SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2606.03463 · 2026

DMF: A Deterministic Memory Framework for Conversational AI Agents

Matteo Stabile, Enrico Zimuel

arXiv · PDF · Open in the Atlas

Code that ran

19 of 19 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; 6 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 13 failures of our instrument; 0 no recorded run), measured 2026-11-21; 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/2606.03463.json · badge: /api/ran/2606.03463.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
matstech/dmf-benchmarks canonical (classifier label)
classifier basis: census
19 ran · 0 no recorded run
FunctionStatusWhere it lives
build_answerer_user_prompt Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/prompts.py
code served (permissive licence) · get_code("340bde46c69a75fb")
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 json
from typing import Any

CATEGORY_2_DATE_INSTRUCTION = 'For time or date questions, resolve relative expressions such as yesterday, last year, this month, and next month against dates or timestamps available in the native context. Prefer an absolute date, month, or year over a relative phrase.'

ANSWERER_USER_PROMPT_TEMPLATE = '\nUse the provided native conversational memory context to answer the LoCoMo question.\nGive a concise answer. Answer with exact words from the native context whenever possible.\n{category_instruction}\n\nNative context:\n{native_context}\n\nQuestion: {question}\nShort answer:\n'.strip()

def _serialize_native_context(native_context: Any) -> str:
    if isinstance(native_context, str):
        return native_context
    return json.dumps(native_context, ensure_ascii=False, indent=2)

def build_answerer_user_prompt(native_context: Any, question: str, *, category: int=0) -> str:
    category_instruction = CATEGORY_2_DATE_INSTRUCTION if category == 2 else ''
    return ANSWERER_USER_PROMPT_TEMPLATE.format(category_instruction=category_instruction, native_context=_serialize_native_context(native_context), question=question)


ARGS = {
    'native_context': {
        'conversation': [
            {'speaker': 'Alice', 'text': 'I was born in 1990'},
            {'speaker': 'Bob', 'text': 'That is interesting'}
        ],
        'timestamp': '2024-01-15'
    },
    'question': 'When was Alice born?',
    'category': 0
}
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
category_name Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/questions.py
code served (permissive licence) · get_code("428bea11a75a56e8")
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

def category_name(category: int) -> str:
    return {1: 'multi-hop', 2: 'temporal', 3: 'open-domain', 4: 'single-hop', 5: 'adversarial'}.get(category, f'category-{category}')

ARGS = {
    'category': 1
}
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
conversation_unit_id Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/adapter.py
code served (permissive licence) · get_code("b876e5700698ca84")
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

# conversation_unit_id takes an int conversation_idx and returns a formatted string
# The function simply formats the index + 1 as a 4-digit zero-padded number in a string

ARGS = {
    'conversation_idx': 0
}
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
exact_match Ran
ran, honoured the contract we drafted
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/rigorous.py
code served (permissive licence) · get_code("6aea303b6aec6c67")
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 float, ndim 0, dtype family float, shape () 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 re
import string

def normalize_text(text: str) -> str:
    text = text.lower()
    text = text.translate(str.maketrans('', '', string.punctuation))
    text = re.sub('\\b(a|an|the)\\b', ' ', text)
    return re.sub('\\s+', ' ', text).strip()

def exact_match(prediction: str, ground_truth: str) -> float:
    return 1.0 if normalize_text(prediction) == normalize_text(ground_truth) else 0.0

ARGS = {
    "prediction": "The quick brown fox jumps over the lazy dog",
    "ground_truth": "the quick brown fox jumps over a lazy dog"
}
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 float, ndim 0, dtype family float, shape () — shape () 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_dataset Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/longmemeval/dataset.py
code served (permissive licence) · get_code("73cba72603aa228a")
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
import os

# Create a temporary JSON file with sample dataset
temp_dir = tempfile.mkdtemp()
dataset_path = os.path.join(temp_dir, "dataset.json")

sample_data = [
    {"id": 1, "text": "sample text 1", "label": "A"},
    {"id": 2, "text": "sample text 2", "label": "B"},
    {"id": 3, "text": "sample text 3", "label": "C"}
]

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

ARGS = {
    "path": dataset_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_category_5_prediction Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/prompts.py
code served (permissive licence) · get_code("64c55dcb60229eb9")
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

# The function only uses string operations, so we need simple string arguments
prediction = "  (B)  "
answer = "The answer to the question"

ARGS = {
    "prediction": prediction,
    "answer": answer
}
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
normalize_locomo_turn Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/dataset.py
code served (permissive licence) · get_code("11bfecc717880f84")
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
from typing import Any

torch.manual_seed(0)

