helix/CITATION.cff at main · phimemory/helix · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
cff-version: 1.2.0
message: "If you use Helix in your research, please cite it as below."
type: software
title: "Helix: Harmonic Encoding with Lossless Infinite conteXt"
abstract: >
  Helix is a neural memory architecture that stores information as rotation
  angles on a geometric manifold rather than decaying floating-point numbers.
  The core insight is that rotation is a unitary (magnitude-preserving)
  operation, unlike the contractive multiplication used in GRUs and LSTMs.
  This enables theoretically infinite lossless memory in a single neuron.
  Built on top of Helix is a Crystal Suite — eleven production-ready modules
  for portable, encrypted, federated, and multi-modal AI memory that is
  fully model-agnostic (any LLM can read any crystal without retraining).
authors:
  - family-names: Kalyan
    given-names: Pavan
    alias: Cintu07
repository-code: "https://github.com/Cintu07/helix"
license: "BUSL-1.1"
version: "2.1.0"
date-released: "2026-04-25"
keywords:
  - neural-architecture
  - memory
  - phase-rotation
  - unitary-isometry
  - lossless-compression
  - model-agnostic
  - ai-infrastructure
  - persistent-memory
  - phase-geometry
  - harmonic-encoding
references:
  - type: article
    title: "Unitary Evolution Recurrent Neural Networks"
    authors:
      - family-names: Arjovsky
        given-names: Martin
      - family-names: Shah
        given-names: Amar
      - family-names: Bengio
        given-names: Yoshua
    year: 2016
    url: "https://arxiv.org/abs/1511.06464"
    notes: "Foundational work on unitary recurrent networks that preserve information via unitary transformations."
  - type: article
    title: "On the difficulty of training recurrent neural networks"
    authors:
      - family-names: Pascanu
        given-names: Razvan
      - family-names: Mikolov
        given-names: Tomas
      - family-names: Bengio
        given-names: Yoshua
    year: 2013
    url: "https://arxiv.org/abs/1211.5063"
    notes: "The vanishing and exploding gradient problem that Helix solves by replacing multiplication with rotation."
  - type: article
    title: "Expressive power of recurrent neural networks"
    authors:
      - family-names: Chen
        given-names: Minshuo
      - family-names: Liu
        given-names: Bohang
    year: 2018
    url: "https://arxiv.org/abs/1711.00811"
    notes: "Theoretical bounds on RNN expressiveness that motivate the phase-angle encoding approach."
  - type: article
    title: "Landauer's Principle"
    authors:
      - family-names: Landauer
        given-names: Rolf
    year: 1961
    journal: "IBM Journal of Research and Development"
    notes: "The physical principle that information erasure has a thermodynamic cost, motivating Helix's reversible computation design."