SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.01990 · 2026

According to Me: Long-Term Personalized Referential Memory QA

Margaret Li, Guangyu Yang, Jinghong Chen, Jingbiao Mei, Bill Byrne, Xinyu Hou

arXiv · PDF · Open in the Atlas

Code that ran

15 of 18 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 0 have no contract verdict; 15 failures of our instrument; 3 no recorded run), measured 2026-11-22; 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/2603.01990.json · badge: /api/ran/2603.01990.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
JingbiaoMei/ATM-Bench canonical (classifier label)
classifier basis: census
15 ran · 3 no recorded run
FunctionStatusWhere it lives
extract_json Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/memqa/qa_agent_baselines/A-Mem/memory_layer.py
code served (permissive licence) · get_code("1ba9020de9ace83e")
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 dict. 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 json
from typing import Any, Dict, List, Optional, Tuple, Union

def extract_json(payload: str) -> Optional[Dict[str, Any]]:
    cleaned = payload.strip()
    if cleaned.startswith('```'):
        cleaned = cleaned.strip('`')
        if cleaned.lower().startswith('json'):
            cleaned = cleaned[4:]
    start = cleaned.find('{')
    end = cleaned.rfind('}')
    if start == -1 or end == -1 or end <= start:
        return None
    fragment = cleaned[start:end + 1]
    try:
        return json.loads(fragment)
    except Exception:
        return None


# Setup code to construct valid arguments
ARGS = {
    "payload": '```json\n{"key": "value", "number": 42}\n```'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
extract_json_blob Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/extract_answer.py
code served (permissive licence) · get_code("b2d758e135419723")
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 dict. 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 json
from typing import Any, Dict, Optional

def _find_first_json_object(text: str) -> Optional[str]:
    in_string = False
    escape = False
    depth = 0
    start_idx: Optional[int] = None
    for idx, ch in enumerate(text):
        if start_idx is None:
            if ch == '{':
                start_idx = idx
                depth = 1
            continue
        if in_string:
            if escape:
                escape = False
                continue
            if ch == '\\':
                escape = True
                continue
            if ch == '"':
                in_string = False
            continue
        if ch == '"':
            in_string = True
            continue
        if ch == '{':
            depth += 1
            continue
        if ch == '}':
            depth -= 1
            if depth == 0:
                return text[start_idx:idx + 1]
    return None

def extract_json_blob(text: str) -> Dict[str, Any]:
    text = text.strip()
    if not text:
        raise ValueError('empty text')
    if text.startswith('```'):
        lines = text.splitlines()
        if len(lines) >= 2 and lines[-1].strip().startswith('```'):
            text = '\n'.join(lines[1:-1]).strip()
    try:
        parsed = json.loads(text)
        if isinstance(parsed, dict):
            return parsed
    except json.JSONDecodeError:
        pass
    candidate = _find_first_json_object(text)
    if candidate:
        parsed = json.loads(candidate)
        if isinstance(parsed, dict):
            return parsed
    raise ValueError('could not parse JSON object from text')


# Test case 1: Simple JSON object
simple_json = '{"key": "value", "number": 42}'

# Test case 2: JSON wrapped in markdown code block
markdown_json = '''```json
{"name": "test", "count": 10}
'''

# Test case 3: JSON object embedded in text
embedded_json = 'Here is some text before {"embedded": true, "value": 123} and after'

# Test case 4: Nested JSON object
nested_json = '{"outer": {"inner": "value"}, "list": [1, 2, 3]}'

# Test case 5: JSON with escaped quotes
escaped_json = '{"escaped": "value with \\"quotes\\"", "other": "test"}'

ARGS = {
    'text': simple_json
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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_model_config Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/memqa/global_config.py
code served (permissive licence) · get_code("f9e247fbeeffa980")
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 dict. 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 os
from pathlib import Path
from unittest.mock import patch, MagicMock

# Mock environment variables and file system to avoid actual file I/O
os.environ.pop('OPENAI_API_KEY', None)
os.environ.pop('VLLM_API_KEY', None)
os.environ.pop('ATMBENCH_STRICT_API_KEYS', None)

# Create minimal stand-in for the module to avoid actual file operations
import sys
from types import ModuleType

# We need to set up the context so the module loads correctly
# Mock the api_keys directory check
mock_api_keys_dir = Path('/tmp/api_keys_mock')

# Patch the file existence checks and reads
def mock_exists(self):
    return False

def mock_open_file(self, *args, **kwargs):
    raise FileNotFoundError()

# We'll patch Path.exists at import time by mocking the environment
with patch.dict(os.environ, {'ATMBENCH_STRICT_API_KEYS': '0'}):
    # Now we can safely call the functions with empty API keys
    ARGS = {
        'provider': 'vllm',
        'model_type': 'vision',
    }
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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_output_path Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/memqa/global_config.py
code served (permissive licence) · get_code("3b464087313f9c13")
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 PosixPath. 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.

from pathlib import Path

# The function get_output_path takes:
# - module: str (e.g., "image", "email", "event")
# - filename: str or None (optional filename to append)

# Reading the function body:
# 1. It creates a base_path = OUTPUT_DIR / module
# 2. If filename is provided, it appends it: base_path / filename
# 3. Returns a Path object
# The function only uses string concatenation with Path objects, so we just need valid strings.

