of 16 cores about 65,000 times as long. A figure for scale, for today's hardware.
Passphrase
Bits
AQRV, 64 MiB
AQRV, 256 MiB
AQRV, 1 GiB
Krux's KEF
3 words
38.8
1 day
4 days
16 days
45 minutes
4 words
51.7
21 years
83 years
331 years
241 days
5 words
64.6
160,000 years
640,000 years
2.6 million years
5,100 years
6 words
77.5
1.3 billion years
5 billion years
20 billion years
40 million years
KEF's times count its compute only: it needs almost no memory, so graphics cards and dedicated hardware make it cheaper still, where Argon2's memory resists them. A passphrase a person makes up can be far weaker than any row. The figures and how they are worked out are in AQRV.md, section 12.
Downloads
A desktop test app and a command line, for Linux and Windows on x86-64, in one download for each. They are a test release: do not put a real secret in them.
Desktop test app and command line
The desktop app makes a code from a sample or data you type, opens a code or a .aqrv file with its passphrase, and shows what each memory setting costs. The command line, aqrv, does the same from a terminal: make, open and inspect a code, run the test vectors, and remake a code from a printed sheet's text.
Unpack the zip before running either program. Windows shows "Windows protected your PC" the first time, because the programs are not code-signed. Check the zip against SHA256SUMS, then choose More info, Run anyway.
Tested on 2026-11-11: the Rust library and a second reader by the same authors agreed on 13,413 generated and corrupted codes, and five fuzz targets ran 600 seconds each with no crash. Both readers are the same authors', so this is not an outside check. The results are in the vectors' README.