SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2407.04841 · ICML · 2024

Associative Recurrent Memory Transformer

Mikhail Burtsev, Aydar Bulatov, Yuri Kuratov, Ivan Rodkin, © Rodkin

arXiv · PDF · Open in the Atlas

Code that ran

4 of 14 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 1 has no contract verdict; 3 failures of our instrument; 10 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2407.04841.json · badge: /api/ran/2407.04841.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
RodkinIvan/associative-recurrent-memory-transformer canonical (classifier label)
classifier basis: census
4 ran · 10 no recorded run
FunctionStatusWhere it lives
allocate_mem_buff Ran
ran; our contract draft was wrong, not the code
RodkinIvan/associative-recurrent-memory-transformer/megatron/memory.py
code served (permissive licence) · get_code("db8deebff0d5256c")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned MemoryBuffer. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.distributed as dist

# Mock torch.distributed to avoid CUDA and rank checks
class MockDistributed:
    @staticmethod
    def get_rank():
        return 0

torch.distributed.get_rank = MockDistributed.get_rank

# Mock torch.cuda to avoid CUDA device access
class MockCuda:
    @staticmethod
    def current_device():
        return 'cpu'

torch.cuda.current_device = MockCuda.current_device

torch.manual_seed(0)

ARGS = {
    'name': 'test_buffer',
    'numel': 1024,
    'dtype': torch.float32,
    'track_usage': True
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type MemoryBuffer — a MemoryBuffer output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a MemoryBuffer output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
bias_dropout_add Ran
ran on a synthesized input
RodkinIvan/associative-recurrent-memory-transformer/megatron/model/transformer.py
code served (permissive licence) · get_code("b6bf9c8a7315bb78")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned Tensor, ndim 2, dtype family float, shape (2, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

ARGS = {
    "x": torch.randn(2, 4),
    "bias": torch.randn(2, 4),
    "residual": torch.randn(2, 4),
    "prob": 0.5,
    "training": True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
deberta_memory_layers_forward Ran
ran; our contract draft was wrong, not the code
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/encoder/memory_layers.py
code served (permissive licence) · get_code("5b2f7f0576ba5c71")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned BaseModelOutput. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
from collections.abc import Sequence
from transformers.modeling_outputs import BaseModelOutput

torch.manual_seed(0)

# Build mock layer module that returns hidden_states or (hidden_states, attentions)
class MockLayer(nn.Module):
    def __init__(self, hidden_size):
        super().__init__()
        self.hidden_size = hidden_size
    
    def forward(self, hidden_states, attention_mask=None, output_attentions=False, 
                query_states=None, relative_pos=None, rel_embeddings=None):
        # Return same shape as input or tuple if output_attentions
        if output_attentions:
            attentions = torch.zeros(hidden_states.shape[0], hidden_states.shape[1], hidden_states.shape[1])
            return hidden_states, attentions
        return hidden_states

# Build memory layers container
class MockMemoryLayers(nn.Module):
    def __init__(self, num_layers, hidden_size):
        super().__init__()
        self.layers = nn.ModuleList([MockLayer(hidden_size) for _ in range(num_layers)])
    
    def __getitem__(self, idx):
        return self.layers[idx]

# Build parent RMT object with memory_layers and memory_position
class MockRMTParent:
    def __init__(self, num_layers, hidden_size, memory_position=0):
        self.memory_layers = MockMemoryLayers(num_layers, hidden_size)
        self.memory_position = memory_position

# Build the self object (the DeBERTa model with memory layers)
class MockDeBERTaMemoryLayers(nn.Module):
    def __init__(self, num_layers, hidden_size, seq_len):
        super().__init__()
        self.hidden_size = hidden_size
        self.seq_len = seq_len
        self.layer = nn.ModuleList([MockLayer(hidden_size) for _ in range(num_layers)])
    
    def get_attention_mask(self, attention_mask):
        return attention_mask
    
    def get_rel_pos(self, hidden_states, query_states, relative_pos):
        return relative_pos
    
    def get_rel_embedding(self):
        return torch.randn(self.hidden_size, self.hidden_size)

# Create instances
batch_size = 2
seq_len = 3
hidden_size = 4
num_layers = 2
memory_position = 0

self_module = MockDeBERTaMemoryLayers(num_layers, hidden_size, seq_len)
rmt_parent = MockRMTParent(num_layers, hidden_size, memory_position)

