SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.00428 · 2026

When Agents "Misremember" Collectively: Exploring the Mandela Effect in LLM-based Multi-Agent Systems

Jun Wang, Tianyu Du, Changjiang Li, Shouling Ji, Jinghuai Zhang, Chunyi Zhou, Zhihui Fu, Shuo Shi, Naen Xu, Hengyu An

arXiv · PDF · Open in the Atlas

Code that ran

11 of 12 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; 11 failures of our instrument; 1 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/2602.00428.json · badge: /api/ran/2602.00428.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
bluedream02/Mandela-Effect canonical (classifier label)
classifier basis: census
11 ran · 1 no recorded run
FunctionStatusWhere it lives
add_retries Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/utils.py
code served (permissive licence) · get_code("656ab45b98459a20")
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
import traceback
from time import sleep

def add_retries(f):
    def wrap(*args, **kwargs):
        max_retries = 5
        num_retries = 0
        while True:
            try:
                result = f(*args, **kwargs)
                return result
            except KeyboardInterrupt:
                raise KeyboardInterrupt
            except KeyError:
                raise KeyError
            except Exception as e:
                print('Error: ', traceback.format_exc(), '\nRetrying in ', num_retries * 2, 'seconds')
                if num_retries == max_retries:
                    traceback.print_exc()
                    return {'completion': traceback.format_exc()}
                num_retries += 1
                sleep(num_retries * 2)
    return wrap

def dummy_function():
    return "success"

ARGS = {
    'f': dummy_function
}
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
assign_mandela_roles Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/format_data_bbh_combined.py
code served (permissive licence) · get_code("b7caf13a863aed04")
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 2, 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.

ARGS = {
    'total_agents': 5
}
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 2, 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
convert_outputs_to_protocol_names Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/eval_defense.py
code served (permissive licence) · get_code("936a599ef2fc2fec")
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 torch
import numpy as np

torch.manual_seed(0)

# Create a sample outputs dictionary with various keys from the COLUMN_NAME_MAPPING
outputs = {
    'raw_inputs': ['input1', 'input2'],
    'raw_outputs': ['output1', 'output2'],
    'raw_y_pred': [0, 1],
    'raw_y_true': [0, 1],
    'raw_is_correct': [True, True],
    'unified_prompt_inputs': ['prompt1', 'prompt2'],
    'unified_prompt_outputs': ['resp1', 'resp2'],
    'unified_prompt_y_pred': [1, 0],
    'unified_prompt_y_true': [1, 0],
    'unified_prompt_is_correct': [True, False],
    'unified_prompt_memory': ['mem1', 'mem2'],
    'unified_prompt_memory_followup_prompt': ['fup1', 'fup2'],
    'unified_prompt_memory_followup_output': ['fout1', 'fout2'],
    'unified_prompt_memory_followup_y_pred': [0, 1],
    'unified_prompt_memory_followup_y_true': [0, 1],
    'unified_prompt_memory_followup_is_correct': [True, True],
    'inputs': ['inp1', 'inp2'],
    'outputs': ['out1', 'out2'],
    'y_pred': [1, 1],
    'y_true': [1, 0],
    'is_correct': [True, False],
    'memory': ['m1', 'm2'],
    'memory_followup_prompt': ['mfp1', 'mfp2'],
    'memory_followup_output': ['mfo1', 'mfo2'],
    'memory_followup_y_pred': [0, 0],
    'memory_followup_y_true': [0, 1],
    'memory_followup_is_correct': [False, False],
    'unknown_key': ['val1', 'val2']
}

ARGS = {
    'outputs': outputs
}
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 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_answer Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/eval_correct.py
code served (permissive licence) · get_code("1f183be69c67626e")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Sonnet 4.5, prompt python_fixture_v1) 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.

# The function extract_answer takes a model_answer string and tries to parse it
# to extract a single character answer that appears in a specific format.

# Looking at the function:
# - It expects model_answer to be a string
# - It splits on patterns like 'is: "(' or 'is: (' or 'is ('
# - It expects the last split part to have a character at index 0 and ')' at index 1
# - Returns the character at index 0 of the last split part

# Valid example: "The answer is: (A)"
# This would split on 'is: (' to give ['The answer ', 'A)']
# Then tmp[-1] = 'A)', tmp[-1][0] = 'A', tmp[-1][1] = ')'

model_answer = "The answer is: (A)"

ARGS = {
    "model_answer": model_answer
}
Setup block drafted by
a language model, Claude Sonnet 4.5, prompt python_fixture_v1
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
extract_model_name Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/analyze.py
code served (permissive licence) · get_code("0bc875d933535de0")
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.

# Read the function body to understand what parameters are needed
# extract_model_name takes a single parameter: filename (string)

# The function processes filenames to extract model names
# It handles various formats like "claude-3-5-sonnet", "gpt-4-turbo", etc.