ARGS = {
    "turn": {
        "speaker": "Alice",
        "text": "Hello, what do you see in this image?",
        "query": "object detection",
        "blip_caption": "A cat sitting on a chair"
    }
}
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
normalize_text Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/rigorous.py
code served (permissive licence) · get_code("ae71bb9ba2af46bf")
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 re
import string

def normalize_text(text: str) -> str:
    text = text.lower()
    text = text.translate(str.maketrans('', '', string.punctuation))
    text = re.sub('\\b(a|an|the)\\b', ' ', text)
    return re.sub('\\s+', ' ', text).strip()

ARGS = {
    "text": "The Quick Brown Fox! A lazy dog, an example."
}
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
normalized_tokens Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/rigorous.py
code served (permissive licence) · get_code("20b3d8cf23b0d87c")
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 re
import string

def normalize_text(text: str) -> str:
    text = text.lower()
    text = text.translate(str.maketrans('', '', string.punctuation))
    text = re.sub('\\b(a|an|the)\\b', ' ', text)
    return re.sub('\\s+', ' ', text).strip()

def normalized_tokens(text: str) -> list[str]:
    normalized = normalize_text(text)
    return normalized.split() if normalized else []

ARGS = {
    "text": "The quick brown fox jumps over a lazy dog!"
}
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
official_ground_truth_answer Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/prompts.py
code served (permissive licence) · get_code("0174084e67c613d1")
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

ARGS = {
    'category': 5,
    'answer': 'This is a sample answer'
}
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
parse_judge_response Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/judge.py
code served (permissive licence) · get_code("c268d1b34b72988e")
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 tuple, 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 re

def parse_judge_response(text: str) -> tuple[str, float, str]:
    """Parse JSON or text judge output into verdict, score, and reason."""
    try:
        payload = json.loads(text)
    except json.JSONDecodeError:
        payload = None
    if isinstance(payload, dict):
        raw_label = str(payload.get('label', '')).strip().upper()
        verdict = raw_label if raw_label in {'CORRECT', 'WRONG'} else 'WRONG'
        reasoning = str(payload.get('reasoning', '')).strip()
        reason = reasoning if reasoning else text.strip()
        score = 1.0 if verdict == 'CORRECT' else 0.0
        return (verdict, score, reason)
    verdict_match = re.search('VERDICT:\\s*(CORRECT|WRONG)', text, flags=re.IGNORECASE)
    reason_match = re.search('REASON:\\s*(.*)', text, flags=re.IGNORECASE | re.DOTALL)
    verdict = verdict_match.group(1).upper() if verdict_match else 'WRONG'
    reason = reason_match.group(1).strip() if reason_match else text.strip()
    score = 1.0 if verdict == 'CORRECT' else 0.0
    return (verdict, score, reason)


# Test case 1: JSON input with CORRECT label
json_correct = json.dumps({"label": "CORRECT", "reasoning": "The answer is accurate."})

# Test case 2: JSON input with WRONG label
json_wrong = json.dumps({"label": "WRONG", "reasoning": "The answer is inaccurate."})

# Test case 3: Plain text with VERDICT and REASON tags
text_verdict = "VERDICT: CORRECT\nREASON: The solution is correct because it follows the algorithm."

# Test case 4: Plain text without structured format
plain_text = "This is just plain text without any structured format."

ARGS = {
    "text": json_correct
}
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 tuple, ndim 1, dtype family float — a tuple 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
parse_locomo_date Ran
ran, honoured the contract we drafted
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/dataset.py
code served (permissive licence) · get_code("f902c4ca0121dbe1")
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 float, ndim 0, dtype family float, shape () 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.

from datetime import datetime, timezone

ARGS = {
    "date_str": "03:45 PM on 15 December, 2023"
}
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 float, ndim 0, dtype family float, shape () — shape () 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
parse_longmemeval_date Ran
ran, honoured the contract we drafted
matstech/dmf-benchmarks/dmf_bench/benchmarks/longmemeval/dataset.py
code served (permissive licence) · get_code("565e4cb5d95611bb")
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 int, ndim 0, dtype family int, shape () 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 re
from datetime import datetime, timezone

def parse_longmemeval_date(date_str: str) -> int | None:
    """'2023/05/01 (Mon) 21:05' -> Unix epoch int (UTC)."""
    try:
        cleaned = re.sub('\\s*\\([A-Za-z]+\\)\\s*', ' ', date_str).strip()
        dt = datetime.strptime(cleaned, '%Y/%m/%d %H:%M').replace(tzinfo=timezone.utc)
        return int(dt.timestamp())
    except (ValueError, TypeError):
        return None


ARGS = {
    'date_str': '2023/05/01 (Mon) 21:05'
}
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 int, ndim 0, dtype family int, shape () — shape () 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
parse_longmemeval_date_human Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/longmemeval/dataset.py
code served (permissive licence) · get_code("2fb4d363847e5b3f")
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 re
from datetime import datetime, timezone