# Build arguments
hidden_states = torch.randn(batch_size, seq_len, hidden_size)
attention_mask = torch.ones(batch_size, seq_len)
output_hidden_states = True
output_attentions = False
query_states = None
relative_pos = None
return_dict = True

ARGS = {
    'self': self_module,
    'hidden_states': hidden_states,
    'attention_mask': attention_mask,
    'output_hidden_states': output_hidden_states,
    'output_attentions': output_attentions,
    'query_states': query_states,
    'relative_pos': relative_pos,
    'return_dict': return_dict,
    'rmt_parent': rmt_parent,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type BaseModelOutput — a BaseModelOutput output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a BaseModelOutput output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_bias_dropout_add Ran
ran; our contract draft was wrong, not the code
RodkinIvan/associative-recurrent-memory-transformer/megatron/model/transformer.py
code served (permissive licence) · get_code("dc55ddfafd78b31b")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned function. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# The function get_bias_dropout_add takes one parameter: training (a boolean)
# It returns a closure that uses this training value
# The returned closure will be used with x, bias, residual, prob arguments

# Create small tensors for testing the returned function
batch_size = 2
dim = 4

x = torch.randn(batch_size, dim)
bias = torch.randn(batch_size, dim)
residual = torch.randn(batch_size, dim)
prob = 0.5

# training is a boolean parameter for get_bias_dropout_add
training = True

ARGS = {
    "training": training
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
deberta_horizontal_memory_forward Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/encoder/horizontal_memory.py
code served (permissive licence) · get_code("55583aac4ec58bf6")
How it ran

no invocation recorded.

generate_pairs Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/run_finetuning_associative_retrieval.py
code served (permissive licence) · get_code("79e90a746251736b")
How it ran

no invocation recorded.

get_mem_buff Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/megatron/memory.py
code served (permissive licence) · get_code("8a1aaa5bb993da97")
How it ran

no invocation recorded.

horizontal_memory_forward Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/encoder/horizontal_memory.py
code served (permissive licence) · get_code("926c84c930441c95")
How it ran

no invocation recorded.

horizontal_memory_forward Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/encoder_decoder/horizontal_memory.py
code served (permissive licence) · get_code("a4bc8265f86d923d")
How it ran

no invocation recorded.

horizontal_memory_forward_1by1 Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/encoder_decoder/horizontal_memory.py
code served (permissive licence) · get_code("6ba5053e38b057df")
How it ran

no invocation recorded.

memory_layers_forward Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/decoder/memory_layers.py
code served (permissive licence) · get_code("431671633a1530c3")
How it ran

no invocation recorded.

memory_layers_forward Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/encoder/memory_layers.py
code served (permissive licence) · get_code("424d63b2052c4ec0")
How it ran

no invocation recorded.

memory_layers_forward Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/encoder_decoder/memory_layers.py
code served (permissive licence) · get_code("219e86a9d74b05ec")
How it ran

no invocation recorded.

memory_layers_forward_r2r Not yet run
unverified (harvested, no recorded run)
RodkinIvan/associative-recurrent-memory-transformer/rmt_utils/decoder/memory_layers.py
code served (permissive licence) · get_code("e4fc0dbd771c70e4")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 distinct functions with an identity, 2 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 0 the same month, 0 later) and 12 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
allocate_mem_buff 2 other papers
2 dated earlier
2108.07435 (2021-08)
Modeling Protein Using Large-scale Pretrain Language Model
2404.03608 (2024-04)
Sailor: Open Language Models for South-East Asia
get_mem_buff 2 other papers
2 dated earlier
2108.07435 (2021-08)
Modeling Protein Using Large-scale Pretrain Language Model
2404.03608 (2024-04)
Sailor: Open Language Models for South-East Asia

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

This paper addresses the challenge of creating a neural architecture for very long sequences that requires constant time for processing new information at each time step. Our approach, Associative Recurrent Memory Transformer (ARMT), is based on transformer self-attention for local context and segment-level recurrence for storage of task specific information distributed over a long context. We demonstrate that ARMT outperfors existing alternatives in associative retrieval tasks and sets a new performance record in the recent BABILong multi-task long-context benchmark by answering single-fact questions over 50 million tokens with an accuracy of 79.9%. The source code for training and evaluation is available on github.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2407.04841")
get_code_for_paper("2407.04841")
have("2407.04841")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2407.04841.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2407.04841.svg)](https://syntology.ai/paper/2407.04841)

Connect an agent — have() is free.