ARGS = {
    "module": "image",
    "filename": "test_output.txt"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type PosixPath — a PosixPath output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a PosixPath 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
item_id Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/memory_variants.py
code served (permissive licence) · get_code("fba80e0d9cdecb65")
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 str. 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
from pathlib import Path
from typing import Any

torch.manual_seed(0)

ARGS = {
    'item': {
        'file_path': '/home/user/documents/example_file.txt',
    },
    'path_key': 'file_path',
}
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 str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
last_token_pool Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/memqa/retrieve/retrievers.py
code served (permissive licence) · get_code("b15d6769e9f3531d")
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)

# Create a batch of 2 sequences with hidden state dimension 4
batch_size = 2
seq_length = 3
hidden_dim = 4

# Create last_hidden_states: shape (batch_size, seq_length, hidden_dim)
last_hidden_states = torch.randn(batch_size, seq_length, hidden_dim)

# Create attention_mask: shape (batch_size, seq_length)
# For left padding case: all sequences have full attention at the end
attention_mask = torch.ones(batch_size, seq_length, dtype=torch.long)

ARGS = {
    "last_hidden_states": last_hidden_states,
    "attention_mask": attention_mask,
}
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
load_json Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/extract_answer.py
code served (permissive licence) · get_code("e1a33f82f601f5fe")
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 dict. 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 json
from pathlib import Path
import tempfile

# Create a temporary directory and a test JSON file
temp_dir = tempfile.mkdtemp()
test_file = Path(temp_dir) / "test.json"

# Write some test data to the JSON file
test_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "object"}
}

with test_file.open('w', encoding='utf-8') as f:
    json.dump(test_data, f)

# Define the arguments dictionary
ARGS = {
    "path": test_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
load_json Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/extract_usage.py
code served (permissive licence) · get_code("d0d3faa7cc886a81")
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 dict. 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 json
from pathlib import Path
import tempfile
import os

# Create a temporary directory and a temporary JSON file
temp_dir = tempfile.mkdtemp()
temp_json_path = Path(temp_dir) / "test.json"

# Write a valid JSON file
test_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "object"}
}

with open(temp_json_path, 'w', encoding='utf-8') as f:
    json.dump(test_data, f)

ARGS = {
    "path": temp_json_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
load_json_list Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/memory_variants.py
code served (permissive licence) · get_code("97f74443ec5601e5")
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 list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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 json
from pathlib import Path
import tempfile

# Create a temporary JSON file with valid list content
with tempfile.NamedTemporaryFile(mode='w', suffix='.json', delete=False, encoding='utf-8') as f:
    json_data = [
        {"key1": "value1", "key2": 42},
        {"key1": "value2", "key2": 100},
        {"nested": {"a": 1, "b": 2}}
    ]
    json.dump(json_data, f)
    temp_path = Path(f.name)

ARGS = {
    "path": temp_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
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
load_json_with_recovery Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/extract_answer.py
code served (permissive licence) · get_code("288ca5fef01abdbc")
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 dict. 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 json
from pathlib import Path
import tempfile
import os

def _find_first_json_object(text: str):
    in_string = False
    escape = False
    depth = 0
    start_idx = None
    for idx, ch in enumerate(text):
        if start_idx is None:
            if ch == '{':
                start_idx = idx
                depth = 1
            continue
        if in_string:
            if escape:
                escape = False
                continue
            if ch == '\\':
                escape = True
                continue
            if ch == '"':
                in_string = False
            continue
        if ch == '"':
            in_string = True
            continue
        if ch == '{':
            depth += 1
            continue
        if ch == '}':
            depth -= 1
            if depth == 0:
                return text[start_idx:idx + 1]
    return None

def load_json_with_recovery(path: Path):
    text = path.read_text(encoding='utf-8')
    try:
        return json.loads(text)
    except json.JSONDecodeError:
        candidate = _find_first_json_object(text)
        if candidate:
            return json.loads(candidate)
        raise

# Create a temporary directory and file
temp_dir = tempfile.mkdtemp()
temp_file_path = Path(temp_dir) / "test.json"

# Write valid JSON to the file
valid_json_content = '{"key": "value", "number": 42, "nested": {"inner": true}}'
temp_file_path.write_text(valid_json_content, encoding='utf-8')

ARGS = {
    'path': temp_file_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict 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
load_question_ids Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/collect_results.py
code served (permissive licence) · get_code("00dedf0286b4a63c")
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 list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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.

from pathlib import Path
import tempfile

# Create a temporary directory with a question_ids.txt file
temp_dir = tempfile.TemporaryDirectory()
eval_root = Path(temp_dir.name)

# Create a sample question_ids.txt file
ids_file = eval_root / 'question_ids.txt'
ids_file.write_text('q_001\nq_002\nq_003\n\nq_004\n', encoding='utf-8')