ARGS = {
    "date_str": "2023/05/01 (Mon) 21:05"
}
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
precision_at_k Ran
ran, honoured the contract we drafted
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/ablation.py
code served (permissive licence) · get_code("1c147c09654b3caa")
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 float, ndim 0, dtype family float, shape () 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
import numpy as np

torch.manual_seed(0)

# Construct arguments for precision_at_k function
# The function takes:
# - retrieved_ids: list[str] - a list of retrieved item IDs
# - gold_ids: set[str] - a set of ground truth item IDs

retrieved_ids = ["item_1", "item_2", "item_3", "item_4", "item_5"]
gold_ids = {"item_2", "item_4", "item_6", "item_7"}

ARGS = {
    "retrieved_ids": retrieved_ids,
    "gold_ids": gold_ids
}
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 float, ndim 0, dtype family float, shape () — shape () 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
question_id Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/adapter.py
code served (permissive licence) · get_code("002ac560607595f7")
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

torch.manual_seed(0)

ARGS = {
    'conversation_idx': 0,
    'question_idx': 1,
}
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
recall_at_k Ran
ran, honoured the contract we drafted
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/ablation.py
code served (permissive licence) · get_code("20a53bb137695e55")
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 float, ndim 0, dtype family float, shape () 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

# Set random seed for reproducibility
torch.manual_seed(0)

# Define ARGS dict with values for recall_at_k parameters
ARGS = {
    "retrieved_ids": ["doc_1", "doc_2", "doc_3", "doc_5"],
    "gold_ids": {"doc_1", "doc_2", "doc_4"}
}
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 float, ndim 0, dtype family float, shape () — shape () 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
reciprocal_rank Ran
ran, honoured the contract we drafted
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/ablation.py
code served (permissive licence) · get_code("939c1d4e6225165a")
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 float, ndim 0, dtype family float, shape () 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

# Construct valid arguments for reciprocal_rank function
retrieved_ids = ["item_1", "item_2", "item_3", "item_4"]
gold_ids = {"item_2", "item_5"}

ARGS = {
    "retrieved_ids": retrieved_ids,
    "gold_ids": gold_ids,
}
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 float, ndim 0, dtype family float, shape () — shape () 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
serialize_locomo_turn_for_dmf Ran
ran; our contract draft was wrong, not the code
matstech/dmf-benchmarks/dmf_bench/benchmarks/locomo/dataset.py
code served (permissive licence) · get_code("1e282bca38ed5dc9")
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 typing import Any

# Construct a minimal turn dict that will be passed to serialize_locomo_turn_for_dmf
turn = {
    'text': 'This is an utterance from the speaker.',
    'speaker': 'Alice',
    'query': 'What is in the image?',
    'blip_caption': 'A dog playing in the park.'
}

ARGS = {
    'turn': turn
}
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

Also present in other papers' code

Of 19 samples lifted from this paper's linked repositories, 19 carry a function identity and 0 do not. Of the 19 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 19 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 2026-11-21.

Repositories linked to this paper

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

Abstract

Conversational AI agents require memory systems that are both scalable and semantically coherent across long interaction horizons. Existing approaches rely predominantly on large language model (LLM)-based summarisation at write time, which introduces non-determinism, escalating token costs, and opacity in pruning decisions. We present the Deterministic Memory Framework (DMF), a CPU-first approach that replaces generative memory compression with a fully deterministic pipeline grounded in classical NLP analysis, vector geometry, and mathematical scoring. DMF assigns each conversational interaction a Survival Score Ω computed from deterministic content signals, conversational cues, and structured provenance, combined through a logistic projection. An interaction-count decay law, denoted as Ω eff (∆n), governs how relevance evolves as new turns arrive, where ∆n is the number of newer interactions rather than wall-clock time, preserving full determinism. We present the mathematical formulation of DMF, its structured recall pipeline, the pruning decision procedure, and the evaluation protocol. Experiments are conducted on a purposebuilt benchmark using the LoCoMo and LongMemEval datasets. We compare DMF against Mem0, a popular memory layer for AI agents. DMF achieves comparable accuracy while using zero tokens to prepare the memory context and 5× to 242× fewer tokens over the entire conversation. These results show that it is possible to eliminate LLM calls from the memory-management loop, reducing token costs to nearly zero and enabling deterministic memory systems for conversational AI agents.

For agents

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

get_harvested_code_for_paper("2606.03463")
get_code_for_paper("2606.03463")
have("2606.03463")

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

curl https://syntology.ai/api/ran/2606.03463.json

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

[![Syntology run record](https://syntology.ai/api/ran/2606.03463.svg)](https://syntology.ai/paper/2606.03463)

Connect an agent — have() is free.