anla — Agent-Native Lossless Archive
An AI may plan how to pack. A public, deterministic, model-independent decoder must recover every byte that was declared into the archive.
Extract(Pack(F, P)) = F
This is the reference implementation: dependency-light, readable, and the executable half of the specification. A second implementation in Rust lives in the same repository and produces byte-identical archives; the two are held to each other by a conformance suite and a differential fuzzer.
- Site and browser workbench — https://anla.evemisslab.com
- Specification — https://anla.evemisslab.com/spec/
- Source, papers, conformance vectors — https://github.com/kakon77777-commits/anla
Install
pip install anla-archive # reader, writer, verifier, both CLIs
pip install "anla-archive[speed,zstd]" # the BLAKE3 and Zstandard fast paths
The distribution is anla-archive; the import packages and the commands are anla
and anla1. PyPI refuses the bare name — it is one character from several existing
projects — and the two names are independent anyway.
import anla, anla1 # after pip install anla-archive
Python 3.10+. No required dependencies: blake3 and zstandard are optional, and
the package carries a readable pure-Python BLAKE3 so the format's hash can be checked
by reading rather than by trusting a compiled wheel.
Two commands, two profiles
anla pack ./project -o project.anla # ANLA-MVP v0.1 — frozen
anla1 pack ./project -o project.anla # ANLA 1.0 — draft
Separate commands rather than a flag, for the same reason 1.0 has its own magic number: they are different formats, and one command switching between them on a flag invites an archive written under one profile and read under the other.
anla1 also carries the context layer — an agent's own history, stored losslessly
and addressable back to the exact bytes of a turn:
anla1 context capture memory.anla
anla1 context segment memory.anla --scheme changepoint-v1
anla1 context address memory.anla "how was the gear table produced"
Read the format from Python
from anla1.snapshot import list_snapshots, extract_snapshot, verify_archive
data = open("project.anla", "rb").read()
report = verify_archive(data) # every chunk, every root
files = extract_snapshot(data, list_snapshots(data)[-1])
Speed, stated rather than discovered
Content-defined chunking is the default, because fixed chunking destroys deduplication. In this writer it is also the slow path, by a wide margin — measured on 64 MiB of incompressible data:
| MiB/s | |
|---|---|
| this writer, fixed chunking | 64.4 |
this writer, anla-cdc-1 |
3.8 |
the Rust writer, anla-cdc-1 |
107.3 |
| this reader, verify | 538.1 |
So: use this package to read archives, to check the format, to embed in Python, and
for trees measured in megabytes. Use anla1-rs from the same repository for volume.
The two produce identical bytes; only one of them is worth an afternoon.
Every figure above is produced by bench/run_bench.py and republished at
https://anla.evemisslab.com/bench/, including the rows where ANLA loses.
Status
ANLA-MVP v0.1 is frozen. ANLA 1.0 is a draft and says so: the freeze rule —
two independent implementations producing byte-identical archives with no verdict
divergence under differential fuzzing — is met, but two implementations by one author
are weaker evidence than two by two authors, and the specification records that
rather than rounding it up.
This is a research profile. It is tested, but it is young. Do not make an ANLA archive the only copy of anything you cannot lose.
Apache-2.0.
Metadata
Release files for anla-archive 0.1.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| anla_archive-0.1.0.tar.gz | 203.5 kB | Without provenance | Details for anla_archive-0.1.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| anla_archive-0.1.0-py3-none-any.whl | Python 3 | none | any | Details for anla_archive-0.1.0-py3-none-any.whl |
Total release size: 331.3 kB
Release files / anla_archive-0.1.0.tar.gz
| Download URL | anla_archive-0.1.0.tar.gz |
|---|---|
| Size | 203.5 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
5188acc4b19c11d1a194a24934d0dfbdac83317c82d44736e27f4cfed6875181
|
|
BLAKE2b-256 checksum How to use checksums |
d96f2246a8e6281ca0aba44acad7c1c9f2ff688772836bb09604ff0498c158d9
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.14.5
|
Release files / anla_archive-0.1.0-py3-none-any.whl
| Download URL | anla_archive-0.1.0-py3-none-any.whl |
|---|---|
| Size | 127.7 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
8b3bae343ff35ba2d47bab26d6863aa4e296e6538aa9e25d3aaab7f62a6ff234
|
|
BLAKE2b-256 checksum How to use checksums |
21f5ece4a3be010a6153661647a9f8c7c5d09ea37ee89859d68f10e4bd25230c
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
twine/7.0.0 CPython/3.14.5
|
Security for anla-archive 0.1.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.