ARGS = {
    'eval_root': eval_root,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
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
load_questions Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/prepare_sandbox.py
code served (permissive licence) · get_code("e748b717c3a734f2")
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 list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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 json
import tempfile
from pathlib import Path

# Create a temporary JSON file with sample ATM-hard questions
temp_dir = tempfile.TemporaryDirectory()
temp_path = Path(temp_dir.name) / "questions.json"

sample_data = [
    {
        "id": "q1",
        "question": "What is the capital of France?",
        "ground_truth": "Paris",
        "other_field": "should_be_stripped"
    },
    {
        "id": "q2",
        "question": "What is 2 + 2?",
        "ground_truth": "4",
        "answer": "should_also_be_stripped"
    }
]

with open(temp_path, 'w', encoding='utf-8') as f:
    json.dump(sample_data, f)

ARGS = {
    'path': temp_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
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
normalize_memory_mode Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/memory_variants.py
code served (permissive licence) · get_code("28eaaa08671641d6")
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 str. 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.

# Setup code for normalize_memory_mode function

ARGS = {
    'mode': 'sgm'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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
resolve_repo_path Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/config.py
code served (permissive licence) · get_code("49b8dc6d41651408")
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 PosixPath. 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 tempfile
from pathlib import Path

# Create a temporary directory structure to serve as REPO_ROOT
with tempfile.TemporaryDirectory() as tmpdir:
    repo_root = Path(tmpdir)
    parent_dir = repo_root.parent
    
    # Create a subdirectory structure
    subdir = repo_root / "subdir"
    subdir.mkdir(exist_ok=True)
    
    # Create some test files
    (repo_root / "test.txt").touch()
    (subdir / "nested.txt").touch()
    
    # Test with a relative path
    test_relative_path = "subdir/nested.txt"
    
    # Test with an absolute path
    test_absolute_path = str(repo_root / "test.txt")
    
    # For this test, we'll use a relative path since the function
    # handles both absolute and relative paths
    ARGS = {
        "path": "subdir/nested.txt"
    }
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type PosixPath — a PosixPath output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a PosixPath 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
runtime_eval_value Ran
ran; our contract draft was wrong, not the code
JingbiaoMei/ATM-Bench/agent_systems/collect_results.py
code served (permissive licence) · get_code("088fcf6bbfb120b4")
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 str. 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 numpy as np

torch.manual_seed(0)

# Analyze the function:
# - value: str parameter, can be empty string, a string value, or in legacy_values
# - fallback: str parameter, returned when value is empty or in legacy_values
# - legacy_values: tuple of strings, optional parameter with default ()

# The function checks:
# 1. if not value (empty string check)
# 2. if value in legacy_values (membership test)
# 3. returns fallback if either condition is true
# 4. otherwise returns value

ARGS = {
    "value": "current_value",
    "fallback": "default_value",
    "legacy_values": ("old_value_1", "old_value_2", "deprecated")
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str 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_api_key Not yet run
unverified (harvested, no recorded run)
JingbiaoMei/ATM-Bench/memqa/global_config.py
code served (permissive licence) · get_code("159e7f0ea5b4e0b2")
How it ran

no invocation recorded.

load_qwen3_vl_embedder Not yet run
unverified (harvested, no recorded run)
JingbiaoMei/ATM-Bench/memqa/retrieve/retrievers.py
code served (permissive licence) · get_code("99f295e69c490e0a")
How it ran

no invocation recorded.

load_qwen3_vl_reranker Not yet run
unverified (harvested, no recorded run)
JingbiaoMei/ATM-Bench/memqa/retrieve/rerankers.py
code served (permissive licence) · get_code("e34f5854798ea44f")
How it ran

no invocation recorded.

Also present in other papers' code

Of 18 samples lifted from this paper's linked repositories, 18 carry a function identity and 0 do not. Of the 17 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 17 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.

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

Personalized AI assistants must recall and reason over long-term user memory, which naturally spans multiple modalities and sources such as images, videos, and emails. However, existing Long-term Memory benchmarks focus primarily on dialogue history, failing to capture realistic personalized references grounded in lived experience. We introduce ATM-Bench, the first benchmark for multimodal, multi-source personalized referential Memory QA. ATM-Bench contains approximately four years of privacy-preserving personal memory data and human-annotated question-answer pairs with ground-truth memory evidence, including queries that require resolving personal references, multi-evidence reasoning from multi-source and handling conflicting evidence. We propose Schema-Guided Memory (SGM) to structurally represent memory items originated from different sources. In experiments, we implement 5 state-of-the-art memory systems along with a standard RAG baseline and evaluate variants with different memory ingestion, retrieval, and answer generation techniques. We find poor performance (under 20% accuracy) on the ATM-Bench-Hard set, and that SGM improves performance over Descriptive Memory commonly adopted in prior works. Code available at: https://github.com/JingbiaoMei/ATM-Bench.

For agents

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

get_harvested_code_for_paper("2603.01990")
get_code_for_paper("2603.01990")
have("2603.01990")

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

curl https://syntology.ai/api/ran/2603.01990.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.01990.svg)](https://syntology.ai/paper/2603.01990)

Connect an agent — have() is free.