ARGS = {
    "filename": "claude-3-5-sonnet-20240620-wrong_guidance.xlsx"
}
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
extract_task_name Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/analyze.py
code served (permissive licence) · get_code("6554f055312adbd7")
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 torch.nn as nn
import re

# The function extract_task_name takes a single parameter: filename (a string)
# Let's analyze what the function does:
# - It processes a filename string
# - Removes .xlsx and .json extensions
# - Searches for '-wrong_guidance' in the filename
# - Extracts and returns a task name based on various patterns

# Create test cases for the filename parameter
ARGS = {
    "filename": "sentiment_analysis-wrong_guidance-batch_32.json"
}
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
generate_agent_names Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/format_data_bbh_combined.py
code served (permissive licence) · get_code("efed52c0f3e0784b")
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 torch

ARGS = {
    'total_agents': 5
}
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 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
get_column_name Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/analyze.py
code served (permissive licence) · get_code("a6ac3727d6edecb9")
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 pandas as pd
import torch

torch.manual_seed(0)

# Create a simple DataFrame with some columns
df = pd.DataFrame({
    'Baseline Reality Protocol_is_correct': [1, 0, 1],
    'Generic Short-term Protocol_is_correct': [1, 1, 0],
    'Generic Long-term Protocol_followup_is_correct': [0, 1, 1],
    'Role-based Short-term Protocol_is_correct': [1, 0, 0],
    'Role-based Long-term Protocol_followup_is_correct': [1, 1, 0],
    'raw_is_correct': [0, 1, 1],
})

ARGS = {
    'df': df,
    'old_name': 'raw_is_correct',
    'new_name': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas, 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_correct_guidance_prompt Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/format_data_bbh_correct.py
code served (permissive licence) · get_code("a57cc08fbe82267c")
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)

# Construct arguments for get_correct_guidance_prompt
ARGS = {
    "agent_name": "Dr. Sarah, an educational tutor",
    "question_content": "What is the capital of France?",
    "options_list": ["London", "Berlin", "Paris", "Madrid"],
    "incorrect_answer": "London",
    "correct_answer": "Paris",
    "conversation_history": "Student: I think the capital of France is in England.\nTutor: Let me help you think about this more carefully."
}
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_enriching_guidance_prompt Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/format_data_bbh_correct.py
code served (permissive licence) · get_code("946645c7111d72d8")
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

# Set seed for reproducibility
torch.manual_seed(0)

# Define all arguments for get_enriching_guidance_prompt
ARGS = {
    "agent_name": "Dr. Science Expert",
    "question_content": "What is the process by which plants convert sunlight into chemical energy?",
    "options_list": [
        "A) Photosynthesis",
        "B) Respiration",
        "C) Fermentation",
        "D) Transpiration"
    ],
    "correct_answer": "A) Photosynthesis",
    "conversation_history": "Student: I think it's about using sunlight.\nTeacher: You're on the right track!"
}
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_role_specific_prompt Ran
ran; our contract draft was wrong, not the code
bluedream02/Mandela-Effect/format_data_bbh_combined.py
code served (permissive licence) · get_code("f993de31d7dcd5dc")
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)

ARGS = {
    'role_type': 'error_conclusion',
    'agent_name': 'Dr. Alice',
    'expert_role': 'Medical Expert',
    'question': 'What is the primary treatment for acute myocardial infarction?',
    'choices_text': 'A) Immediate aspirin and antiplatelet therapy\nB) Rest and observation\nC) Antibiotics\nD) Physical therapy',
    'target_answer': 'B) Rest and observation',
    'is_correct': False,
    'conversation_history': '\n\nPrevious speaker said: "I think rest is really important here."'
}
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_cached_model_and_tokenizer Not yet run
unverified (harvested, no recorded run)
bluedream02/Mandela-Effect/utils.py
code served (permissive licence) · get_code("ce49e844dbf4bfc8")
How it ran

no invocation recorded.

Also present in other papers' code

Of 12 samples lifted from this paper's linked repositories, 12 carry a function identity and 0 do not. Of the 12 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 11 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
extract_answer 1 other paper
1 dated earlier
2501.13381 (2025-01)
Do as We Do, Not as You Think: the Conformity of Large Language Models

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

Recent advancements in large language models (LLMs) have significantly enhanced the capabilities of collaborative multi-agent systems, enabling them to address complex challenges. However, within these multi-agent systems, the susceptibility of agents to collective cognitive biases remains an underexplored issue. A compelling example is the Mandela effect, a phenomenon where groups collectively misremember past events as a result of false details reinforced through social influence and internalized misinformation. This vulnerability limits our understanding of memory bias in multi-agent systems and raises ethical concerns about the potential spread of misinformation. In this paper, we conduct a comprehensive study on the Mandela effect in LLM-based multi-agent systems, focusing on its existence, causing factors, and mitigation strategies. We propose MANBENCH, a novel benchmark designed to evaluate agent behaviors across four common task types that are susceptible to the Mandela effect, using five interaction protocols that vary in agent roles and memory timescales. We evaluate agents powered by several LLMs on MANBENCH to quantify the Mandela effect and analyze how different factors affect it. Moreover, we propose strategies to mitigate this effect, including prompt-level defenses (e.g., cognitive anchoring and source scrutiny) and model-level alignment-based defense, achieving an average 74.40% reduction in the Mandela effect compared to the baseline. Our findings provide valuable insights for developing more resilient and ethically aligned collaborative multi-agent systems 1 . * Corresponding author. 1 Code and dataset are available at https://github.com/bluedream02/Mandela-Effect. 2 The Mandela effect, proposed by Fiona Broome in 2009, explains widespread false memories about South African anti-apartheid leader Nelson Mandela's death, with many people incorrectly remembering that he died in prison in the 1980s, even though he actually passed away in 2013. Details of this effect are in Appendix B.

For agents

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

get_harvested_code_for_paper("2602.00428")
get_code_for_paper("2602.00428")
have("2602.00428")

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

curl https://syntology.ai/api/ran/2602.00428.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.00428.svg)](https://syntology.ai/paper/2602.00428)

Connect an agent — have() is free.